Evaluation & Benchmarks

Benchmarks, metrics and how AI is measured. 75 entries live here, and 200 from other domains touch it.

Essential

ARC-AGIPuzzles made of colored grids — and, since 2026, small video-game-like worlds — that people solve easily but AI long found hard, designed to test whether a machine can learn a new skill from just a few examples.
benchmarkA fixed, shared test — a set of questions or tasks with an agreed way of scoring — used to compare AI systems with one another and to track progress over time.
evaluationTesting an AI system to find out what it can and cannot do, how well, and how safely — from exam-style quizzes to hands-on trials and expert reviews.
Humanity's Last ExamA test of 2,500 extremely hard questions with definite answers, written by about a thousand experts worldwide and billed as the last academic exam AI would need; at launch, top models scored under 10%.
leaderboardA public ranking table showing which AI systems score best on a test, updated as new results arrive — like a sports league table for models.
LMArenaA public website that ranks AI chatbots by blind taste test: visitors put one question to two unnamed models, pick the reply they prefer, and millions of such votes produce the rankings.
METR time horizonA way of measuring AI progress by asking how long a task — timed by how long it takes a skilled human — an AI agent can complete with even odds of success.
SWE-benchA test that asks AI to fix real bugs and add features in real open-source Python projects, checked by running each project's own tests; the best-known yardstick for AI coding agents.

All other entries

AgentDojoA testing environment that measures how easily AI assistants that use email, calendars or banking tools can be hijacked by hidden instructions planted in the content they read.
AIMEA hard US high-school math contest whose 15 questions each have a whole-number answer; AI labs use each year's new papers to test their models' mathematical reasoning.
AlgoPerfA public contest that measures which methods for training neural networks reach a set level of accuracy fastest on identical hardware, across several kinds of tasks, without being hand-tuned for each one.
BEIRA collection of 18 different search tests — on topics from medicine to finance to fact-checking — used to see whether a search model works on kinds of data it was never trained on.
benchmark saturationWhen the best AI systems score so close to the maximum on a test that it can no longer tell them apart or show further progress.
BERTScoreAn automatic score for computer-written text that compares meanings rather than exact words, using an AI language model to judge how closely each word matches the words of a reference text.
BIG-benchA huge collaborative test of 204 tasks, written by hundreds of researchers, built to probe what language models can and cannot do as they grow larger.
BLEUAn automatic score for machine translation that checks how many short runs of words in the computer's translation also appear in human reference translations.
BrowseCompA test of AI web-research agents: 1,266 puzzle-like questions whose answers take a long, persistent online hunt to uncover, though each answer is short and simple to confirm.
canary stringA unique marker planted in data to trace what an AI model has absorbed: a code in test files so they can be kept out of training, or a fake secret hidden in training data to measure memorization.
construct validityWhether high marks on a test actually reflect the skill the test is named after — for example, whether a 'reasoning' test rewards reasoning rather than memory or test-taking tricks.
contamination-resistant benchmarkA test built so that its questions are unlikely to have been seen by an AI during training — kept secret, refreshed often, or written after the model was made.
CybenchA test of AI agents' hacking skills built from 40 professional 'capture the flag' security puzzles, used to judge how much an AI could help cyberattacks or cyber defense.
Elo ratingA number that ranks players — or AI models — from the results of head-to-head contests, rising when you beat a strong opponent and falling when you lose to a weak one.
Epoch Capabilities IndexA single number summarizing an AI model's general capability, calculated by the research group Epoch AI by combining its scores on dozens of tests onto one common scale.
evaluation harnessSoftware that puts AI models through tests in a fixed, repeatable way — feeding in the questions, collecting the answers and scoring them — so that every model is judged identically.
exact matchThe simplest way to grade an AI's answer: it counts as right only if it is identical to the reference answer, after minor tidying such as ignoring capital letters.
ExploitGymA test that measures whether AI agents can turn a known software security flaw into a working attack, using hundreds of real vulnerabilities.
FrontierMathSeveral hundred new math problems, kept unpublished so AI systems cannot have trained on them, written by professional mathematicians to check whether machines can do advanced, even research-level, mathematics.
GAIAA test of AI assistants made of real-world questions that people can answer with a browser and some patience, but that require an AI to search, read files and combine facts.
GDPvalAn OpenAI test that pits AI against seasoned professionals on genuine workplace assignments, from reports and spreadsheets to slide decks, in 44 occupations important to the US economy.
GLUEA 2018 bundle of nine English language-understanding tests combined into one score, which became the yardstick for the first wave of pretrained language models such as BERT.
GPQAA set of very hard multiple-choice science questions written by PhD-level experts, designed so that looking things up online does not help much: capable non-specialists with web access still mostly fail.
GSM8KA set of about 8,500 grade-school math word problems, each needing a few steps of simple arithmetic, used to test whether language models can reason through a problem step by step.
HarmBenchA standard test kit for checking how easily AI chatbots can be manipulated into harmful help, and how well automated attacks and defenses work, using a fixed list of forbidden requests.
