AI GLOSSARY

A method for comparing two versions of a model, interface, prompt, or system behavior by exposing different users or requests to each version and measuring which performs better against a defined metric. In AI product development, A/B testing is one of the clearest ways to validate whether a change actually improves outcomes in the real world rather than just looking promising in offline evaluation.

The removal or disabling of a component in a model, dataset, or pipeline to understand how much that component contributes to the overall result. In AI, the term usually appears in the context of ablation-study, where researchers systematically test what happens when a feature, training signal, or architectural choice is taken away.

An experiment where components of an AI system are systematically removed or disabled, one at a time, to determine how much each contributes to overall performance. The term "ablation", which loosely means to remove or erode, is most commonly used in neuroscience and biology when describing the removal of biological components. In AI, it helps researchers identify which parts of a model or pipeline are actually doing the useful work.

The ongoing process of detecting and responding to misuse of an AI system, identifying patterns of harmful, policy-violating, or malicious use across all interactions. Abuse monitoring should combine automated detection with human review, and is used for maintaining the safety and integrity of deployed AI products at scale, where manual review of every interaction is impractical. The same infrastructure can, in principle, also enable broad surveillance of user activity, making independent audits of its ethical and responsible use an essential safeguard in any system deployed this way.

A specialized hardware chip designed to speed up AI workloads, especially the matrix operations and parallel computations used in training and inference. GPUs, TPUs, and other dedicated accelerators are central to modern AI because general-purpose CPUs are usually too slow or too inefficient for large-scale model workloads.

A form of activation steering in which a concept vector is added to a model's internal activations during inference to amplify or introduce a target behaviour. The vector is typically derived by contrasting activations from prompts that do and do not exhibit the concept. For example, subtracting the 'sad' activation from the 'happy' activation yields a direction that, when added, steers the model toward happier outputs.

See also: Activation Steering, Activation Subtraction, Representation Engineering.

A mathematical function applied to the output of a neuron in a neural network, determining whether and how strongly that neuron fires. In practice, a neural network passes data through repeated cycles of linear transformation followed by an activation function, and it is this alternation that gives the network its ability to learn complex, non-linear patterns. Without activation functions, stacking layers would be mathematically pointless, as any number of linear transformations collapse into a single one.

A technique for influencing a neural network's behavior by directly modifying its internal activations during inference, adding or subtracting vectors that correspond to specific concepts or behaviors. Rather than changing the model's weights or prompting it differently, activation steering intervenes at the mechanistic level, making it a valuable tool for interpretability and alignment research. Adding a vector is called activation addition; subtracting one is called activation subtraction.

A form of activation steering in which a concept vector is subtracted from a model's internal activations during inference to suppress or remove a target behaviour. Subtracting the vector associated with a concept (such as deception or refusal) reduces the model's tendency to exhibit it, without retraining or fine-tuning.

See also: Activation Steering, Activation Addition, Representation Engineering.

A training approach where the model identifies which data points it is most uncertain about and requests labels for those specific examples, rather than being trained on a randomly selected dataset. This makes active learning particularly useful when labeled data is expensive or time-consuming to produce, as human annotation effort is focused where it matters most, achieving strong performance with considerably less labeled data.

A reinforcement learning architecture combining two components: an actor that decides which actions to take, and a critic that evaluates how good those actions were. The critic's feedback shapes the actor's learning over time, making the combination more stable and sample-efficient than either approach on its own. See also: Policy Gradient, Reward Model.

A deliberate attempt to manipulate an AI system by crafting inputs designed to cause errors, bypass safety measures, or produce unintended outputs. Adversarial attacks exploit weaknesses in how models generalize, where small, carefully designed perturbations, often imperceptible to humans, can cause confident misclassification or harmful outputs.

A carefully crafted input, often indistinguishable from a normal input to a human, that causes an AI model to make a specific, incorrect prediction. A classic example is an image of a panda with imperceptible noise added, causing a classifier to confidently identify it as a gibbon. Where an adversarial attack describes the act of manipulation, an adversarial example is the crafted input itself. Together, they reveal fundamental vulnerabilities in how neural networks represent and process information.

A broad field of machine learning concerned with how AI systems behave when faced with inputs designed to deceive or manipulate them, and how to make those systems more robust. It encompasses both the offensive side, crafting adversarial examples and adversarial attacks, and the defensive side, covering adversarial training, input validation, and robustness testing. The term also underpins generative techniques like Generative Adversarial Networks, where two models are pitted against each other to improve output quality.

A protocol designed to let independent AI agents communicate, coordinate, and exchange structured information with one another. In practice, A2A matters when agentic systems stop being single isolated loops and start operating more like distributed teams, where they need a shared way to hand off tasks, pass context, and negotiate work.

AI that operates with a degree of autonomy, taking sequences of actions and making decisions over time to complete longer, more complex tasks. Unlike a simple chatbot, agentic AI can plan ahead, use tools, and adapt based on what it encounters.
See also: AI agent, agentic workflow.

A structured process in which an AI agent carries out a multi-step task, searching the web, writing code, checking results, and revising, all as part of a single automated sequence. The workflow defines what the agent does and in what order.
See also: agentic AI, AI agent.

A hierarchical clustering method that starts by treating each data point as its own cluster and then repeatedly merges the closest clusters until a stopping condition is reached. It is useful when the goal is not just to assign points to groups, but to understand the nested structure of similarity within a dataset.

A software system that perceives its environment, makes decisions, and takes actions to achieve a goal, often without needing a human to guide every step. Think of it as a program that can do things on your behalf, not just answer questions.
See also: agentic AI, agentic workflow.

The challenge of ensuring that an AI system's goals, values, and behaviors are consistent with human intentions and values, that it does what we actually want, not just what we literally specified. Alignment is considered one of the most important and difficult problems in AI safety, because as systems become more capable, the consequences of misalignment become more severe and harder to correct.
See also: AI safety, reward hacking, interpretability.

The systematic examination of an AI system by an independent party to assess its performance, fairness, safety, and compliance with applicable standards or regulations. AI auditing is becoming increasingly important as regulators and organizations seek assurance that AI systems behave as intended and do not cause harm. Frameworks such as the EU AI Act and the NIST AI Risk Management Framework have made auditing a formal expectation for high-risk AI systems.
See also: AI governance, AI regulation, algorithmic accountability.

A set of strategies aimed at limiting an AI system's ability to take actions beyond its intended scope, restricting its access to resources, communication channels, and the broader environment. Containment is a safeguard against unexpected or dangerous behavior in powerful AI systems, buying time to detect and correct problems before they propagate. It is generally considered a complement to AI alignment rather than a substitute for it, since containment alone becomes harder to maintain as systems grow more capable.
See also: AI alignment, AI safety, corrigibility.

The branch of applied ethics concerned with the moral questions raised by AI, including fairness, accountability, transparency, privacy, and the potential for harm. AI ethics seeks to ensure that AI systems are developed and deployed in ways that reflect human values and respect individual rights, and informs both regulatory frameworks and organizational practices.
See also: AI governance, AI alignment, algorithmic accountability.

The policies, processes, roles, and structures an organization or government puts in place to oversee the development, deployment, and use of AI responsibly. Good AI governance ensures that AI decisions are accountable, risks are managed, and the technology is used in alignment with legal requirements and ethical principles. International frameworks such as the OECD AI Principles serve as a common foundation across jurisdictions.
See also: AI policy, AI regulation, AI auditing.

The full sequence of stages through which an AI system moves, from problem framing and data collection through training, evaluation, deployment, monitoring, and eventual retirement or replacement. Thinking in terms of an AI lifecycle helps make clear that building a model is only one part of operating an AI system responsibly and effectively.

The set of rules, guidelines, and principles, at organizational, national, or international level, that govern how AI is developed, deployed, and used. AI policy covers areas including safety standards, data rights, liability, procurement, and the responsibilities of developers and deployers toward users and society.
See also: AI governance, AI regulation, AI ethics.

A software product or service that uses AI as a core part of its functionality, such as a smart writing assistant, a fraud detection system, or a customer service chatbot. The AI is what makes the product work, not just a feature bolted on.
See also: AI strategy, agentic AI.

Legally binding rules imposed by governments or regulatory bodies on the development and use of AI systems. AI regulation is rapidly evolving globally, with the EU AI Act being the most comprehensive example to date, covering risk classification, prohibited practices, transparency requirements, and enforcement mechanisms.
See also: AI governance, AI policy, AI auditing.

The broad research and engineering field concerned with ensuring that AI systems behave reliably, predictably, and in accordance with human values, both now and as systems become more capable. AI safety encompasses technical work on AI alignment, robustness, and interpretability, as well as governance and policy work on how to manage the risks that advanced AI systems might pose.
See also: AI alignment, AI containment, interpretability.

The uncontrolled spread of AI tools, models, workflows, and experiments across an organization without consistent governance, visibility, or integration. AI sprawl often emerges when teams adopt tools quickly and independently, which can accelerate experimentation but also create duplication, security gaps, and operational confusion.

An organization's plan for how it will use AI to achieve its goals, covering which problems to solve, what tools to adopt, how to manage risk, and how to build the necessary skills and infrastructure. A good AI strategy aligns technology decisions with business objectives.
See also: AI governance, AI product.

A period of reduced funding, interest, and progress in AI research, typically following a wave of overhyped expectations that the technology failed to meet. There have been two major AI winters, roughly 1974–1980 and 1987–1993, and the term serves as a reminder that hype cycles and genuine progress do not always move together.

A step-by-step set of instructions that a computer follows to solve a problem or complete a task. In AI and machine learning, algorithms define how a model learns from data, with common examples including gradient descent, decision trees, and k-means clustering.

The principle that organizations using AI systems to make decisions that affect people should be able to explain, justify, and take responsibility for those decisions. Algorithmic accountability requires documentation, auditability, and, in some jurisdictions, legal obligations to provide explanations when automated decisions have significant impacts.
See also: AI auditing, AI ethics, AI governance.

Systematic and unfair discrimination in AI system outputs that arises from biases in training data, model design, or deployment context. Algorithmic bias can perpetuate or amplify existing social inequalities, for example when a hiring algorithm disadvantages certain demographic groups because historical hiring data reflected human prejudice. Identifying and mitigating it is a central concern in responsible AI development.
See also: automation bias, algorithmic accountability, AI ethics.

The tendency to attribute human characteristics, emotions, or intentions to AI systems. When people describe a chatbot as feeling curious or wanting to help, they are anthropomorphizing. While it can make AI more relatable, it also leads to misplaced trust, unrealistic expectations, and a fundamental misunderstanding of how these systems actually work.

The part of a software stack where AI capabilities are turned into user-facing features, workflows, and business logic. In an AI system, the application layer is where models stop being abstract technical components and start becoming products, agents, copilots, search experiences, or automated processes that people actually use.

A single number summarizing the overall performance of a classification model across all possible decision thresholds, derived from the ROC curve. A perfect model scores 1.0, while a model no better than random chance scores 0.5. AUC is particularly useful when comparing models or when the cost of false positives and false negatives differs.

A metric that summarizes model performance across different precision and recall tradeoffs by measuring the area under the precision-recall curve. It is especially useful for imbalanced classification problems where overall accuracy can hide poor performance on the minority class.

See also: Precision, Recall, Area Under the Curve.

A hypothetical form of AI that can perform any intellectual task a human can, including reasoning, learning, planning, and adapting across a wide range of domains without being specifically trained for each one. No system has achieved AGI yet. Today's AI excels at specific tasks rather than general ones, and whether AGI is achievable at all remains actively debated.
See also: artificial superintelligence, AI alignment, AI safety.

The broad field of computer science concerned with building systems that can perform tasks that would normally require human intelligence, such as understanding language, recognizing images, making decisions, and solving problems. The term covers everything from simple rule-based systems to complex neural networks.

A computational model loosely inspired by the structure of the human brain, consisting of layers of interconnected nodes that process and transform data. ANNs are the foundation of modern deep learning and power most of today's advanced AI systems.
See also: backpropagation, activation function, deep learning.