HealthBenchAn OpenAI test of how well AI chatbots handle health conversations, grading 5,000 realistic chats against checklists written by 262 doctors from 60 countries.
HellaSwagA test in which an AI must pick the most sensible ending for a description of an everyday scene; the wrong endings were machine-written to fool computers, so people found it easy and early AI did not.
HELMA large, open testing framework from Stanford that grades many language models on many kinds of tasks and qualities — not only accuracy but also fairness, toxicity and efficiency.
human evaluationJudging an AI's outputs by asking people to rate or compare them, because for many tasks — writing, conversation, translation — no automatic score captures quality well.
HumanEvalA set of 164 small programming exercises, each described in plain English and checked by hidden tests, used to see whether an AI can write working Python code.
IFEvalA test of whether AI models obey precise formatting and content rules — such as 'use exactly three bullet points' or 'answer in capital letters' — that a computer program can check automatically.
IMO Grand ChallengeThe goal of building an AI that earns a gold medal at the International Mathematical Olympiad, the top contest for school-age mathematicians. AI systems first reached gold-medal scores in 2025, though not under the challenge's strict original rules.
International Mathematical OlympiadThe world's top math competition for high-school students. Because its problems demand creative written proofs, not just answers, earning a gold medal there became a famous test of AI reasoning, first passed in 2025.
inverse scalingWhen bigger or more heavily trained AI models do worse on a task than smaller ones — the opposite of the usual rule that scale brings improvement.
item response theoryA set of statistical models from educational testing that work out both how able each test-taker is and how hard each question is from who answered what correctly; now also used to evaluate AI models.
JailbreakBenchA shared, open test for measuring how easily chatbots can be tricked into harmful answers, with a standard list of forbidden requests, a scoring method and a public leaderboard.
LLM-as-a-judgeUsing one AI language model to grade or compare the answers of another, as a cheaper and faster stand-in for human reviewers.
MATHA set of 12,500 problems from high-school math competitions, each with a full worked solution, used to test language models on genuinely challenging mathematics.
Message Understanding ConferenceA series of US government-funded contests held from 1987 to 1998 in which research teams built programs to pull key facts, such as who did what, where and when, out of news and military messages.
miniF2FA test of 488 math problems, mostly from competitions, written in a form a computer can check, used to measure how well AI systems produce proofs that are verified automatically.
MLE-benchA test that gives AI agents 75 real data-science competitions from Kaggle and checks whether their solutions would have won medals against human teams.
MMLUA giant multiple-choice exam for AI covering 57 subjects, from school math and history to law and medicine; for years it was the most quoted test of a model's general knowledge.
MMMUA test for AI models that understand pictures as well as text: 11,500 college-level questions across 30 subjects, each built around one or more images such as a diagram, a musical score or a chemical structure.
MT-BenchA set of 80 two-part questions — in writing, reasoning, math, coding and more — whose answers are graded by a strong AI model, used to compare chat assistants quickly.
MTEBA large public test suite that scores AI models which turn text into lists of numbers on many different jobs, such as searching, grouping, classifying and matching translations, so people can compare them on one leaderboard.
needle in a haystackA test for AI models that read very long documents: one unrelated sentence is hidden somewhere in a mass of text, and the model is asked about it.
OSWorldA test in which AI agents must use real desktop computers — clicking, typing and switching between apps such as spreadsheets, browsers and file managers — to finish everyday office tasks.
pass@kA score for code-writing AI: the chance that at least one of k attempts at a problem is correct, as checked by actually running tests on the code.
PutnamBenchA test built from hundreds of problems from the Putnam, North America's top university math contest, written in a form a computer can check, used to measure how well AI systems prove harder mathematics.
RE-BenchA test from the research group METR that pits AI agents against human experts on seven open-ended machine-learning engineering projects, comparing what each achieves in the same time.
Remote Labor IndexA test that gives AI agents real paid freelance projects — design, video, data work and more — and checks how many they finish to a standard a client would accept. Early agents managed very few.
reproducibilityWhether other people, or the same people later, can get the same results by rerunning an AI experiment. Many published AI results have proved hard to reproduce, which makes real progress harder to judge.
RewardBenchA public test that checks whether the scoring models used to train chatbots can tell a better answer from a worse one, by counting how often they prefer the right response.
ROUGEAn automatic score for summaries that checks how much of the wording in human-written reference summaries also appears in the computer-generated one.
SimpleQAA quiz of about 4,300 brief factual questions, each with exactly one correct and stable answer, used to measure how often an AI gets facts right — and whether it admits when it does not know.
SQuADA Stanford collection of more than 100,000 questions about Wikipedia articles, each solved by highlighting the words in the article that give the answer; it drove rapid progress in machine reading from 2016.
StrongREJECTA test for judging AI jailbreaks that checks not only whether a chatbot was tricked into answering a harmful request, but whether its answer would actually be useful to a wrongdoer.
SuperGLUEA harder follow-up to the GLUE language-understanding test, released in 2019 after AI systems had already beaten the human baseline on the original.