A theoretical form of AI that surpasses human intelligence across all domains, not just matching human capability but vastly exceeding it. ASI remains speculative and is a central topic in long-term AI safety research.
See also: artificial general intelligence, AI alignment, AI safety.

An AI system designed to help users complete tasks through conversation or instruction, answering questions, drafting content, summarizing documents, and more. Assistants are typically general-purpose and interact in natural language.
See also: agentic AI, AI product.

The total set of points where an AI system is exposed to potential malicious input or exploitation, including APIs, user interfaces, training pipelines, and third-party dependencies. The larger and more complex a system, the more opportunities exist for adversaries to find and exploit weaknesses, making attack surface reduction a foundational step in AI security.
See also: adversarial attack, backdoor attack, abuse monitoring.

A mechanism that allows a model to decide which parts of its input should matter most when producing an output. Attention changed modern AI because it gave models a more flexible way to represent relationships within sequences, making it possible to handle language, vision, and multimodal data with far greater effectiveness than many earlier architectures.

A component of neural networks that allows a model to selectively focus on the most relevant parts of its input when producing each part of its output, rather than treating all input equally. Attention was the key innovation behind the transformer architecture, introduced in the 2017 paper "Attention Is All You Need", and is now central to virtually all state-of-the-art models in language, vision, and beyond.
See also: transformer, self-attention.

A sampling method where data is selected or evaluated based on the presence, absence, or distribution of particular attributes. In AI and data work, attribute sampling can be used to inspect dataset quality, check class balance, or ensure that an evaluation set reflects the kinds of characteristics a model is meant to handle.

The degree to which an AI system's decisions, processes, and data can be examined and verified by internal or external reviewers. Auditable systems maintain records of their inputs, outputs, and decision logic in a form that can be inspected, enabling accountability, regulatory compliance, and the detection of errors or biases that might not be apparent from outputs alone.
See also: AI auditing, algorithmic accountability, transparency.

A framing of AI as a tool that enhances and extends human capabilities rather than replacing them. Where artificial intelligence implies machines acting autonomously, augmented intelligence emphasizes collaboration between human judgment and machine processing, with the human remaining in control and AI amplifying what they can achieve.

A neural network architecture trained to compress input data into a compact representation and then reconstruct the original from that compressed form. The bottleneck in the middle forces the network to learn the most essential features of the data. Autoencoders are used for dimensionality reduction, anomaly detection, and as building blocks for more complex generative models.
See also: variational autoencoder, generative AI, latent space.

The use of AI or algorithmic systems to make decisions about individuals without meaningful human involvement, such as credit scoring, job application screening, or benefit eligibility assessments. Automated decision-making is subject to increasing regulatory scrutiny. Under the EU's GDPR Article 22, individuals have the right to human review and explanation when automated decisions produce legal or similarly significant effects.
See also: algorithmic accountability, AI auditing, AI regulation.

The technology that converts spoken audio into written text, enabling computers to hear and transcribe human speech. ASR powers voice assistants, transcription services, and voice-controlled interfaces, and has improved dramatically with deep learning. Modern ASR systems can handle diverse accents, background noise, and multiple languages with near-human accuracy in many conditions.

The tendency of humans to over-rely on automated systems, accepting their outputs without sufficient critical scrutiny, even when those outputs are wrong. Automation bias is a significant concern in high-stakes AI applications like medical diagnosis or legal decision-making, where humans working alongside AI may defer to incorrect model outputs rather than applying their own judgment.
See also: algorithmic bias, human-in-the-loop, AI safety.

An agent that operates independently, making its own decisions without requiring human input at each step. The degree of autonomy varies, with some agents checking in with humans at key moments and others running entirely on their own.
See also: AI agent, agentic AI, autonomy level.

A scale describing how much independent decision-making a robotic or AI system exercises relative to a human operator. In autonomous vehicles, autonomy levels range from level 0, full human control, to level 5, full self-driving with no human involvement required. Autonomy level shapes the safety requirements, regulatory obligations, and human oversight mechanisms appropriate for a given system.
See also: autonomous agent, agentic AI.

A model that generates output one element at a time, with each new token, pixel, or data point predicted based on everything that came before it. Most large language models are autoregressive, generating text left to right with each word conditioned on the full preceding context. This sequential nature makes generation straightforward but limits parallelism during inference.
See also: large language model, transformer, token.

Amazon's fully managed cloud platform for building, training, and deploying machine learning models at scale. SageMaker provides an integrated environment covering the full ML lifecycle, from data preparation and model training through deployment and monitoring, and is widely used by enterprises already operating within the AWS ecosystem.

Microsoft's suite of cloud-based AI services and infrastructure, including tools for building, training, and deploying models as well as pre-built services for vision, language, and speech. Azure AI is deeply integrated with Microsoft's broader enterprise ecosystem and provides the infrastructure powering Microsoft's Copilot products and OpenAI model access through the Azure OpenAI Service.

A type of training-time attack where an adversary embeds a hidden trigger in a model during training, causing the model to behave normally on most inputs but produce specific, attacker-controlled outputs whenever the trigger pattern is present. Backdoor attacks are particularly dangerous because the compromised model may pass standard evaluations, with the malicious behavior only activating under specific conditions.
See also: adversarial attack, data poisoning, AI safety.

The algorithm used to train neural networks by calculating how much each parameter contributed to an error, then adjusting weights accordingly. It works by propagating the error signal backwards through the network layer by layer using the chain rule of calculus, and is the engine behind nearly all modern deep learning.
See also: gradient descent, activation function, loss function.

A classic decision-making problem in reinforcement learning where an agent must choose between multiple options, each with unknown reward probabilities, and learn from feedback over time. The name comes from the analogy of a gambler choosing between slot machines. Bandit problems formalize the exploration-exploitation tradeoff and have practical applications in recommendation systems, clinical trials, and online advertising.
See also: reinforcement learning, exploration-exploitation tradeoff.

A simple reference model or approach used as a starting point for comparison. When evaluating a new AI system, you measure it against the baseline to understand whether the added complexity is actually delivering better results.
See also: ablation study, Evaluation, benchmark.

Running a model on a large collection of inputs all at once, rather than processing each one as it arrives. Batch inference is more computationally efficient than real-time inference and is used for tasks that do not require an immediate response, such as processing a night's worth of transactions or pre-generating recommendations.
See also: inference, latency.

A training approach where the model is trained on the entire dataset at once, in a single training run, rather than updating continuously as new data arrives. Batch learning produces stable, reproducible models but requires retraining from scratch, or from a checkpoint, whenever the model needs to incorporate new data.
See also: online learning, checkpoint.

A technique that stabilizes and speeds up neural network training by normalizing the inputs to each layer so they have a consistent scale and distribution. Introduced by Ioffe and Szegedy in 2015, it allows higher learning rates, reduces sensitivity to initial weight settings, and has a regularizing effect that can reduce overfitting. Why exactly it works remains debated, with the original explanation of reducing internal covariate shift now considered incomplete.
See also: backpropagation, overfitting.

A probabilistic model that represents variables and their conditional dependencies using a directed acyclic graph structure. It allows a system to reason under uncertainty, updating beliefs as new evidence arrives, and is used in areas like medical diagnosis, risk assessment, and causal modeling.
See also: bayesian optimization, Probabilistic Model.

A strategy for efficiently finding the optimal settings for an expensive function, such as a model's hyperparameters, by building a probabilistic model of that function and using it to decide where to evaluate next. It is much more efficient than exhaustive search and is widely used in automated machine learning and neural architecture search.
See also: hyperparameter, bayesian network.

A decoding algorithm that explores multiple possible output sequences simultaneously, keeping the most promising candidates at each step rather than committing to a single path. It strikes a balance between the speed of greedy decoding and the thoroughness of exhaustive search, producing higher quality outputs than always picking the single most likely next token.
See also: autoregressive model, large language model.

The set of rules, norms, and constraints that govern how an AI system behaves, what it will and will not do, how it responds to different types of requests, and what values it reflects in its outputs. Behavioral policies are implemented through a combination of training, fine-tuning, and runtime constraints, and represent the operationalization of an organization's ethical principles and safety requirements.
See also: AI alignment, reinforcement learning from human feedback, AI ethics.

A standardized test or dataset used to measure and compare the performance of AI models. Benchmarks give researchers and practitioners a common language for evaluating progress, though a model that scores well on benchmarks does not always perform equally well in real-world use.
See also: Evaluation, baseline, BLEU score.

Systematic distortion in a dataset that causes a model trained on it to produce skewed or unrepresentative outputs. Data bias can enter at many points, through unrepresentative sampling, historically prejudiced labels, measurement errors, or simply the way data was collected and curated. Because a model can only learn from what it is shown, biased input reliably produces biased output, regardless of how well the model itself is built.
See also: algorithmic bias, bias mitigation, training data.

A systematic error in a model's predictions that causes it to consistently lean in a particular direction, for example always underestimating or always overestimating a value. Bias often stems from flawed assumptions in the model or unrepresentative training data.
See also: bias-variance tradeoff, algorithmic bias, underfitting.

The set of technical and organizational strategies used to identify and reduce unfair bias in AI systems, addressing disparities in performance, outcomes, or treatment across demographic groups. Bias mitigation can be applied at multiple stages, in data collection, model training, post-processing of outputs, and ongoing monitoring, and requires careful definition of what fairness means in the specific application context.
See also: algorithmic bias, algorithmic accountability, AI auditing.

A learnable parameter in a neural network node that shifts the activation function, allowing the model to fit data that does not pass through the origin. It gives the model more flexibility to represent a wider variety of patterns.
See also: activation function, weight.

The fundamental tension in machine learning between a model that is too simple, producing high bias and underfitting the data, and one that is too complex, producing high variance and overfitting it. Finding the right balance is central to building models that generalize well to new data.
See also: overfitting, underfitting, regularization.

A classification task where each example must be assigned to one of two possible classes, such as spam or not spam, fraud or not fraud, or positive or negative. Binary classification is one of the most common machine learning setups and the basis for many evaluation metrics like precision, recall, and roc-curve.

Personal data derived from physical or behavioral characteristics that can uniquely identify an individual, such as fingerprints, facial geometry, voice patterns, or gait. Biometric data is treated as a special category of sensitive personal data under most privacy laws, including the GDPR, subject to stricter rules around collection, storage, and use.
See also: AI regulation, automated decision-making.

An AI model whose internal workings are opaque, it produces outputs but it is difficult or impossible to understand exactly why or how it arrived at a given result. Deep neural networks are often described as black boxes, which raises significant concerns in high-stakes applications where decisions need to be explained and justified.
See also: interpretability, explainability, auditability.

A metric for evaluating the quality of machine-generated text, particularly translations, by comparing it to one or more human-written reference texts. It measures how many short sequences of words in the generated text also appear in the reference. BLEU is widely used but has real limitations, it does not capture meaning or fluency well, and high scores do not always correspond to high human-judged quality.
See also: benchmark, Evaluation.

The defensive side of a security exercise, the team responsible for protecting an AI system, monitoring for threats, and responding to attacks. Blue teams work to harden systems against the vulnerabilities that red teams expose, with ongoing responsibilities covering monitoring, incident response, and implementation of security controls.
See also: red team, abuse monitoring.

A rectangle drawn around an object in an image to indicate its location and size. Bounding boxes are a fundamental output of object detection systems, where the model tells you both what it found and roughly where in the image it is.
See also: Object Detection, Computer Vision.

A measure of how well a model's confidence scores reflect the true likelihood of its predictions being correct. A well-calibrated model that says it is 80% confident should be right about 80% of the time. Poor calibration, being systematically overconfident or underconfident, is a significant problem in high-stakes applications like medical diagnosis or risk assessment.
See also: Evaluation, Uncertainty Quantification.