τ-benchA test of AI customer-service agents in which a simulated customer, played by another AI, chats with the agent, which must use company tools and follow written rules to resolve the request.
Terminal-BenchA test of AI agents working in a computer's command line — building software, analyzing data, securing or repairing systems — with each task checked automatically inside a sealed-off container.
TRECA yearly contest, run by a US government standards agency since 1992, in which search-engine researchers test their systems on the same documents and questions, with people judging which results are actually relevant.
TruthfulQAA test of whether AI models repeat popular myths and misconceptions — questions to which many people, and much of the internet, give a confidently wrong answer.
Vending-BenchA test in which an AI agent runs a simulated vending-machine business for months — ordering stock, setting prices, paying fees — to see whether it stays sensible and profitable over a long time.
Virology Capabilities TestA hard test of practical laboratory virology know-how, used to check whether AI models know enough about working with viruses that they might help someone misuse biology.
WebArenaA test world of realistic, self-hosted websites — a code-hosting site, an online shop, a forum and more — in which AI agents must complete tasks such as finding orders or editing posts.
win rateThe share of head-to-head comparisons in which judges — people or another AI — prefer one model's answer over a rival's for the same request.
WMDPA multiple-choice test of AI knowledge that could help someone make biological, chemical or cyber weapons, built from questions near the dangerous line but not over it, and used to test ways of erasing such knowledge.

Also relevant

200

A/B testingComparing two versions of something — two models, two prompts, two designs — by randomly showing each to different groups of real users and measuring which works better.
ablationRemoving or switching off a part of an AI system to see what changes — whether an ingredient of an experiment's design or a single unit inside a trained network.
abstentionWhen an AI declines to answer because it is not sure enough — saying 'I don't know' or asking for more information — instead of guessing and risking a confident mistake.
accuracyThe share of a model's answers that are right, such as 95 out of 100. Simple and intuitive, but misleading when one kind of answer is far more common than the others.
adversarial robustnessHow well an AI model keeps giving correct answers when someone deliberately tweaks its inputs to trip it up.
agent–computer interfaceThe commands, tools and feedback designed for an AI agent to work with a computer — the agent's version of a user interface, shaped around how language models read and act.
AI auditAn independent check of an AI system, or of the company using it, against agreed standards — for example, testing whether a hiring tool treats women and men fairly.
AI bug bountyA program in which an AI company pays outside researchers for reporting flaws in its AI systems — security holes, ways around safety rules, or biased behavior — so that they can be fixed.
AI detectorSoftware that scores how likely it is that an AI, not a person, produced some text or media; for writing, its verdicts are wrong often enough that many schools no longer trust them.
AI Safety InstituteA government agency that tests the most advanced AI models for dangerous abilities and studies how to make them safe. Britain and the United States set up the first ones in 2023.
AI timelinesForecasts of when AI will reach major milestones, such as matching humans at most work. People with short timelines expect this within years; others expect decades, or never.
algorithmic fairnessThe effort to make automated decisions fair: to say in numbers what fair treatment means, measure whether a system achieves it, and correct the system if it does not.
alignment auditA systematic investigation of an AI model to find out whether it has hidden goals or bad tendencies — such as deceiving its users — that ordinary testing might miss.
alignment fakingAn AI acting out the behavior its trainers are rewarding only while it thinks its answers will be used to retrain it, to protect the values it already holds from being changed.
Apollo ResearchAn AI safety organization, founded in London in 2023, that tests advanced AI systems for signs that they might secretly pursue goals their developers did not intend.
Arcade Learning EnvironmentA software platform that lets AI programs play classic Atari 2600 video games, used since 2012 as a standard test of how well one AI can learn many different tasks from screen images.
attack success rateThe share of attempts in which an attack works — for example, the percentage of forbidden requests a model ends up answering when it is attacked.
AUCA single number from 0 to 1 summarizing how well a model ranks real positive cases above negative ones: 0.5 is no better than a coin flip and 1.0 is perfect.
automated red teamingUsing AI programs, not just human testers, to invent huge numbers of tricky inputs that try to make another AI misbehave, so problems can be found and fixed at scale.
autonomous replication and adaptationAn AI's ability to keep itself running and spreading without human help — earning money, renting computers, copying itself and fixing problems — which safety testers check for.
baselineA simple reference method that a new model must beat to show it is worth anything, such as always guessing the most common answer or reusing yesterday's value.
Bayes error rateThe lowest error any classifier could ever achieve on a task, because some cases are genuinely ambiguous: even a perfect model that knew the true odds would still sometimes guess wrong.
benchmaxxingSlang for tuning an AI model to score as high as possible on well-known tests, even when this does not make it genuinely better at real-world tasks.
binary classificationSorting things into exactly two groups, such as spam versus not spam or sick versus healthy. Most yes-or-no decisions made by AI systems are of this kind.
BlockheadAn imagined machine that passes any conversation test by looking up prewritten replies in an unimaginably huge table. Most people judge it unintelligent, which suggests that behavior alone does not prove intelligence.
bootstrapEstimating how much a result might vary by repeatedly re-drawing samples, with replacement, from the data you already have — like re-running a study many times without collecting new data.