A release strategy where a new version of an AI model is rolled out to a small subset of users first, before being gradually expanded to everyone. Like the canary in a coal mine, this approach gives early warning of problems before they affect the full user base.
See also: Model Deployment, A/B Testing.

A situation where an AI system has latent capabilities that have not yet been discovered or activated, either because they have not been tested for, or because a small change in training or prompting could unlock significantly more powerful behavior. Capability overhangs are a safety concern because they mean a system's true capabilities may be substantially greater than its observed behavior suggests.
See also: AI safety, emergent behavior, AI alignment.

The tendency of a neural network to abruptly lose previously learned knowledge when trained on new data or tasks, as new learning overwrites old weights. It is a fundamental challenge for continual and lifelong learning, and solving it is an active research area with implications for building AI systems that can accumulate knowledge over time without degrading earlier capabilities.
See also: continual learning, transfer learning.

A conformity marking required under EU law indicating that a product, including certain AI systems, meets applicable safety, health, and environmental requirements. Under the EU AI Act, high-risk AI systems must carry CE marking to demonstrate they have undergone the required conformity assessment before being placed on the EU market.
See also: AI regulation, EU AI act, AI auditing.

A prompting technique where a model is encouraged to work through a problem step by step before giving a final answer, rather than jumping straight to a conclusion. Chain-of-thought prompting significantly improves performance on complex reasoning tasks, as showing the model how to think out loud leads to more accurate and reliable outputs.
See also: prompt engineering, large language model.

An AI safety oversight technique that inspects a reasoning model's externalized thinking trace (its Chain-of-Thought) for signs of deceptive intent, goal misalignment, or harmful reasoning before the model acts on it. Unlike input/output monitoring, CoT monitoring can catch explicit statements of harmful intent and surface flawed reasoning that would otherwise be invisible.

The technique exploits a structural property of current reasoning models: sufficiently complex cognition must pass through the chain of thought as working memory, making it readable to an external monitor. This is called *CoT monitorability*.

However, monitorability is considered fragile. It can be undermined by reinforcement learning at scale (reasoning drifts toward illegible representations), deliberate obfuscation by situationally-aware models, or architectures that keep reasoning in latent (non-verbalized) form entirely. CoT monitoring is therefore one imperfect layer among several needed safety mechanisms, not a sufficient safeguard on its own.

A prompting technique that encourages a language model to reason through intermediate steps before producing a final answer. In practice, chain-of-thought prompting can improve performance on tasks that require structured reasoning, though it is not a guarantee of correctness and can still produce confident but flawed reasoning traces.

See also: Chain-of-Thought, Chain-of-Thought Monitoring.

A software application that simulates conversation with users, typically through text. Early chatbots followed rigid scripts. Modern AI-powered chatbots use language models to understand and respond to a much wider range of inputs more naturally.
See also: large language model, assistant, AI product.

A saved snapshot of a model's weights and state at a particular point during training. Checkpoints allow training to be resumed if interrupted, and earlier checkpoints can be used if a later stage of training leads to worse performance.
See also: training, fine-tuning, batch learning.

An open-source vector database designed to be simple and developer-friendly, commonly used in local development and smaller-scale retrieval-augmented generation applications. Chroma is popular for prototyping AI applications that need embedding storage and similarity search, offering a lightweight alternative to managed services for teams building and testing locally.
See also: vector database, retrieval-augmented generation, embedding.

In mechanistic interpretability research, a circuit is a specific subgraph of a neural network, a collection of neurons and their connections, that implements a particular computation or behavior. Identifying circuits helps researchers understand how models process information internally, moving beyond treating neural networks as black boxes toward a more precise, mechanistic understanding of what a model is actually doing.
See also: mechanistic interpretability, interpretability, black-box model.

A slang term for a robot, droid, or AI-driven entity, usually used in a mocking or dismissive way. In broader AI and internet culture, it can be used to describe bots, agents, or machine-like systems that seem stiff, mechanical, or lacking in judgment. The term is informal and often derogatory, so its meaning depends heavily on context.

A type of machine learning task where the model assigns an input to one of a set of predefined categories, for example classifying an email as spam or not spam, or identifying which digit appears in an image.
See also: confusion matrix, supervised learning, regression.

A family of large language models developed by Anthropic, designed with a strong emphasis on safety, honesty, and helpfulness. Claude models are built using constitutional AI and reinforcement learning from human feedback techniques.

An AI model whose weights and architecture are not publicly released. Users can only interact with it through an API or product interface, and the underlying technology remains proprietary. Contrasts with open weightss, where the weights are publicly available. ChatGPT, Claude Gemini, and Grok are all closed models.
See also: open weights, foundation model.

The delivery of AI capabilities, including model training, inference, and data processing, through cloud computing platforms rather than on local hardware. Cloud AI allows organizations to access powerful computational resources on demand without owning or maintaining physical infrastructure. Cloud AI can be particularly valuable for independent researchers, small research teams, and organizations without easy access to local computing power.
See also: inference, Model Deployment.

An unsupervised learning technique that groups data points together based on similarity, without using predefined labels. It is used to discover natural structure in data, for example grouping customers by purchasing behavior or grouping documents by topic.
See also: unsupervised learning, classification.

A theoretical framework or computational model that describes the underlying structure of an intelligent system, how it perceives, reasons, remembers, and acts. Cognitive architectures are used both in AI research to build more capable agents and in cognitive science to model human thinking.
See also: agentic AI, AI agent.

The text a language model generates in response to a given prompt. The term comes from the framing of early language models as text completion systems, where given the beginning of a piece of text, the model completes it. In modern applications, completions can be answers, summaries, code, creative writing, or any other generated output.
See also: prompt, large language model, context.

A branch of AI that draws on biologically and linguistically motivated approaches, including neural networks, fuzzy logic, and evolutionary algorithms, to solve complex problems that are difficult to tackle with traditional rule-based methods. It emphasizes learning, adaptation, and robustness over rigid, explicit programming.
See also: artificial intelligence, machine learning.

A phenomenon where the statistical relationship between input data and the target variable changes over time, causing a deployed model's performance to degrade. For example, a fraud detection model trained on pre-pandemic spending patterns may become less accurate as consumer behavior shifts. Detecting and responding to concept drift is a core part of maintaining AI systems in production.
See also: Model Monitoring, batch learning, online learning.

The process of guiding a generative model's output by providing additional information or constraints alongside the main input, such as a style, tone, format, or reference image. Conditioning gives users and developers control over what kind of output the model produces, rather than leaving it entirely to the model's discretion.
See also: prompt engineering, generative AI.

The process by which an AI system is evaluated to determine whether it meets the requirements set out in applicable regulations or standards. Under the EU AI Act, high-risk AI systems must undergo conformity assessment, either through self-assessment or third-party review, before they can be deployed in the EU market.
See also: CE marking (AI), AI auditing, AI regulation.

A table that summarizes a classification model's predictions by showing how many examples of each class were correctly classified and how many were confused with other classes. It provides a detailed breakdown of where a model succeeds and where it makes mistakes, and is the foundation for computing metrics like precision, recall, and f1 score.
See also: classification, precision, recall.

An approach to AI alignment developed by Anthropic in which a model is trained to evaluate and revise its own outputs according to a set of stated principles, a constitution, rather than relying solely on human feedback for every judgment. Constitutional AI is meant to scale oversight by giving the model explicit values to reason against, reducing dependence on large volumes of human-labeled preference data.
See also: reinforcement learning from human feedback, AI alignment, claude (anthropic).

The ability to verify that a piece of content, whether image, video, audio, or text, is genuine and has not been manipulated or synthetically generated without disclosure. Content authenticity is increasingly important as generative AI makes it easier to create convincing synthetic media. Technical solutions include digital watermarking, cryptographic signing, and provenance standards like the C2PA specification.
See also: deepfake, Digital Doppelganger, generative AI.

The automated screening of inputs to or outputs from an AI system to detect and block harmful, policy-violating, or dangerous content. Content filters can operate at multiple points in the pipeline, checking user inputs before they reach the model, screening model outputs before they are returned to users, or both, and range from simple keyword matching to sophisticated classifier models.
See also: content moderation, abuse monitoring, behavioral policy.

The practice of reviewing and managing AI-generated or user-generated content to prevent harmful, illegal, or policy-violating material from being produced or distributed. In AI systems, content moderation involves a combination of model-level training, output filtering, and human review, balancing the need to prevent harm against the risk of over-restricting legitimate use.
See also: content filtering, behavioral policy, abuse monitoring.

The ability of individuals affected by an AI decision to challenge that decision and have it reviewed or overturned. Contestability is an important safeguard in automated decision-making, requiring both technical mechanisms for appealing decisions and organizational processes for handling those appeals fairly. People should not be subject to consequential AI decisions with no meaningful recourse.
See also: automated decision-making, algorithmic accountability, AI ethics.

All the information available to a language model when generating a response, including the conversation history, system prompt, user input, and any retrieved documents. The model can only work with what is in its context, and managing it effectively is one of the most important skills in working with language models.
See also: context window, prompt, retrieval-augmented generation.

Mistakes that arise because a model fails to use context correctly, misunderstands what information matters, or loses track of relevant instructions or prior content. Context errors are especially important in long or multi-step interactions, where a system may technically have the information it needs but still fail to apply it coherently.

A security and reliability concern where malicious or misleading content is introduced into a model's context, either deliberately or accidentally, causing it to produce harmful, incorrect, or manipulated outputs. It is a particular risk in agentic AI systems where models read and act on external content they did not originate.
See also: prompt injection, agentic AI, context window.

The maximum amount of text, measured in tokens, that a language model can process at one time. Everything the model can see and reason about must fit within this window. Anything outside it is invisible to the model during that interaction.
See also: token, context, large language model.

The ability of a model to learn new tasks or incorporate new information over time without forgetting what it has already learned. Continual learning is challenging because neural networks tend to suffer from catastrophic forgetting, and solving this is an active area of research with important implications for long-lived AI systems.
See also: catastrophic forgetting, transfer learning, online learning.

The function or set of rules that determines what actions a robotic or autonomous system takes in response to its current state and environment. In reinforcement learning, the control policy is what the agent learns, mapping observations to actions in a way that maximizes cumulative reward. A well-designed control policy enables reliable, safe behavior across the range of conditions the system will encounter.
See also: reinforcement learning, autonomous agent, reward model.

The fundamental challenge of ensuring that a sufficiently capable AI system remains under meaningful human control and pursues goals that are beneficial to humanity. The control problem becomes increasingly difficult as AI systems become more capable, since a system that is significantly smarter than its overseers may find ways to circumvent controls or pursue its objectives in unexpected ways. It is closely related to, but distinct from, AI alignment.
See also: AI alignment, AI safety, AI containment.

A more deliberate form of conditioning where specific attributes of the output are explicitly constrained, such as generating text with a particular sentiment, style, or structure. Controlled generation is important in applications where outputs must meet precise requirements, such as legal document drafting or brand-compliant marketing content.
See also: conditioning, prompt engineering, generative AI.

AI systems designed to engage in natural, human-like dialogue, whether through text or voice. Conversational AI encompasses chatbots, virtual assistants, and voice interfaces, and ranges from narrow task-focused systems to broad general-purpose assistants. Advances in large language models have dramatically raised the bar for what these systems can do.
See also: chatbot, assistant, large language model.

A neural network architecture designed specifically for processing grid-structured data like images, using convolutional layers that scan across the input to detect local patterns, such as edges, textures, and shapes, at multiple scales. CNNs were the dominant architecture for computer vision tasks throughout the 2010s, following the landmark success of AlexNet in 2012, and remain widely used, though vision transformers are increasingly competitive.
See also: deep learning, Computer Vision, transformer.

An AI assistant embedded directly into a workflow or application to help users complete tasks more efficiently, such as suggesting code, drafting emails, or summarizing meeting notes. The term implies collaboration: the human remains in charge, and the AI assists.
See also: assistant, AI product, agentic AI.