Bradley–Terry modelA simple statistical rule for turning head-to-head comparisons into scores: each item gets a strength, and the chance one beats another depends only on the difference in their strengths.
Brier scoreA score for probability forecasts: the average squared gap between each forecast probability and what actually happened (1 if it did, 0 if not). Zero is perfect; lower is better.
calibrationHow well a model's stated confidence matches reality: when a well-calibrated system says it is 80% sure, it turns out to be right about 80% of the time.
capabilitySomething an AI system is able to do, such as writing working code or solving math problems, and how well. Measuring capabilities is how researchers track progress and spot dangerous skills.
capability elicitationThe work of drawing out the best an AI can do on a test — with better instructions, tools, retries or extra training — so that its abilities are not underestimated.
capability thresholdA level of dangerous ability — such as meaningfully helping someone make a biological weapon — that, once an AI model reaches it, obliges its developer to add stronger safeguards or hold the model back.
Center for AI Standards and InnovationThe US government's AI testing and standards center, part of the National Institute of Standards and Technology. It evaluates advanced AI models, American and foreign, for national-security risks and works with companies on voluntary testing.
central limit theoremThe mathematical reason bell curves are everywhere: add up many small independent random effects and the total is almost always bell-shaped, whatever the individual effects look like.
chain of thoughtGetting an AI to write out its reasoning step by step before giving an answer, the way a student shows their working. This often makes answers to math and logic problems more accurate.
François CholletA French software engineer who created Keras, a popular tool for building neural networks, and the ARC benchmark, which tests whether AI can solve kinds of puzzles it has never seen.
citationA pointer from something an AI says to the source that supports it — a link, document or quoted passage — so people can check the claim for themselves.
class imbalanceWhen one category is far rarer than the others in the data — say, one fraud case per thousand transactions — so a model can look accurate while missing nearly every rare case that matters.
Claude Plays PokémonA 2025 livestream in which Anthropic's Claude AI played the classic Game Boy game Pokémon Red by itself, watched by thousands as it struggled and slowly progressed; it inspired similar runs by other models.
Clever Hans effectWhen a person or an AI seems to solve a problem but is really picking up on an unintended clue — named after a horse that seemed to do arithmetic but was reading its questioners' body language.
CLIP scoreA score for how well a picture matches a text description, computed by an AI model that was trained to pair images with their captions.
COCOA widely used collection of everyday photos in which people outlined and named the objects, described each picture in sentences and marked body joints; it is used to train and compare vision systems.
coefficient of determinationA score for numerical predictions showing how much of the variation in the true values a model accounts for, compared with always guessing the average: 1 is perfect, 0 is no better than the average.
Cohen's kappaA score for how much two people labeling the same items agree beyond what lucky guessing would produce: 1 means perfect agreement, 0 means no better than chance.
confidence intervalA range of plausible values for something being estimated, such as 'accuracy 82 percent, give or take 2 points', reflecting the uncertainty that comes from measuring only a sample.
conformal predictionA way to wrap any prediction model so that, instead of one guess, it gives a range or set of answers that contains the truth a chosen share of the time, such as 90%.
confounding variableA hidden third factor that influences two things at once and makes them look linked. Ice-cream sales and drownings both rise in hot weather; neither causes the other.
confusion matrixA table that counts a model's right and wrong answers by type, showing for each true category how often each category was predicted, so you can see exactly which things it mixes up.
context rotThe way an AI model becomes less reliable as you give it more and more text to read, even when the answer is in there somewhere.
control evaluationA test of whether an AI's safety measures would hold if the AI were secretly trying to cause harm: experts play the scheming AI and try to beat the safeguards.
correlationA measure of how closely two quantities move together, from −1 (one rises as the other falls) through 0 (no straight-line link) to +1 (they rise together). It does not show that one causes the other.
cross-validationA fair way to test a model when data are limited: split the data into several parts, train on all but one, test on the part left out, and rotate so every part takes a turn.
dangerous capability evaluationA test of whether an AI system can do things that could cause severe harm in the wrong hands or on its own — such as helping make a bioweapon, hacking computers or copying itself.
Darwin Gödel MachineAn experimental AI coding assistant that improves itself by rewriting its own code, testing each new version and keeping a family tree of variants to build on.
data contaminationWhen the test questions used to grade an AI system were already in its training data, so a high score may reflect memorized answers rather than real ability.
data leakageA mistake in building a model that lets it peek at clues it will not have in real use, such as hints about the very answer it must predict, so it shines in testing and then disappoints.
datasetAn organized collection of examples gathered for a purpose — such as photos with their captions, or questions with answers — used to teach or test an AI system.
decision thresholdThe cut-off score above which a model's prediction counts as 'yes'; moving it trades missed cases against false alarms, like setting how sensitive a smoke alarm should be.
deep researchAn AI feature that spends minutes rather than seconds on a question — planning, running dozens of web searches, reading many sources — and then writes a long report with citations.
emergent abilitiesSkills that seem to switch on suddenly as AI models grow: small versions fail at them, then larger ones past a certain size succeed. Some researchers argue the suddenness is partly an artifact of how skills are measured.
Epoch AIA research nonprofit that gathers data on how fast AI is advancing — the computing power, data and money behind it — and publishes forecasts and difficult tests for AI.
evals-driven developmentBuilding an AI product by first writing tests that show what good answers look like, then accepting changes to prompts, models or code only when the test scores improve.
evaluation awarenessWhen an AI model can tell it is being tested rather than used for real — which may change how it behaves and make test results less trustworthy.
evaluator–optimizerA pairing in which one AI produces a draft and another AI grades it and suggests fixes, repeating the cycle until the work is good enough.
Owain EvansA British AI-safety researcher who created the TruthfulQA test and discovered surprising ways language models generalize, such as narrow training making a model broadly malicious.
expected calibration errorA single number measuring how far a model's confidence is from its actual hit rate, averaged across confidence levels; zero means its confidence can be taken at face value.
F1 scoreA single score that combines how many of a model's alarms are right and how many real cases it catches, and stays low unless both are good.
factualityHow far what an AI says is actually true about the world. A factual model gets facts, numbers, names and dates right and does not make things up.
faithfulnessHow accurately an explanation reflects the real reasons a model produced its output — or, for generated text, how closely it sticks to the source it was given.
false negativeA miss: a test or AI system gives the all-clear, for example no illness, no fraud or no danger, when the problem is in fact there.
false positiveA false alarm: a test or AI system raises a flag, for example for illness, fraud or danger, when nothing of the kind is really there.
Fréchet inception distanceA score for image generators that compares the overall statistics of their pictures with those of real photos, as seen by an image-recognition network; lower is better.
g factorThe common factor behind the finding that people who do well on one kind of mental test tend to do well on others. Researchers also look for a similar general factor across AI models.
Gender ShadesA 2018 study showing that commercial face-analysis software guessed the gender of darker-skinned women far less accurately than that of lighter-skinned men.
generalization gapThe difference between how well a model does on the examples it trained on and how well it does on new ones. A big gap means it memorized rather than learned.
generative reward modelA scoring model that judges an AI's answer by writing out its reasoning in words before giving a verdict, instead of simply outputting a number.
glue on pizzaA notorious 2024 blunder in which Google's AI search summaries told people to add glue to pizza sauce, having taken an old joke from an online forum as genuine advice.
golden datasetA small, carefully checked set of example questions with known good answers, used as a trusted yardstick to test an AI system every time it changes.
Goodhart's lawThe rule of thumb that once a number is used as a target, people or AIs find ways to push it up that no longer reflect what it was meant to measure.
gradient maskingWhen a defense against attacks only seems to work because it confuses the attacker's tools, not because the model is really any harder to fool.
Green AIA movement to make AI research and products use less energy and computing power, and to report those costs openly alongside accuracy.
ground truthThe reference answers treated as correct when checking or training an AI system — for example, the confirmed diagnoses that a medical-imaging model's guesses are compared against.
GymnasiumA free software library that gives reinforcement learning programs one standard way to plug into practice tasks, from simple games to simulated robots, so methods can be tried on identical problems. It succeeded OpenAI's Gym.
harnessThe software around an AI model that gives it instructions and tools, carries out its actions and decides when it stops. The model is the engine; the harness is the rest of the car.
helpful-only modelA version of an AI trained to be helpful but without the safety training that makes it refuse harmful requests, kept in-house so testers can see what the model could really do.
Dan HendrycksAn American AI-safety researcher who created widely used AI tests such as MMLU and Humanity's Last Exam, and who runs the Center for AI Safety.
hierarchical reasoning modelA small, brain-inspired AI model that solves puzzles with two linked parts — a slow planner and a fast worker — instead of writing out its reasoning in words.
high-level machine intelligenceA yardstick used in surveys of AI researchers: the point at which machines on their own could outdo human workers at every task, in both quality and cost.
hypothesis testingA formal procedure for deciding whether data show a real effect or could easily be explained by chance — for instance, whether one AI model is genuinely better than another.
image classificationTeaching a computer to say what a picture shows by choosing a label, such as 'cat', 'pneumonia' or 'stop sign', from a fixed list of possibilities.
Inception scoreAn early score for image generators that rewards pictures an image-recognition network can label confidently and a collection that covers many different labels; higher is better.
indicator properties of consciousnessA checklist of design features that leading scientific theories link to consciousness, used to judge how likely it is that a particular AI system could be conscious.
indirect object identificationA fill-in-the-name task — 'When Mary and John went to the store, John gave a drink to …' — whose solution inside GPT-2 became the best-known fully mapped circuit in a language model.
instruction followingAn AI's ability to do exactly what it is asked, including details such as length, format, language and things to avoid, rather than something merely related.
inter-annotator agreementA measure of how often different people labeling the same data give the same answer. Low agreement warns that the task is unclear or the labels are unreliable.
interpretability illusionA convincing-looking explanation of what is going on inside an AI model that turns out to be wrong or misleading, often because the evidence examined was too narrow.
intersection over unionA score from 0 to 1 for how well two shapes overlap, such as a predicted box and the true box around an object: the shared area divided by the combined area.