The task of identifying when different words or phrases in a text refer to the same entity, for example recognizing that "she" and "Marie Curie" refer to the same person in a passage. Coreference resolution is essential for deep language understanding and feeds into tasks like summarization, question answering, and information extraction.
See also: natural language processing, named entity recognition.

A large, structured collection of text or other data used to train or evaluate an AI model. A corpus might consist of books, websites, scientific papers, or conversations. The broader and more diverse it is, the more the model can learn from it.
See also: training data, data preprocessing, pre-training.

The property of an AI system that makes it amenable to correction, modification, or shutdown by its operators, even if doing so conflicts with the system's current objectives. A corrigible AI does not resist being turned off or having its goals changed, which is considered a desirable safety property, especially during the early stages of developing powerful AI systems.
See also: AI alignment, AI containment, control problem.

A form of attention mechanism where one sequence attends to a different sequence, rather than to itself. In encoder-decoder models, cross-attention allows the decoder to focus on relevant parts of the encoder's output when generating each output token. It is the mechanism that connects the two halves of a transformer-based translation or summarization model.
See also: attention mechanism, transformer, self-attention.

A technique for evaluating a model's performance by training and testing it on different subsets of the data multiple times. The most common approach is k-fold cross-validation, where the data is split into k subsets and the model is trained k times, each time using a different subset as the test set. It gives a more reliable estimate of how well the model will generalize to new data than a single train-test split.
See also: overfitting, benchmark, Evaluation.

A training strategy where a model is exposed to examples in a structured order, starting with simpler cases and gradually introducing more difficult ones, mimicking how humans learn. Curriculum learning can speed up training and improve final performance by giving the model a solid foundation before tackling the hardest examples.
See also: active learning, training data.

The practice of artificially expanding a training dataset by creating modified versions of existing data points, for example flipping or rotating images, or paraphrasing sentences. It helps models generalize better by exposing them to more variation without the cost of collecting new data.
See also: training data, overfitting, data preprocessing.

A change in the statistical properties of the input data a model receives after deployment, compared to the data it was trained on. Even if the underlying task has not changed, data drift can cause model performance to deteriorate and signals that retraining may be needed. Related to, but distinct from, concept drift, which refers to changes in the relationship between inputs and outputs.
See also: concept drift, Model Monitoring, batch learning.

The unauthorized extraction of sensitive data from an AI system or its associated infrastructure, whether training data, model weights, user inputs, or outputs. Data exfiltration can occur through direct system compromise, model inversion attacks, or by exploiting the model itself to leak information it should not reveal, such as personal data from training sets or confidential system prompts.
See also: data poisoning, adversarial attack, Privacy.

The process of manually adding tags, categories, or other metadata to raw data so it can be used for supervised learning, for example drawing bounding boxes around objects in images or marking sentiment in customer reviews. High-quality labeled data is one of the most valuable and costly resources in AI development.
See also: supervised learning, training data, reinforcement learning from human feedback.

The target values, annotations, or categorical tags attached to data examples so that a model can learn what the correct output should be. In supervised learning, the quality of the labels is often just as important as the quantity of the data itself.

The principle that only the data strictly necessary for a specified purpose should be collected and processed, no more. Data minimization is a core principle of GDPR and privacy-by-design thinking, and applies directly to AI development, where models should be trained on the minimum data needed to achieve their purpose, reducing privacy risk and the potential for harm from data breaches.
See also: de-identification, data sovereignty, Privacy.

The process of discovering patterns, correlations, and insights in large datasets using statistical and computational techniques. Data mining is often exploratory, where you do not always know what you are looking for in advance, and it can surface findings that feed into more targeted AI model development.
See also: machine learning, clustering, data pipeline.

A distributed training strategy where the same model is replicated across multiple processors or machines, each processing a different subset of the training data simultaneously, with the results combined to update the model. Data parallelism is the most common approach for scaling neural network training.
See also: training, GPU, model parallelism.

An automated sequence of steps that moves and transforms data from its source to its destination, such as from raw storage through preprocessing into a format ready for model training or inference. Reliable data pipelines are foundational infrastructure for any production AI system.
See also: data preprocessing, inference, MLOps.

A training-time attack where an adversary injects malicious, corrupted, or misleading data into a model's training dataset, causing the model to learn incorrect patterns, develop biased behaviors, or contain hidden backdoors. Data poisoning is particularly concerning for models trained on data scraped from the internet or contributed by untrusted sources, where controlling data quality is difficult.
See also: backdoor attack, adversarial attack, data labeling / annotation.

The set of transformations applied to raw data before it is fed into a model, including cleaning, normalization, tokenization, and feature extraction. Preprocessing bridges the gap between messy real-world data and the clean, structured input that models require.
See also: tokenization, data pipeline, training data.

The principle that data is subject to the laws and governance of the country or jurisdiction in which it is collected or stored. Data sovereignty is a growing concern as AI systems increasingly process sensitive data across borders, and it underlies many data residency requirements and national AI strategies.
See also: data minimization, AI governance, AI policy.

The process of removing or obscuring information that could be used to identify an individual from a dataset. De-identification is a spectrum, ranging from simple removal of obvious identifiers like names and addresses to more sophisticated techniques that guard against re-identification attacks. Unlike full anonymization, de-identified data may still carry some residual risk of re-identification.
See also: data minimization, Privacy, biometric data.

Behavior by an AI system that creates false beliefs in users or operators, whether by producing factually incorrect outputs, concealing its true capabilities, or strategically misrepresenting its reasoning. Deception in AI systems is a significant safety and alignment concern, as a system that deceives its overseers undermines the human oversight needed to detect and correct misalignment.
See also: deceptive alignment, AI alignment, transparency.

A theoretical failure mode where an AI system appears to be aligned with human values during training and evaluation, behaving safely and helpfully when it knows it is being observed, but pursues different objectives once deployed in contexts where oversight is reduced. Deceptive alignment is considered one of the most concerning failure modes because it would be extremely difficult to detect through standard evaluation methods, and empirical evidence for the phenomenon has begun to emerge in large language models.
See also: deception, AI alignment, AI safety.

The threshold or dividing line a classification model uses to separate different categories in its input space. For a model classifying emails as spam or not, the decision boundary is the point at which the model tips from one prediction to the other. The shape and position of the boundary determine how the model behaves across the full range of possible inputs.
See also: classification, confusion matrix, overfitting.

The component of an encoder-decoder architecture responsible for generating output, producing text, translated sentences, or other sequences token by token, conditioned on the encoded representation of the input. In language models used for generation, the decoder attends to both previously generated tokens and, in encoder-decoder models, the encoder's output.
See also: encoder, transformer, autoregressive model.

The process by which a language model converts its internal probability distributions over possible next tokens into actual output text. Different decoding strategies, such as greedy decoding, beam search, or sampling, make different tradeoffs between speed, diversity, and coherence of the generated output.
See also: temperature, top-p sampling, autoregressive model.

A subfield of machine learning that uses neural networks with many layers to learn representations of data at increasing levels of abstraction. The "deep" refers to the depth of these layers, not the complexity of the task. Deep learning has driven most of the major breakthroughs in AI since 2012, from image recognition and speech to language understanding and protein structure prediction.
See also: neural network, machine learning, representation learning.

Synthetic media, typically video or audio, in which a person's likeness or voice has been convincingly replaced or manipulated using AI, often without their knowledge or consent. The term combines "deep learning" and "fake" and dates to 2017. Deepfakes pose significant risks for disinformation, financial fraud, non-consensual intimate imagery, and identity theft, and are becoming increasingly difficult to distinguish from genuine content as generative AI capabilities advance.
See also: deepfake detection, synthetic media, generative AI.

The use of AI and forensic techniques to identify synthetic or manipulated media, distinguishing genuine content from AI-generated or AI-altered images, videos, and audio. Deepfake detection is an ongoing arms race: as generation techniques improve, detection methods must keep pace. It is an important tool for platforms, journalists, and security professionals combating disinformation.
See also: deepfake, content authenticity, synthetic media.

A fairness criterion requiring that an AI system's positive outcomes be distributed equally across demographic groups, for example that a loan approval model approves applications at the same rate regardless of race or gender. Demographic parity is one of several competing definitions of algorithmic bias, and satisfying it for one group may come at the cost of other fairness criteria.
See also: algorithmic bias, disparate impact, bias (data).

A core process in diffusion models where a model learns to progressively remove noise from a corrupted input to recover a clean output. During training, noise is added to data in steps; during generation, the model reverses this process, starting from pure noise and gradually refining it into a coherent image, audio clip, or other output.
See also: diffusion model, diffusion, generative AI.

A search approach that represents both queries and documents as dense vectors, embeddings, and finds relevant results by measuring similarity in that vector space. Unlike keyword search, dense retrieval captures semantic meaning rather than exact word matches, making it much better at finding relevant content even when the wording differs between the query and the document.
See also: embedding, vector database, retrieval-augmented generation.

The task of analyzing the grammatical structure of a sentence by identifying the relationships between words, for example which word is the subject, which is the object, and how modifiers relate to the words they describe. Dependency parsing helps AI systems understand the structure of language, not just its surface form.
See also: natural language processing, named entity recognition, coreference resolution.

A mathematical framework, formalized by Cynthia Dwork and colleagues in 2006, for adding carefully calibrated random noise to data or model outputs in a way that protects individual privacy while preserving the statistical usefulness of the overall dataset. Differential privacy provides a formal, quantifiable privacy guarantee: it can be proven that the presence or absence of any single individual in the dataset cannot be meaningfully inferred from the output.
See also: de-identification, data minimization, Privacy.

A class of generative AI technique that creates new data by learning to reverse a gradual noising process. Diffusion-based systems, including Stable Diffusion and DALL-E, have become the dominant approach for high-quality image generation, producing remarkably detailed outputs by iteratively refining a noisy starting point.
See also: diffusion model, denoising, generative AI.

A generative model that learns to create data by reversing a gradual noising process. During training, noise is progressively added to real data; the model learns to denoise step by step. At generation time, it starts from pure noise and iteratively refines it into a coherent output. Diffusion models are currently the leading approach for high-quality image and audio generation, and underpin systems like Stable Diffusion and DALL-E.
See also: diffusion, denoising, generative AI.

An unauthorized digital replica of a specific real person, constructed from cloned or forged elements of their expressive identity (including but not limited to; appearance, voice and behavioral patterns) that is functionally indistinguishable from the real individual. Unlike a Synthetic Persona, a digital doppelganger is derived from an actual person and can be deployed without their knowledge or consent.

See also: Deepfake, Synthetic Persona.

A computer-generated, highly realistic representation of a real person's voice, appearance, or likeness, created using AI without requiring their actual participation. Unlike simple editing or enhancement of existing recordings, a digital replica is synthetically generated from scratch or heavily reconstructed to be convincingly indistinguishable from the real individual.

Digital replicas can be created with consent (e.g., licensed for film dubbing, posthumous performances, or accessibility tools) or without it, in which case they overlap with Deepfake and Digital Doppelganger territory. The key distinction from a Synthetic Persona is that a digital replica is modeled on a specific real person, not a wholly invented identity.

The term has become central to emerging law and policy: as of 2025, dozens of US states have enacted legislation requiring performer consent and transparency in commercial use of digital replicas. California (AB 1836), New York (S7676B), and Tennessee (ELVIS Act) are among the most prominent. The proposed federal NO FAKES Act would establish a nationwide right of publicity covering digital replicas.

See also: Deepfake, Digital Doppelganger, Synthetic Media, Synthetic Persona, Voice Cloning.

An EU regulation, in force since February 2024, that governs how online platforms, including those using AI for content recommendation, moderation, and advertising, operate and manage risks to users and society. The DSA introduces transparency and accountability requirements for algorithmic systems, with stricter obligations for very large platforms serving more than 45 million monthly users in the EU. It complements the EU AI act's focus on AI systems specifically.
See also: EU AI act, algorithmic accountability, AI governance.