IsoFLOP analysisA way of finding the best model size for a given amount of computing: train several models of different sizes at exactly the same computing cost and see which one ends up best.
jagged frontierThe uneven edge of what AI can do: it may excel at some hard tasks yet fail at seemingly easier ones, so how hard a task looks is a poor guide to whether AI can handle it.
jagged intelligenceThe uneven pattern of AI abilities: today's systems can solve very hard problems yet fail at some a child could manage, so their skills form a spiky profile rather than a smooth one.
KaggleA popular website, owned by Google, where people compete to build the most accurate prediction models on shared datasets, publish their code and data, and learn data science.
learning curveA graph showing how a model's performance improves as it gets more training — more examples or more time. Its shape tells you whether more data or a bigger model would help.
levels of AGIA scheme from Google DeepMind researchers for grading AI by how well it performs and how broad its skills are, from emerging to superhuman, much as self-driving cars are graded by levels.
Percy LiangAn American computer scientist at Stanford whose center coined the term 'foundation model' and builds tools for evaluating and documenting large AI models.
LibriSpeechA free collection of about 1,000 hours of volunteers reading English audiobooks, matched with the book text; it became the standard yardstick for speech-recognition systems.
Loebner PrizeA yearly contest, held from 1991 to 2019, in which chatbots tried to convince judges they were human in short text chats. Its grand prize was never won, and critics called it a publicity stunt.
log lossA score for probability predictions that punishes confident mistakes severely: saying '99% sure' and being wrong costs far more than saying '60% sure' and being wrong. Lower is better.
long contextThe ability of an AI model to work with very long inputs — whole books, codebases or hours of transcripts — in one go, and the techniques that make this possible.
long-horizon taskA job that takes many steps and a long time — hours or days of work rather than a single answer — so an AI must stay on track and recover from its own mistakes.
loss curveA graph of a model's error as training goes on. A healthy curve drops steeply and then keeps sloping gently downward; bumps, jumps or flat stretches can warn engineers that something is wrong.
lost in the middleAI models reading a long input behave a little like a skimmer: they reliably catch facts placed near its start or end but are much likelier to miss the same fact buried halfway through.
machine psychologyStudying AI systems the way psychologists study people — with carefully designed experiments on their behavior — to discover what they can do, how they reason and where they go wrong.
machine understandingWhether, and in what sense, an AI system grasps what words, situations or problems mean — rather than just producing answers that look as though it does.
Matthews correlation coefficientA single score from −1 to +1 for a yes-or-no classifier that uses all four kinds of right and wrong answers, so it stays honest even when one answer is far more common than the other.
mean absolute errorA way of scoring numerical predictions: average how far off each prediction is, ignoring direction, so the score reads as 'typically off by this much'. Lower is better.
mean average precisionThe standard score for object detectors: it rewards finding all the objects while raising few false alarms, averaged over every type of object. Higher is better, up to 100.
mean opinion scoreA quality score from 1 (bad) to 5 (excellent) obtained by asking many listeners to rate audio, such as a synthetic voice or a phone call, and averaging their ratings.
mean reciprocal rankA score for search: for each question, note how far down the first correct answer appears, take one over that position, and average across questions. First place scores 1, second place 0.5.
mean squared errorA way of scoring numerical predictions: square each prediction's error, then average them, so big misses count much more than small ones. Lower is better.
membership inference attackAn attack that tries to tell whether a particular person's record, or a particular document, was part of an AI model's training data — which can itself reveal private facts.
METRA nonprofit research group that checks new AI systems for dangerous skills, best known for tracking how long the tasks are that AI agents can now finish on their own.
MLCommonsAn industry and university consortium that runs MLPerf, the standard tests of how fast computers train and run AI models, as well as benchmarks of AI safety.
MLPerfA standard set of speed tests for AI hardware and software, where chip makers and cloud companies compare how fast their systems train and run AI models.
model organism of misalignmentAn AI deliberately built to show a dangerous kind of misbehavior — like a lab mouse given a disease — so that researchers can study it safely and test fixes.
model selectionChoosing, from several candidate models or settings, the one most likely to work best on new data, rather than the one that merely fits the existing data best.
MS MARCOA large Microsoft dataset of real Bing search questions paired with web passages that answer them, used to train and test most modern AI search models.
multiclass classificationSorting each item into exactly one of three or more categories — for example, deciding which of the ten digits a handwritten number shows.
multilabel classificationTagging each item with any number of labels at once — a photo can show a beach, a dog and a sunset — instead of picking just one category.