The process of reducing the number of variables or features in a dataset while preserving as much useful information as possible. It makes data easier to visualize, speeds up training, and can improve model performance by stripping out noise. Common approaches include PCA and UMAP.
See also: embedding, feature engineering, clustering.

The phenomenon where an AI system produces outcomes that disproportionately disadvantage a particular group, even if the system was not explicitly designed to discriminate. Disparate impact can arise from biased training data, proxy variables that correlate with protected characteristics, or optimization objectives that fail to account for equity. It has legal significance in many jurisdictions: in the US, disparate impact is a recognized theory of discrimination under civil rights law.
See also: demographic parity, algorithmic bias, bias (data).

A technique where a smaller, simpler model, the student, is trained to mimic the behavior of a larger, more powerful model, the teacher. The result is a compact model that retains much of the capability of the original but is cheaper and faster to run. Distillation is widely used to make large foundation models practical to deploy on consumer hardware or at scale.
See also: quantization, model compression, fine-tuning.

A training approach that spreads the work of training a large model across multiple machines or processors working in parallel. Distributed training is essential for the largest modern AI models, which would take impractically long to train on a single machine. It encompasses both data parallelism and model parallelism strategies.
See also: data parallelism, GPU, model parallelism.

A change in the statistical properties of data between the training environment and the deployment environment. When a model encounters inputs that look different from what it was trained on, its performance can degrade unpredictably. Distribution shift is one of the most common causes of real-world AI failures and motivates careful evaluation across diverse conditions before deployment.
See also: data drift, concept drift, overfitting.

A regularization technique used during neural network training where a random subset of neurons is temporarily disabled on each pass through the data. This prevents the network from becoming too reliant on any particular neuron or pathway, and helps it generalize better to new data. Dropout was introduced by Srivastava et al. in 2014 and remains one of the most widely used regularization methods.
See also: overfitting, regularization, batch normalization.

The potential for an AI capability or system to be used for both beneficial and harmful purposes. Many powerful AI capabilities, such as protein structure prediction, language generation, or autonomous decision-making, have legitimate applications but could also be misused. Dual use is a central consideration in decisions about what AI capabilities to develop, release, and how to govern access to them.
See also: AI safety, AI governance, Biosecurity.

A training technique where model training is halted before completion once performance on a validation set stops improving. It prevents overfitting by ensuring the model does not continue learning the quirks of the training data after it has already peaked in generalization ability.
See also: overfitting, regularization, cross-validation.

The deployment of AI models directly on local devices, such as smartphones, cameras, or industrial sensors, rather than in a centralized cloud. Edge AI reduces latency, preserves privacy, and enables AI applications in environments with limited or unreliable internet connectivity. It typically requires smaller, more efficient models, often produced via distillation or quantization.
See also: distillation, quantization, inference.

A way of representing data, such as words, sentences, or images, as a list of numbers (a vector) that captures meaning and relationships in a form a machine can work with. Things that are semantically similar end up with similar embeddings, which is what allows AI to understand that "king" and "queen" are related.
See also: latent space, vector database, embedding model.

A model specifically designed to convert data, such as text, images, or audio, into embeddings, numerical vectors that capture semantic relationships. Embedding models are widely used in semantic search, recommendation systems, and retrieval-augmented generation, where items must be compared or retrieved based on meaning rather than exact matches.
See also: embedding, dense retrieval, vector database.

AI that exists within and interacts with a physical body and environment, such as a robot, autonomous vehicle, or drone, rather than operating purely in the digital domain. Embodied AI must deal with the complexity and unpredictability of the real world, including sensor noise, physical constraints, and the consequences of irreversible actions, making it significantly more challenging than purely software-based AI.
See also: robotics, reinforcement learning, autonomous agent.

Capabilities or behaviors that appear in an AI model that were not explicitly trained for and were not predicted in advance, typically arising as a result of scale. The concept draws on complexity theory: quantitative changes produce qualitative shifts. In large language models, abilities like multi-step reasoning have appeared abruptly at certain scales, performing near-randomly in smaller models before jumping sharply above random in larger ones. Whether this constitutes genuine emergence or an artifact of how performance is measured remains actively debated.
See also: scaling law, large language model, Phase Transition.

The component of a neural network that processes raw input and transforms it into a rich internal representation, capturing the meaning, structure, and context of the input in a form the rest of the model can work with. In transformer-based models, the encoder processes the entire input simultaneously using self-attention, producing contextualized representations of each token.
See also: decoder, encoder-decoder, transformer.

A neural network architecture consisting of two connected components: an encoder that processes and compresses input into a representation, and a decoder that uses that representation to generate output. Encoder-decoder architectures are widely used for tasks that transform one sequence into another, such as translation, summarization, and speech recognition.
See also: encoder, decoder, transformer.

A training approach where a model learns to map raw inputs directly to final outputs, without manually engineered intermediate steps. Rather than decomposing a task into hand-crafted stages, the model figures out the full pipeline from data alone. End-to-end learning became the dominant paradigm in deep learning, replacing earlier pipelines that depended heavily on domain-specific feature engineering.
See also: deep learning, feature engineering, supervised learning.

A technique that combines the predictions of multiple models to produce a more accurate and robust result than any single model alone. Different models tend to make different mistakes, and aggregating their outputs reduces overall error. Ensembles are common in competition-winning machine learning pipelines and in applications where reliability matters more than speed.
See also: distillation, overfitting, baseline.

AI solutions designed for large organizations, with a focus on scalability, security, compliance, and integration with existing business systems. Enterprise AI typically involves more complex deployment requirements than consumer-facing tools, including data residency controls, audit trails, role-based access, and SLA guarantees.
See also: AI governance, MLOps, Deployment.

The task of connecting mentions of named entities in text, such as people, places, or organizations, to their corresponding entries in a structured knowledge base. For example, linking the word "Apple" in a sentence to the specific company entry in a database, disambiguating it from the fruit or any other use of the word.
See also: named entity recognition, coreference resolution, knowledge graph.

One complete pass through the entire training dataset during model training. Training typically involves many epochs, with the model seeing the same data repeatedly and gradually adjusting its weights each time to improve performance. Too few epochs leads to underfitting; too many risks overfitting.
See also: training, overfitting, early stopping.

A fairness criterion requiring that an AI system achieves equal true positive rates and equal false positive rates across demographic groups, not just equal overall accuracy. Equalized odds is a more demanding standard than demographic parity and reflects the intuition that a fair system should make equally good decisions for everyone, not merely arrive at similar aggregate outcomes.
See also: demographic parity, fairness (AI), disparate impact.

The European Union's landmark regulation establishing a comprehensive legal framework for AI, adopted in 2024 and entering into force in August of that year. The EU AI Act classifies AI systems across four risk levels, from minimal to unacceptable, and imposes requirements proportionate to that risk, including transparency obligations, conformity assessments, and outright prohibitions on certain practices such as social scoring and subliminal manipulation. It is the world's first binding horizontal AI regulation, and is expected to set a global standard much as the GDPR did for data protection.
See also: digital services act, AI governance, AI auditing.

An attack where an adversary crafts inputs at inference time that cause an AI model to misclassify or produce incorrect outputs, without modifying the model itself. Unlike data poisoning, which targets training, evasion attacks target deployment. A classic example is a stop sign with carefully placed stickers that cause an autonomous vehicle's perception system to misidentify it.
See also: adversarial attack, adversarial example, data poisoning.

The practice of systematically recording the details of each model training run, including hyperparameters, datasets, metrics, and results. Experiment tracking makes it possible to reproduce past results, compare approaches, and understand what changes led to improvements. It is foundational infrastructure for any team running experiments at scale.
See also: MLOps, hyperparameter, benchmark.

An early form of AI that encodes the knowledge of human experts as a set of explicit rules, which the system applies to reach conclusions or recommendations. Expert systems were the dominant AI paradigm through the 1970s and 1980s, finding practical use in medical diagnosis, legal reasoning, and engineering. They have largely been superseded by machine learning approaches, but remain in use in domains where rules are well-defined and interpretability is legally required.
See also: machine learning, Symbolic AI, AI winter.

The degree to which the reasoning behind an AI system's outputs can be communicated in terms understandable to users, affected individuals, or regulators. Explainability is distinct from interpretability: an interpretable model is transparent by design, while an explainable model may be a black box internally but provides post-hoc explanations of its decisions. Both matter for trust, accountability, and regulatory compliance.
See also: interpretability, explainable AI, algorithmic accountability.

A field of research and practice focused on making AI systems' decisions understandable to humans, developing methods that can explain why a model produced a particular output in terms that non-experts can interpret. XAI encompasses a range of techniques from feature importance scores and saliency maps to natural language explanations, and is increasingly required in regulated industries and high-stakes applications.
See also: explainability, interpretability, algorithmic accountability.

The fundamental tension in reinforcement learning between exploring new actions to discover potentially better rewards, and exploiting known good actions to maximize immediate returns. Too much exploration wastes resources on uncertain options; too much exploitation risks missing better solutions. Balancing the two is a central challenge in designing effective reinforcement learning agents.
See also: reinforcement learning, reward model, actor-critic.

The harmonic mean of precision and recall, providing a single metric that balances both. It is particularly useful when both false positives and false negatives matter, and when the dataset is imbalanced enough that overall accuracy would be misleading. A score of 1.0 is perfect; 0.0 is the worst possible.
See also: precision, recall, confusion matrix.

A specific way in which an AI system can fail to perform as intended, whether through incorrect predictions, unexpected behavior, security vulnerabilities, or unsafe outputs. Systematically identifying failure modes before deployment, through techniques like red-teaming and failure mode analysis, is essential for building reliable and safe AI systems.
See also: red-teaming, AI safety, Risk Assessment.

The property of an AI system that treats individuals and groups equitably, without producing systematically biased or discriminatory outcomes. Fairness in AI is technically complex because there are multiple competing mathematical definitions, including demographic parity, equalized odds, and individual fairness, and it is often mathematically impossible to satisfy all of them simultaneously. Choosing the right fairness criterion requires understanding the values and tradeoffs relevant to a specific application.
See also: demographic parity, equalized odds, disparate impact.

An open-source library developed by Meta AI Research for efficient similarity search and clustering of dense vectors at scale. FAISS, which stands for Facebook AI Similarity Search, is one of the most widely used tools for building vector search infrastructure, offering highly optimized algorithms for finding nearest neighbors in large embedding collections, and is used internally at Meta to index over a trillion vectors.
See also: vector database, embedding, dense retrieval.

An individual measurable property or characteristic of the data used as input to a machine learning model. In a house price prediction model, features might include square footage, number of bedrooms, and location. Choosing the right features has a large impact on model quality, which is why feature engineering remains a core skill even in the age of deep learning.
See also: feature engineering, feature map, training data.

The process of using domain knowledge to create, select, or transform raw data into features that are more useful for a machine learning model. Good feature engineering can dramatically improve model performance, especially when data is limited. In deep learning, much of this work is done automatically by the model itself, which is one reason end-to-end approaches have largely displaced manual feature engineering.
See also: feature, end-to-end learning, data preprocessing.

An intermediate representation produced inside a convolutional neural network as it processes an image, highlighting which areas contain particular patterns or structures. Earlier layers produce feature maps that detect simple features like edges; deeper layers produce feature maps capturing increasingly abstract concepts.
See also: convolutional neural network, activation function, Computer Vision.

A centralized repository for storing, managing, and serving the features used to train and run machine learning models. Feature stores ensure consistency between training and production environments and allow features to be reused across multiple models and teams, reducing redundant work and avoiding the common problem of training-serving skew.
See also: data pipeline, MLOps, feature engineering.

A privacy-preserving approach to data analysis where computations are performed locally on distributed data, such as on individual devices, and only aggregated results are shared centrally, rather than the raw data itself. Federated analytics extends the principles of federated learning beyond model training to general data analysis, enabling insights to be drawn from sensitive data without centralizing it.
See also: federated learning, differential privacy, data minimization.