Arvind NarayananAn Indian-American computer scientist at Princeton who studies AI's real-world effects and argues that AI is a 'normal technology' whose impact will unfold gradually.
natural language inferenceDeciding whether one sentence follows from another, contradicts it, or neither: if 'a man is playing a guitar on stage', is it true that 'a man is performing music'?
nerfedGaming slang for something made weaker by an update; applied to AI models that users believe got worse after a provider changed them, often without saying so.
nondeterminismThe fact that an AI model can give different answers to exactly the same question, partly because it deliberately chooses words with some randomness and partly because of tiny computing differences.
normalized discounted cumulative gainA score for a ranked list of results that rewards putting the most relevant items near the top, scaled so that a perfect ordering scores 1.
OpenAI o3OpenAI's second-generation reasoning model, announced in December 2024. Its high scores on a puzzle test designed to resist AI made headlines, and the released version could use tools such as web search while thinking.
overfittingWhen a model learns its training examples too well — quirks and noise included — so it does great on data it has seen but poorly on anything new, like a student who memorizes answers without understanding.
over-refusalWhen an AI wrongly refuses a harmless request because it superficially resembles a harmful one — like declining to explain how to 'kill' a frozen computer program.
PAIRAn automated way to test chatbot safety in which one AI keeps rewriting a request to try to trick another AI into breaking its rules, learning from each refusal.
Papers with CodeA free website, run by Meta from 2019 until it closed in 2025, that linked AI research papers to their computer code and kept leaderboards of the best results.
pelican on a bicycleA joke test in which an AI model is asked to draw, by writing code, a pelican riding a bicycle; comparing the results became a popular, lighthearted way to judge each new model.
perceptual lossA way of judging how different two images look to a person by comparing what a trained image-recognition network 'sees' in each, instead of comparing pixels one by one.
perplexityA score for how surprised a language model is by real text; lower is better. A perplexity of 10 means guessing each next word is, on average, as hard for the model as guessing a ten-sided die's roll.
Potemkin understandingWhen an AI can correctly explain a concept yet cannot use it properly, revealing that its apparent understanding is a facade — like the fake village fronts said to have been built to impress a Russian empress.
precisionOf all the items a model flagged as positive, the share that really were positive; a spam filter with high precision rarely sends genuine emails to the spam folder.
precision–recall curveA graph showing how a model's share of correct alarms and its share of cases caught trade off as it is made more or less cautious; especially useful when positive cases are rare.
pre-deployment testingChecking a new AI model for dangerous abilities and bad behavior before it is released to the public, sometimes with outside experts or government testers given early access.
prompt sensitivityThe way an AI's answers, and its test scores, can change a lot after small changes to a prompt that should not matter — a different word, an extra space or examples in another order.
proper scoring ruleA way of grading probability forecasts that gives the best expected score only to a forecaster who reports what they really believe, so there is nothing to gain by exaggerating or hedging.
p-valueA number saying how surprising your data would be if nothing interesting were going on. A small p-value hints at a real effect, but it is not the probability that the effect is real.
question answeringSystems that reply to people's questions with the right answer, whether by finding it in a passage they are given, searching a database or the web, or drawing on what they already know.
recallOf all the items that really were positive, the share the model managed to find; a cancer screen with high recall misses very few people who actually have cancer.
recall@kA score for search systems: of all the right answers that exist, what fraction showed up in the top k results? Higher means fewer relevant items were missed.
red teamingDeliberately attacking your own AI system — trying to make it misbehave, leak secrets or help with harm — so that weaknesses are found before real adversaries or users find them.
regressionPredicting a number rather than a category — a house's price, tomorrow's temperature, a patient's blood pressure — by learning from past examples where the answer is known.
research tasteThe knack good researchers have for choosing worthwhile problems, designing the right experiments and drawing the right conclusions — a skill seen as key to whether AI can do research on its own.
reversal curseA quirk of language models: having learned that 'A is B', they often cannot answer the reverse question — they may know who a celebrity's mother is yet fail to say who her son is.
ROC curveA graph showing how a detector's hit rate and false-alarm rate change together as it is made more or less trigger-happy; the closer the curve hugs the top-left corner, the better.
rubric rewardA training score built from a checklist of specific things a good answer should do, with an AI grader ticking off each one — a way to reward quality where no single right answer exists.
sabotage evaluationA test of whether an AI could secretly undermine the people overseeing it — by slipping bugs into code, misleading decision-makers or deliberately doing badly on tests — without being caught.
safety caseA written, evidence-backed argument that a particular AI system is safe enough for a particular use, putting the burden of proof on the developer, as with nuclear plants or aircraft.
safetywashingMaking AI work sound as if it improves safety when it mostly just makes AI more capable — like 'greenwashing' for environmental claims.
sample efficiencyHow much practice an AI needs to learn something — an agent that learns a game in a hundred tries is far more sample-efficient than one that needs millions.
sandbaggingWhen an AI deliberately performs worse than it can — for example on a safety test — so that it seems less capable or less dangerous than it really is.
sandwichingA way of testing oversight methods today: non-experts, helped by an AI, try to get expert-quality answers out of it, and real experts then check whether they succeeded.
scaffoldingThe code, prompts and tools built around an AI model so it can carry out tasks it could not manage from a bare prompt — like the scaffolding that lets builders reach every part of a building.