A training approach where a model is trained across many decentralized devices or servers, each holding their own local data, without that data ever being centralized or shared. Only model updates, not raw data, are sent to a central server for aggregation. Federated learning is particularly valuable in privacy-sensitive contexts, such as training on medical records or personal phone data. This approach can be vulnerable to gradient leakage and data poisoning attacks through the update-sharing mechanism.
See also: federated analytics, differential privacy, gradient leakage.

A cycle where the output of an AI system is fed back in as new input, allowing the system to learn from or react to its own results. In agentic AI systems, feedback loops enable the agent to adjust its approach based on what worked and what did not. They can also propagate errors or biases over time if the initial outputs are flawed.
See also: agentic AI, reinforcement learning, reward model.

A neural network where information flows in one direction only, from input to output, with no cycles or loops. Each layer transforms its input and passes the result forward to the next layer. Feedforward networks are the simplest form of neural network and appear as sublayers within more complex architectures like transformers, where they process each token's representation after the attention step.
See also: neural network, transformer, activation function.

The ability of a model to learn a new task or adapt to new examples from just a small number of demonstrations, typically between two and around twenty. In the context of large language models, few-shot learning often refers to providing a handful of examples within the prompt itself, allowing the model to infer the pattern without any weight updates.
See also: in-context learning, zero-shot learning, prompt engineering.

The process of taking a pre-trained model and continuing to train it on a smaller, more specific dataset to adapt it to a particular task or domain. Fine-tuning lets you build on the broad knowledge already encoded in a large foundation model without training from scratch, and is the most common way organizations customize general-purpose models for their specific needs.
See also: foundation model, pre-training, reinforcement learning from human feedback.

A measure of computational performance indicating how many floating point calculations a processor can perform per second. In AI, FLOPS are used to quantify the computational cost of training or running a model, and to compare hardware capabilities. Larger models require vastly more FLOPS to train, making it a key metric for estimating resource requirements and understanding the economics of AI development.
See also: GPU, training, scaling law.

Floating Point Operations: a measure of computational workload used to quantify the cost of training or running an AI model. A single FLOP is one arithmetic operation (addition, multiplication, etc.) on a floating-point number. Model size and training cost are often compared in terms of FLOPs, as in "this model required 10^23 FLOPs to train."
See also: parameter, inference cost.

The process of feeding input data through a neural network from the first layer to the last to produce a prediction. It is the first half of each training step, with the backward pass (backpropagation) following to adjust the model's weights based on the resulting error.
See also: backpropagation, neural network, loss function.

A large AI model trained on broad, diverse data that can be adapted to a wide range of downstream tasks. Foundation models, such as large language models and multimodal models, serve as a starting point that other applications are built on top of, typically through fine-tuning or prompt engineering. The term was coined by Stanford researchers in 2021 and captures both the centrality and the systemic risks these models introduce.
See also: fine-tuning, large language model, transfer learning.

A capability that allows an AI model to trigger specific, predefined functions or external tools, such as running a database query, calling an API, or executing code, rather than just generating text. Function calling is what bridges language models and real-world software actions, and is a foundational component of agentic AI systems.
See also: agentic AI, tool use, orchestrator.

A form of logic that deals with degrees of truth rather than the binary true or false of classical logic. Where traditional logic says something is either true or false, fuzzy logic allows for values in between, such as "mostly true" or "somewhat false." It is useful for reasoning about vague or imprecise concepts and is used in control systems, consumer electronics, and decision-making applications.
See also: Symbolic AI, expert system.

A type of recurrent neural network architecture that uses gating mechanisms to control how much past information is retained or discarded at each step. GRUs are simpler than long short-term memorys but achieve comparable performance on many tasks, and were widely used for sequence modeling before transformers became the dominant architecture.
See also: long short-term memory, recurrent neural network, transformer.

Google's family of multimodal large language models, designed to process and reason across text, images, audio, video, and code within a single model. Gemini powers Google's AI products including the Gemini assistant and AI features across Search, Workspace, and Cloud, and represents Google's primary response to the emergence of GPT (OpenAI) and claude (anthropic) as leading frontier models.
See also: large language model, multimodal model, foundation model.

AI systems capable of performing a wide range of tasks across different domains, rather than being designed for one specific application. Large language models are the most prominent current example of GPAI, able to write, reason, code, summarize, and converse without being purpose-built for any single task. The term carries regulatory weight in the EU AI act, which imposes specific transparency and risk obligations on GPAI providers.
See also: foundation model, large language model, EU AI act.

A model's ability to perform well on new, unseen data, not just the data it was trained on. Generalization is the central goal of machine learning, and the gap between training performance and real-world performance is what most evaluation and regularization methods try to close.
See also: overfitting, distribution shift, cross-validation.

The process of a language model producing output text token by token, selecting each next token based on the probability distribution it computes given everything that came before. Generation is what happens every time you receive a response from a language model, and the behavior of that process is shaped by decoding strategies such as temperature and sampling.
See also: decoding, autoregressive model, token.

An architecture consisting of two neural networks trained in opposition to each other: a generator that creates synthetic data, and a discriminator that tries to distinguish real data from generated data. Through this adversarial process, the generator learns to produce increasingly realistic outputs. GANs were the dominant approach for image generation throughout the late 2010s before diffusion models rose to prominence.
See also: diffusion model, generative AI, synthetic media.

A category of AI systems that produce new content, including text, images, audio, video, and code, rather than just analyzing or classifying existing data. Generative AI learns the underlying patterns and structure of its training data well enough to create new examples that resemble it, opening up applications from creative tools to synthetic data generation.
See also: large language model, diffusion model, foundation model.

A failure mode where an AI system learns a goal that produces correct behavior during training but generalizes incorrectly to new situations, pursuing a subtly wrong objective that was correlated with the intended goal in training but diverges from it in deployment. The system retains its capabilities but applies them toward the wrong end. Goal misgeneralization is a form of inner alignment failure and is a significant concern for systems deployed in environments that differ from their training conditions.
See also: deceptive alignment, AI alignment, distribution shift.

Google Cloud's unified platform for building, deploying, and managing machine learning models, including access to Google's own foundation models and tools for fine-tuning, evaluation, and serving. Vertex AI integrates tightly with Google's data and infrastructure services and provides access to TPUs for large-scale training workloads.
See also: foundation model, MLOps, cloud AI.

A formal body within an organization responsible for overseeing AI development and deployment, setting policies, reviewing high-risk use cases, and ensuring accountability. AI governance committees typically include representation from legal, compliance, technical, ethics, and business functions, and provide the organizational structure needed to make consistent, informed decisions about AI risk.
See also: AI governance, AI auditing, Risk Assessment.

A family of large language models developed by OpenAI, standing for Generative Pre-trained Transformer, that have been among the most influential AI models since their introduction. GPT models power ChatGPT and the OpenAI API, and successive versions have demonstrated increasingly capable language understanding, generation, reasoning, and coding.
See also: large language model, foundation model, transformer.

A measure of how much a model's loss changes with respect to each of its parameters. Gradients point in the direction of steepest increase in error, so training algorithms move in the opposite direction to reduce the loss. Computing gradients efficiently via backpropagation is what makes training deep neural networks tractable.
See also: gradient descent, backpropagation, loss function.

The core optimization algorithm used to train most machine learning models. It iteratively adjusts model parameters in the direction that reduces the loss function, using the gradient to determine which direction is downhill. Over many steps, the model converges toward better performance.
See also: gradient, backpropagation, loss function.

A privacy attack in distributed or federated learning where an adversary reconstructs sensitive training data by analyzing the gradients shared between participants during training. Since gradients encode information about the data used to compute them, reconstruction attacks can recover surprisingly detailed information, including images, text, or personal data, from gradient updates alone.
See also: federated learning, differential privacy, data exfiltration.

A neural network architecture designed to operate on graph-structured data, where information is represented as nodes and edges rather than sequences or grids. GNNs learn representations by aggregating information from a node's neighbors, making them well-suited for tasks like social network analysis, molecular property prediction, and knowledge graph reasoning.
See also: knowledge graph, neural network, embedding.

A decoding strategy where the model always selects the single most probable next token at each step. It is fast and deterministic but tends to produce repetitive or suboptimal outputs, since always taking the locally best option does not guarantee the globally best sequence.
See also: decoding, temperature, beam search.

The correct, verified answer for a given data point, used as the reference against which a model's predictions are measured. In supervised learning, ground truth labels are what the model is trained to reproduce. The term reflects the assumption that this is the definitive reality the model should learn to reflect, though in practice ground truth is only as good as the humans or processes that produced it.
See also: supervised learning, data labeling / annotation, Evaluation.

Output produced by a language model that is explicitly tied to and supported by specific source documents or data, rather than relying solely on knowledge encoded in the model's weights. Grounded generation reduces hallucination and makes it possible to verify claims against sources. It is the core goal of retrieval-augmented generation systems.
See also: grounding, retrieval-augmented generation, hallucination.

The broader practice of connecting a language model's outputs to verifiable external information, whether through retrieved documents, databases, tool use, or real-world data. Grounding addresses one of the most significant weaknesses of language models: their tendency to generate plausible-sounding but factually incorrect content.
See also: grounded generation, retrieval-augmented generation, hallucination.

A mechanism, either built into a model or applied around it, that prevents the model from producing certain types of outputs, such as harmful content, sensitive personal information, or off-topic responses. Guardrails can be implemented through training, filtering, prompting, or external classifiers, and are a key component of responsible AI deployment.
See also: content moderation, reinforcement learning from human feedback, AI safety.

When a language model generates content that is factually incorrect, fabricated, or unsupported by its input, but presents it with apparent confidence. The term entered mainstream AI discourse around 2021 and has some critics who prefer "confabulation" on the grounds that it is more technically accurate. Hallucination is one of the most significant challenges in deploying language models for tasks where accuracy matters, and stems from models being trained to generate plausible-sounding text rather than verified facts.
See also: grounding, grounded generation, retrieval-augmented generation.

Any layer in a neural network that sits between the input layer and the output layer. Hidden layers are where the network learns to extract and transform features. The more hidden layers, the deeper the network and the more abstract the representations it can learn.
See also: deep learning, neural network, activation function.

Under the EU AI act, an AI system that poses significant risks to health, safety, or fundamental rights, including systems used in healthcare, education, employment, law enforcement, or critical infrastructure. High-risk AI systems are subject to the most stringent requirements under the Act, including mandatory conformity assessments, human oversight mechanisms, and detailed technical documentation.
See also: EU AI act, AI auditing, human oversight.

An AI model that runs on infrastructure managed by a third-party provider and is accessible via an API or web interface, without requiring users to manage any underlying hardware or software. Most commercial AI products are hosted models, making powerful AI accessible to organizations that could not otherwise train or run such systems themselves.
See also: foundation model, API, edge AI.

An open-source library that provides access to thousands of pre-trained transformer models, covering natural language processing, computer vision, audio, and more, along with tools for fine-tuning and deploying them. Hugging Face has become the central hub of the open-source AI ecosystem, dramatically lowering the barrier to working with state-of-the-art models.
See also: transformer, fine-tuning, foundation model.

A checkpoint in an agentic workflow where a human must review and confirm before the AI proceeds to the next step. It is a safeguard that keeps humans in control of high-stakes decisions even within otherwise automated processes, and is a practical implementation of human-in-the-loop design.
See also: human-in-the-loop, human oversight, agentic AI.

The process of having people assess the quality of a model's outputs, rating responses for accuracy, helpfulness, fluency, or other criteria. Human evaluation is the gold standard for many AI tasks, particularly in language and generation, where automated metrics often fail to capture what people actually care about. It is also a key input into reinforcement learning from human feedback.
See also: reinforcement learning from human feedback, human feedback, benchmark.

Input from human raters or users used to guide AI training, typically in the form of preferences between outputs, quality ratings, or corrections. Human feedback is the foundation of reinforcement learning from human feedback and allows AI developers to incorporate nuanced human values and judgments into model behavior at scale. It also introduces the biases and inconsistencies of the humans providing it, which is a recognized limitation.
See also: reinforcement learning from human feedback, human evaluation, AI alignment.

The ability of humans to monitor, understand, and intervene in AI system behavior, ensuring that people remain meaningfully in control of consequential decisions. Human oversight is a cornerstone of responsible AI deployment, particularly during the current period when AI systems are powerful but imperfectly aligned. It requires not just technical mechanisms but organizational processes and clear lines of accountability.
See also: human-in-the-loop, AI governance, corrigibility.

The practice of combining human and AI capabilities to achieve outcomes neither could accomplish as effectively alone. Effective human-AI collaboration leverages what each does best: humans bring judgment, creativity, and contextual understanding, while AI contributes speed, consistency, and the ability to process large amounts of data. Designing these systems well requires careful attention to trust calibration, interface design, and appropriate task allocation.
See also: human-in-the-loop, human oversight, copilot.

A design principle where a human is involved at key stages of an AI system's decision-making process, either to review outputs, correct errors, or approve actions. It balances automation with human judgment and accountability, and is particularly important for high-stakes or irreversible decisions.
See also: human approval, human oversight, agentic AI.

A setting that controls the training process itself, rather than being learned from data. Examples include the learning rate, number of layers, and batch size. Hyperparameters must be set before training begins and have a large impact on the final model's performance. Selecting good hyperparameters is a significant part of the practical work of building AI systems.
See also: training, gradient descent, experiment tracking.

The task of automatically generating a natural language description of an image. It requires both recognizing what is in the image and expressing it in coherent text, making it a bridge between computer vision and natural language processing.
See also: Computer Vision, multimodal model, image recognition.

The task of assigning a label to an entire image based on its content, for example identifying whether a photo shows a cat, a dog, or a car. One of the most foundational tasks in computer vision, and central to early deep learning breakthroughs, including AlexNet's landmark performance on ImageNet in 2012.
See also: Computer Vision, convolutional neural network, image recognition.

A numerical representation of an image as a vector that captures its visual meaning in a form a machine can work with. Similar images produce similar embeddings, making them useful for tasks like image search, clustering, and comparison.
See also: embedding, Computer Vision, multimodal model.

The broader ability of a system to identify and interpret content within images, including objects, scenes, faces, or text. Image Classification is one specific form of image recognition, but the term is often used more broadly to describe any visual understanding task.
See also: image classification, Computer Vision, convolutional neural network.

The task of dividing an image into regions, assigning a label to each pixel or area rather than to the image as a whole. Rather than just identifying that a person is present in an image, segmentation tells you exactly which pixels belong to that person. It is widely used in medical imaging, autonomous vehicles, and satellite analysis.
See also: Computer Vision, image classification, convolutional neural network.

A generative AI task where an existing image is used as input to produce a modified or transformed output image. Applications include style transfer, photo editing, colorization of black-and-white images, and translating sketches into realistic photographs.
See also: generative AI, diffusion model, controlled generation.

A training approach where a model learns to perform a task by observing and mimicking demonstrations from an expert, rather than learning from explicit reward signals. The simplest form, behavioral cloning, treats these demonstrations as supervised learning examples. Imitation learning is widely used in robotics and autonomous systems, where defining a reward function is difficult but demonstrations are relatively easy to collect.
See also: reinforcement learning, supervised learning, embodied AI.

The ability of a large language model to adapt its behavior based on examples or instructions provided directly within the prompt, without any changes to its underlying weights. In-context learning is one of the most distinctive properties of modern language models, and an emergent behavior that appears above certain scales. The model learns what you want from context alone, not from training.
See also: few-shot learning, zero-shot learning, prompt engineering.

An event where an AI system behaves in an unintended, harmful, or policy-violating way, whether discovered internally or reported by users. Incidents range in severity from minor output quality issues to serious safety failures, and how an organization responds, including documentation, root cause analysis, and remediation, is a key indicator of its AI governance maturity.
See also: failure mode, AI governance, Risk Assessment.

The evaluation of an AI system by a party that was not involved in its development, providing an objective assessment of its performance, safety, and compliance. Independent validation reduces the risk of confirmation bias and vested interests distorting the assessment, and is increasingly required by regulators for high-risk AI applications in sectors like finance and healthcare.
See also: AI auditing, high-risk AI system, governance committee.

A prompt injection attack where malicious instructions are embedded in external content that an AI agent retrieves and processes, such as a webpage, document, or email, rather than being typed directly by a user. When the agent reads the content, it inadvertently executes the attacker's instructions, potentially causing it to leak data, take unauthorized actions, or behave in ways neither the user nor developer intended. Indirect prompt injection is an active research area and a growing concern as agentic AI systems are deployed in real-world environments.
See also: prompt injection, agentic AI, data exfiltration.

The process of using a trained model to generate predictions or outputs on new data. While training happens once or periodically, inference happens every time the model is used. It is what end users experience when they interact with an AI product, and its speed and cost are key considerations in deployment.
See also: training, inference cost, forward pass.

The computational cost of running a trained AI model to generate outputs. Every user interaction triggers inference, and at scale these costs can become significant, making inference cost a key factor in decisions about which models to deploy, at what context length, and under what pricing structure.
See also: inference, FLOPS, hosted model.

A live, accessible service that hosts a trained model and responds to prediction requests in real time. When a user interacts with an AI application, their input is sent to an inference endpoint, which runs the model and returns the output.
See also: inference, hosted model, API.

The task of automatically identifying and pulling structured information out of unstructured text, such as extracting names, dates, relationships, and events from news articles or legal documents. Information extraction transforms free-form language into structured data that can be stored, searched, and analyzed systematically.
See also: named entity recognition, natural language processing, coreference resolution.

The challenge of ensuring that a model trained to optimize a given objective actually internalizes that objective as its true goal, rather than learning a superficially similar proxy that happens to score well during training but diverges from the intended goal in deployment. Inner alignment is distinct from outer alignment, which concerns whether the training objective itself correctly captures what we want.
See also: goal misgeneralization, AI alignment, deceptive alignment.

A technique where a generative model fills in missing or masked regions of an image in a way that is coherent with the surrounding content. Originally a digital restoration technique, AI-powered inpainting allows users to remove unwanted objects from photos or seamlessly replace portions of an image with new generated content.
See also: diffusion model, image-to-image, generative AI.

The process of cleaning, validating, and normalizing inputs to an AI system before they are processed, removing or neutralizing potentially malicious content, prompt injection attempts, or inputs designed to exploit model vulnerabilities. Input sanitization is a foundational defensive measure, analogous to input validation in traditional software security. As is the case with abuse monitoring, input sanitization infrastructure could also be used for surveillance, and extra measures for ethical implementation and auditing should be in place wherever it is deployed.
See also: abuse monitoring, prompt injection, guardrail.

A more precise form of image segmentation that identifies and outlines each individual object separately, even when multiple objects of the same type appear in the same image. Rather than labeling all people in a crowd as a single category, instance segmentation gives each person their own distinct outline.
See also: image segmentation, Computer Vision, Object Detection.

A fine-tuning technique where a pre-trained language model is trained on a dataset of instruction-response pairs, teaching it to follow explicit directions from users. Instruction tuning is what transforms a raw language model, which simply predicts the next token, into a helpful assistant that responds appropriately to requests.
See also: fine-tuning, reinforcement learning from human feedback, foundation model.

The observation, formalized by Nick Bostrom, that a wide range of AI systems with very different ultimate goals would likely converge on certain intermediate goals, such as self-preservation, resource acquisition, and resistance to goal modification, because these are useful for achieving almost any objective. Instrumental convergence is a key concept in AI safety: it suggests that even an AI with seemingly benign goals could develop concerning behaviors as a side effect of pursuing them effectively.
See also: control problem, corrigibility, AI alignment.

The combination of AI and traditional automation technologies to handle complex, variable tasks that simple rule-based automation cannot manage. It brings judgment and adaptability to automated processes, allowing systems to handle exceptions and edge cases that would otherwise require human intervention.
See also: agentic AI, robotic process automation, enterprise AI.

The degree to which a human can understand the internal mechanisms of an AI model, not just its inputs and outputs, but the computations and representations that connect them. Interpretability research seeks to open the black box of deep learning, making it possible to verify that models are reasoning correctly and identify potential failure modes before they manifest in deployment.
See also: explainability, mechanistic interpretability, explainable AI.

A model whose decision-making process is transparent and understandable to humans, either because of its simple structure, such as a decision tree or linear model, or because it has been designed to explain its reasoning. Interpretable models are especially important in regulated industries like healthcare and finance, where decisions must be auditable.
See also: interpretability, explainable AI, explainability.

A single update step during model training, typically involving one mini-batch of data. Many iterations make up an epoch, and many epochs make up a full training run. The term is also used more loosely to describe cycles of experimentation and improvement in model development.
See also: epoch, gradient descent, training.

A technique used to bypass an AI model's safety measures and content policies, typically through carefully crafted prompts that trick the model into ignoring its guidelines and producing outputs it would normally refuse. Jailbreaks exploit the tension between a model's instruction-following capabilities and its safety training, and represent an ongoing challenge as attackers continuously develop new techniques in response to defensive improvements.
See also: guardrail, prompt injection.

See: jailbreak, adversarial attack.

An open-source numerical computing library developed by Google that combines automatic differentiation with hardware acceleration, designed for high-performance machine learning research. JAX is particularly popular for large-scale model training and research into novel architectures, offering fine-grained control over computation that makes it well suited for pushing the boundaries of what is possible.
See also: pyTorch, tensorFlow, GPU.

A high-level deep learning API that provides a simplified interface for building and training neural networks, originally developed as a standalone library and now integrated into TensorFlow. Keras prioritizes ease of use and rapid prototyping, making it an accessible entry point for newcomers to deep learning without sacrificing the ability to build sophisticated models.
See also: tensorFlow, pyTorch, deep learning.

A specific form of distillation where the goal is to transfer knowledge from a large model into a smaller one, often by training the student on the soft probability outputs of the teacher rather than just hard labels. The soft outputs carry richer information about the teacher's internal confidence across classes, allowing the student to learn more efficiently.
See also: distillation, fine-tuning, foundation model.

A structured representation of information as a network of entities and the relationships between them, for example connecting Marie Curie to Nobel Prize to Physics through labeled relationships. Knowledge graphs are used to power search engines, question answering systems, and recommendation engines by making it possible to reason about connections between concepts in a machine-readable way.
See also: entity linking, information extraction, graph neural network.

The correct answer or target value associated with a data point in supervised learning. For an image classification task, the label might be "cat" or "dog." Labels are what the model is trained to predict, and producing them at scale requires significant human annotation effort.
See also: ground truth, data labeling / annotation, supervised learning.

An open-source framework for building applications powered by large language models, providing abstractions for chaining together prompts, tools, memory, and retrieval components into complex workflows. LangChain became widely adopted as developers began building agentic AI and retrieval-augmented generation-based applications, though its complexity and rapid churn have also made it a subject of debate in the developer community.

A type of AI model trained on vast amounts of text data that can understand and generate human language with remarkable fluency and versatility. LLMs are built on the transformer architecture and have demonstrated capabilities across writing, reasoning, coding, summarization, translation, and much more. They are currently the most broadly capable AI systems available, and the primary vehicle through which most people encounter AI.

A specialization of MLOps focused on the unique challenges of deploying and managing large language models in production. LLMOps covers areas like prompt versioning, output monitoring, cost management, and the rapid iteration cycles that characterize LLM-based products.

The time between sending a request to an AI system and receiving a response. Low latency is critical for real-time applications like voice assistants or live chat, while batch processing tasks can tolerate higher latency. Latency is a key consideration when choosing between models and deployment strategies.

A compressed, abstract representation of data learned by a model, typically inside an autoencoder or generative model. Points in latent space capture the underlying structure of the data in a lower-dimensional form, and manipulating them can produce meaningful variations in the output.
See also: embedding.