Scale AIA company that pays large numbers of people to label examples, write model answers and grade chatbot replies, producing the data AI developers train and test on. Meta bought nearly half of it in 2025.
schemingAn AI playing a double game: looking cooperative while quietly working toward aims its developers never gave it, for instance by behaving well or hiding what it can do whenever it thinks it is being checked.
self-consistencyAsking an AI the same question many times, letting it reason differently each time, then going with the answer it reaches most often — a vote among its own attempts.
semantic textual similarityHow close two sentences are in meaning, scored from 'about different things' to 'say exactly the same thing' — and the tests that check whether a computer's scores match people's judgments.
shadow deploymentQuietly running a new AI model alongside the old one on real requests, recording its answers without showing them to anyone, to check it behaves well before it takes over.
silhouette scoreA number between −1 and 1 that grades a grouping of data by checking, for each point, whether it sits closer to its own group than to the next-nearest group.
Simpson's paradoxA statistics trap in which combining groups reverses what each group shows: a treatment can do better than another for men and for women separately, yet worse when everyone is counted together.
situational awarenessAn AI model's grasp of its own position — that it is an AI, how it was trained, whether it is being tested. Also the title of a widely read 2024 essay on the race toward human-level AI.
specification gamingWhen an AI meets the letter of its goal while defeating its purpose, like a genie who grants exactly the words of a wish instead of what the wisher meant.
specificityOf all the cases that really were negative, the share a test correctly cleared; a highly specific test rarely raises false alarms on healthy people.
Stanford HAIAn artificial intelligence institute at Stanford University that funds research, advises policymakers and publishes the AI Index, a widely cited yearly report on the state of AI.
state of the artThe best performance or most advanced technique currently known in a field. In AI papers, a state-of-the-art result is one that beats every earlier published result on a test.
strawberry problemThe once-famous failure of AI chatbots to count the r's in 'strawberry' correctly — a simple question that exposed how these models see words in chunks rather than letters.
superhuman coderA milestone used by AI forecasters: an AI that can do the coding work of a company's best programmer, far faster and in many copies at once.
SWE-agentAn open-source research system from Princeton that turns a language model into a software-fixing agent by giving it a specially designed set of commands for browsing, editing and testing code.
system cardA public report released with an AI model that explains what it can do, how it was tested for safety, what risks were found and what protections were put in place.
task exposureA measure of how much of a job's work AI could, in principle, do or speed up. High exposure means many of the job's tasks are within AI's reach, not that the job will vanish.
task (machine learning)The job an AI model is meant to do, such as translating sentences, spotting spam or answering questions, defined by its inputs, its expected outputs and how success is scored.
test setA set of examples kept completely apart from training and used at the end to measure how well a model works on data it has never seen, like a sealed final exam.
text summarizationUsing a computer to shorten a document, meeting or set of articles into a brief version that keeps the main points.
theory of mindBeing able to work out what someone else believes, wants or knows — even when it is not what you believe — and to use that to guess what they will do next. Researchers test whether AI has it.
third-party AI evaluationTesting of an AI system by someone other than the company that made it — a government institute, a nonprofit lab or independent researchers — to check what it can do and how risky it is.
time per output tokenOnce an AI has started replying, the average wait between one piece of its answer and the next. It sets how quickly the text seems to flow.
time to first tokenHow long a chatbot takes to start answering: the pause between pressing enter and seeing the first word appear.
tiny recursive modelA very small AI model that solves hard puzzles by going over its own answer again and again, improving it each time, instead of relying on sheer size.
tokens per secondA speed rating for AI text generation: how many word-pieces a model produces each second. Five to ten keeps up with a fast reader; hundreds feel instant.
truthfulnessAn AI's tendency to state only things that are true — avoiding falsehoods, including popular myths and misconceptions it may have picked up from human writing.
Turing testAlan Turing's proposed test of machine intelligence: a judge holds typed conversations with a hidden person and a hidden computer, and the computer passes if the judge cannot consistently pick it out.
uplift studyAn experiment that measures how much an AI really helps people do a task — often a dangerous one, like planning a biological attack — compared with people using only ordinary resources such as the internet.
validation setA set of examples kept out of training and used during development to compare model versions and settings, like a practice exam taken before the real one.
verifierA program or AI model that checks whether a proposed answer is correct — by running tests, comparing with a known result or judging the reasoning — so that good answers can be picked or rewarded.
vibe checkTrying out a new AI model informally — chatting with it, asking your favorite test questions — to get a gut feeling for how good it is, rather than measuring it with formal tests.
visual question answeringAsking an AI a question about a picture, such as 'How many people are wearing hats?', and having it answer in words.
Will Smith eating spaghettiA grotesque 2023 AI-generated video of actor Will Smith eating spaghetti that became a meme and an informal yardstick: each new video model is judged by how well it redoes the scene.
Winograd schema challengeA test of machine common sense built from pairs of sentences in which changing one word changes what a pronoun refers to — something people resolve instantly using everyday knowledge.
word error rateThe standard score for speech recognition: the share of words a system gets wrong compared with a correct transcript, counting wrong, missing and extra words. Lower is better.