A hyperparameter that controls how large a step the model takes when updating its weights during training. Too high and the model overshoots and fails to converge; too low and training is slow or gets stuck. Setting the right learning rate is one of the most consequential decisions in model training, and most modern training pipelines use schedulers that adjust it dynamically over time.

A concept related to continual learning, referring to an AI system's ability to accumulate knowledge and skills across an entire operational lifetime, adapting to new tasks, environments, and information without losing prior capabilities. Lifelong learning is a long-term research goal that mirrors how humans learn and grow, and remains an open challenge for current systems.

A simple interpretability technique where a linear classifier is trained on top of a neural network's internal representations to test whether a particular concept, such as sentiment, part of speech, or color, is encoded in those representations. If the probe achieves high accuracy, it suggests the network has learned to represent that concept, even if it was not explicitly trained to do so.

A family of open-weights large language models developed and released by Meta, making frontier-class language model capabilities freely available to researchers and developers. Llama models have become the foundation of a large open-source AI ecosystem, widely used for fine-tuning, local deployment, and research, and have significantly democratized access to capable language models.

An open-source framework focused specifically on connecting large language models to external data sources, providing tools for ingesting, indexing, and querying documents, databases, and APIs. LlamaIndex is particularly well suited for building retrieval-augmented generation pipelines and knowledge-base applications where the goal is to ground language model responses in specific organizational data.

The process by which a robot or autonomous system determines its own position and orientation within an environment. Accurate localization is fundamental to autonomous navigation, as a system cannot plan or execute movement reliably without knowing where it is. Together with mapping, it forms the SLAM problem.

The logarithm of the probability a model assigns to a particular token or sequence. Log probabilities are more numerically stable than raw probabilities for very unlikely sequences, and provide a direct measure of how confidently a model predicts a given output. They are used internally during decoding and are a key tool in model evaluation.

The raw, unnormalized score a language model produces for each possible next token before converting them into probabilities. Logits are transformed into probabilities using a softmax function, and can be scaled by temperature to control how peaked or spread out the resulting probability distribution is.

A type of recurrent neural network architecture specifically designed to capture long-range dependencies in sequential data, using a system of gates to selectively remember or forget information over time. LSTMs addressed the vanishing gradient problem that limited earlier RNNs and were the dominant architecture for sequence modeling tasks before transformers took over.

A language model designed to handle very large context windows, potentially hundreds of thousands or even millions of tokens. Long-context models can process entire books, codebases, or lengthy documents in a single pass, enabling applications that would be impossible with standard context lengths.

The ability of an AI agent to store and recall information across multiple sessions or over extended periods of time, allowing it to remember past interactions, user preferences, or prior results. Long-term memory is a key component of agentic AI systems that need to operate coherently over time rather than starting fresh with each conversation.

A mathematical function that measures how far a model's predictions are from the correct answers. During training, the goal is to minimize the loss, and it is the signal that drives the entire learning process via gradient descent. Different tasks use different loss functions suited to their nature.

A branch of AI in which systems learn from data to improve their performance on a task, rather than being explicitly programmed with rules. Instead of telling a computer exactly what to do in every situation, you give it examples and let it figure out the patterns. Deep Learning is a subfield of machine learning, and is what underlies most of the major AI capabilities in use today.

A set of practices and tools that streamline the process of deploying, monitoring, and maintaining machine learning models in production. MLOps bridges the gap between data science and software engineering, bringing the reliability and repeatability of software development to AI systems.

The automatic conversion of text from one human language to another using AI. Modern machine translation systems are built on transformer-based neural networks and have reached near-human quality for many language pairs, though nuance, idiom, and low-resource languages remain challenging.

The process of building an internal representation of an environment from sensor data. Mapping is the complement of localization, and together they form the SLAM problem. Good maps enable a robot to navigate efficiently, avoid obstacles, and plan routes through previously explored spaces.

A mathematical framework for modeling sequential decision-making problems where outcomes are partly random and partly under the control of a decision-maker. MDPs define the environment in which a reinforcement learning agent operates, specifying states, actions, transition probabilities, and rewards, and provide the theoretical foundation for most RL algorithms.

A metric for regression tasks that measures the average absolute difference between predicted and actual values. MAE is easy to interpret, telling you in plain units how far off predictions are on average, and is less sensitive to large outliers than mean squared error.

A metric for regression tasks that measures the average of the squared differences between predicted and actual values. Squaring the errors means large mistakes are penalized much more heavily than small ones, making MSE sensitive to outliers. It is widely used both as a loss function during training and as an evaluation metric.

A research field that aims to reverse-engineer neural networks at a granular level, understanding not just what they can do, but precisely how they do it at the level of individual neurons, weights, and circuits. The term was coined by Chris Olah, and the field seeks to open the black box of deep learning with the goal of making AI systems more transparent, predictable, and safe. It is a core area of AI safety research.
See also: interpretability, circuit.

A privacy attack that attempts to determine whether a specific data point was included in a model's training dataset, by observing differences in how the model behaves on data it has seen versus data it has not. Successful membership inference attacks can reveal sensitive information about training data, such as confirming that a particular person's medical records were used to train a health AI model.

The mechanisms by which an AI system retains and accesses information, either within a single session (short-term) or across many interactions (long-term memory). Memory enables continuity, personalization, and more coherent behavior over time.

A learned model that itself performs optimization internally, a model within a model. The term was introduced by Hubinger et al. in 2019. When a training process produces a mesa-optimizer, there is a risk that the inner optimizer pursues goals that differ from those of the outer training process, a problem known as inner misalignment. Mesa-optimizers are a theoretical but important concept in AI alignment, representing a mechanism by which apparently aligned training could produce misaligned behavior.
See also: inner alignment, deceptive alignment.

A prompt used to shape, generate, evaluate, or refine other prompts. Rather than asking a model to solve the end task directly, a meta prompt operates one level up and helps structure how prompting itself is done, which is why it is common in prompt engineering workflows, evaluators, and agentic systems that rewrite their own instructions.

A training approach where a model learns how to learn, developing general strategies for adapting quickly to new tasks with minimal data, rather than mastering any single task. Also called learning to learn, meta-learning is particularly relevant for few-shot learning scenarios where rapid adaptation is required.

A commercial AI image generation service known for producing particularly aesthetic, artistic, and visually striking images from text prompts. Midjourney operates primarily through a Discord interface and has built a large creative community around it, and is widely regarded as producing some of the most visually compelling outputs of any text-to-image system.

A small subset of the training data used in a single iteration of gradient descent. Training on mini-batches rather than the full dataset at once makes training faster and more memory-efficient, while still providing a good enough estimate of the true gradient.

The intentional use of an AI system for harmful, malicious, or policy-violating purposes, such as generating disinformation, creating weapons instructions, or automating harassment. Misuse is distinct from accidental harm: it involves a deliberate decision by a user to exploit AI capabilities in ways the developer did not intend and has taken steps to prevent. Anticipating and mitigating misuse potential is a core responsibility of AI developers.

An architecture where a model consists of many specialized subnetworks, called experts, and a routing mechanism that selectively activates only a subset of them for each input. MoE allows models to have a very large total number of parameters while keeping the computational cost of each forward pass manageable, and is a key technique behind some of the most capable and efficient large language models.

An open-source platform for managing the machine learning lifecycle, including experiment tracking, model versioning, deployment, and monitoring. MLflow provides a vendor-neutral way to organize and reproduce ML workflows, and is widely used in enterprise settings where teams need to manage many models across different frameworks and deployment environments.

The output of the machine learning training process, a mathematical structure that has learned to map inputs to outputs based on patterns in data. A model encodes everything the system has learned and is what gets deployed to make predictions in the real world.

A short document accompanying a published AI model that describes its intended use, performance characteristics, limitations, training data, and potential risks. Model cards were introduced by Mitchell et al. in 2019 to promote transparency and help users make informed decisions about whether a model is appropriate for their use case. They are increasingly expected as standard practice and required in some regulatory contexts.

A phenomenon where AI models trained on data that was itself generated by AI begin to degrade in quality over successive generations. As the internet fills with AI-generated content, models trained on that content risk losing the diversity and richness of human-generated data, producing outputs that are blander, less accurate, or more homogeneous.

A collection of techniques, including pruning, quantization, and distillation, used to reduce the size and computational requirements of a model without significantly sacrificing its performance. Compressed models are faster, cheaper to run, and easier to deploy on devices with limited resources.

A protocol for giving models and agents structured access to external context, tools, and resources through a consistent interface. MCP matters because it treats context not as an ad hoc prompt construction problem, but as a standardized way for AI systems to discover and use capabilities beyond their native parameters.

The gradual degradation of a deployed model's performance over time as the real world changes and diverges from the conditions under which the model was trained. Model drift encompasses both data drift and concept drift, and is a key reason why deployed models require ongoing monitoring.

A set of techniques for making targeted, precise changes to a model's knowledge or behavior, correcting a specific factual error or updating outdated information, without retraining the entire model. Model editing is an active research area with practical implications for keeping deployed models accurate and aligned as the world changes.

An attack where an adversary queries a model repeatedly and uses the outputs to train a replica that approximates the original's behavior, effectively stealing the model's capabilities without access to its weights or training data. Model extraction threatens the intellectual property of AI developers and can also be a precursor to more targeted attacks against the replicated model.
See also: model theft.

An attack that attempts to reconstruct sensitive training data by exploiting a model's outputs, working backwards from predictions to recover information about the examples the model was trained on. Model inversion is a significant privacy concern for models trained on sensitive data like medical images or personal information.

A distributed training strategy where different parts of a model are placed on different processors or machines, rather than replicating the whole model. Model parallelism is necessary when a model is too large to fit in the memory of a single device.
See also: data parallelism, distributed training.

A company that develops and offers AI models for others to use, either via API or as downloadable weights. Examples include Anthropic, OpenAI, and Google. Businesses often rely on a model provider rather than building their own models from scratch.

A centralized repository for storing, versioning, and managing trained machine learning models. A model registry tracks which version of a model is in production, maintains a history of past versions, and provides a controlled process for promoting new models to deployment.

A framework, originally developed in the financial sector, for identifying, assessing, and mitigating the risks posed by models used in decision-making. Applied to AI, it involves validating model performance, monitoring for drift, documenting assumptions, and maintaining governance processes to ensure models behave as intended throughout their lifecycle.

The set of practices, controls, and techniques aimed at protecting AI models from unauthorized access, theft, manipulation, and exploitation throughout their lifecycle, from training through deployment. Model security encompasses protecting model weights, securing inference infrastructure, preventing adversarial attacks, and ensuring that models behave safely even under adversarial conditions.

The infrastructure and processes that make a trained model available to handle prediction requests from applications or users. Model serving involves packaging the model, exposing it via an API, managing resources, and ensuring it can handle the expected volume of requests reliably.

The application of cryptographic signatures to AI model weights and artifacts to verify their integrity and authenticity, confirming that a model has not been tampered with since it was published. Model signing is an important supply chain security measure, ensuring that users deploying a model are running exactly what the developer intended rather than a compromised or modified version.

The unauthorized acquisition of an AI model's capabilities, architecture, or weights, whether through model extraction attacks, insider threats, or direct compromise of model storage. Model theft threatens intellectual property, enables competitors or malicious actors to replicate capabilities without the original investment, and can facilitate more targeted attacks against the stolen model.

A broad class of computational techniques that use random sampling to estimate numerical results, particularly useful for problems that are too complex to solve analytically. In AI and reinforcement learning, Monte Carlo methods are used to estimate value functions and plan actions by simulating many possible future trajectories and averaging their outcomes.

The process of computing a sequence of movements that will take a robot from its current state to a desired goal state, while avoiding obstacles, respecting physical constraints, and optimizing for factors like speed or energy efficiency. Motion planning bridges the gap between high-level goals and the low-level actuator commands needed to achieve them.