AI glossary

373 words you'll meet in the AI world, from the basics to security, privacy law, business and coding. Every term has its own page. 293 come straight from our storybooks.

Levels: Beginner Intermediate Advanced· Jump to the A–Z index

🧠 AI fundamentals

The basic words everyone should know.

AccuracybeginnerHow often something is right.
AI modelbeginnerThe trained program that makes the predictions. GPT, Gemini, Claude and Llama are families of models.
AlgorithmbeginnerA step-by-step set of instructions for solving a problem.
Artificial general intelligence (AGI)advancedA hoped-for kind of AI that could match people across most kinds of thinking work. No system does this today.
Artificial Intelligence (AI)beginnerComputer programs that can learn and make smart guesses.
BenchmarkintermediateA standard test used to compare AI models.
DatabeginnerFacts, numbers, words, pictures or sounds that a computer can store and learn from.
Deep learningintermediateMachine learning that uses neural networks with many layers.
Machine learningbeginnerA kind of AI that learns patterns from data instead of following rules a person wrote.
Narrow AIbeginnerAI built to do one kind of task well.
ParameterintermediateOne of the internal numbers a model adjusts while training. Large models have billions.

📈 How AI learns

Data, training, testing and neural networks.

AnomalybeginnerSomething that doesn’t fit the usual pattern.
Artificial neuronintermediateA tiny calculator that multiplies its inputs by weights, adds them, and decides.
AttentionadvancedThe part of a transformer that lets each word look at the other words that matter for its meaning.
BackpropagationadvancedThe method networks use to work out the slope for every weight at once.
ClassifierintermediateAn AI that sorts things into groups, like wet or dry.
ClusteringbeginnerPutting similar things into groups.
ConfidenceintermediateHow sure a model says it is about an answer.
Confidence scorebeginnerHow sure an AI is about its answer.
Error, or lossintermediateA number that measures how wrong a model’s answers are.
FeaturebeginnerA clue the AI looks at, like colour or shape.
GeneralisingintermediateDoing well on new examples, not just old ones.
Gradient descentintermediateImproving a model by stepping each weight a little downhill on the error, again and again.
Held-out setintermediateData kept aside, like a sealed exam paper.
InputintermediateA number that goes into a neuron, like how clear the sky is.
LabelbeginnerThe right answer given with an example.
LayerintermediateA group of neurons working side by side, passing answers to the next group.
Learning rateintermediateHow big each step is.
Local minimumadvancedA low point that isn’t the lowest point.
Neural networkintermediateMany artificial neurons connected in layers.
OverfittingintermediateMemorising the training examples so closely that new examples go wrong.
OvershootintermediateStepping so far that you go past the best answer.
PatternbeginnerSomething that repeats, which helps you guess.
PredictionbeginnerThe AI’s best guess.
RegressionintermediatePredicting a number, rather than choosing a category.
Reinforcement learningintermediateLearning by trial and error, getting rewards for good actions.
Slope, or gradientintermediateWhich way, and how steeply, the error changes when a weight changes.
Spurious patternintermediateAn accident in the data that looks like a real rule.
Supervised learningbeginnerLearning from examples that come with answers.
TamperingintermediateSecretly changing a system so it gives the answers you want.
Test dataintermediateExamples kept hidden, used only to check the model at the end.
ThresholdintermediateThe line the total must cross for the neuron to say “yes”.
TrainingintermediateThe whole process of adjusting a model’s weights using examples.
Training databeginnerThe examples an AI learns from.
TransformeradvancedThe neural network design behind most modern chatbots, which uses attention to weigh which words matter.
UnderfittingintermediateA rule too simple to capture the pattern, even in the training data.
Unlabelled databeginnerExamples with no right answers attached.
Unsupervised learningbeginnerFinding patterns without being told the answers.
WeightintermediateHow much an input counts. It can be positive or negative.
Weighted sumintermediateEach input times its weight, all added up.

💬 Chatbots and generative AI

LLMs, prompts, images, voices and why they make mistakes.

AI watermarkintermediateA visible label or hidden signal showing that content was made with AI.
Base modeladvancedA model after pre-training, before it learns to be an assistant.
ChatbotbeginnerAn app you can talk or type to, often powered by a language model.
ContextbeginnerThe words before, which help make a better guess.
Context windowintermediateHow much text a model can consider at once, including your conversation and files.
DimensionadvancedOne of the numbers in a vector; one direction on the map.
DistanceintermediateHow far apart two words are on the map. Small distance, similar meaning.
EmbeddingintermediateA list of numbers that gives a word (or picture, or sound) an address on a map of meanings.
Fact-checkbeginnerTo check whether something is true using a trusted source.
Fine-tuningintermediateExtra training on carefully chosen examples to teach a new skill.
Generative AIbeginnerAI that makes new pictures, text, music or video.
GroundingintermediateTying an AI's answer to real sources, so it can be checked.
GuardrailsintermediateProtections that help an AI refuse harmful requests.
HallucinationbeginnerWhen an AI says something made up as if it were true.
Human feedbackintermediatePeople comparing answers so the model learns which are better.
Image generationbeginnerAI that creates new pictures from a text description.
Knowledge cutoffbeginnerThe date after which a model has no training information.
Language modelbeginnerA program that has learned patterns in language and predicts words.
Large language model (LLM)beginnerAn AI model trained on huge amounts of text to understand and write language.
MultilingualintermediateWorking across many languages.
MultimodalintermediateAI that works with more than one kind of input or output, like text, pictures, voice and video.
Next-word predictionbeginnerGuessing the word most likely to come next.
NoisebeginnerRandom dots with no picture in them.
Open-weight modelintermediateA model whose trained weights are published, so anyone can download and run it.
Pre-trainingintermediateThe first stage, learning from a huge amount of text by guessing the next word.
ProbabilitybeginnerHow likely something is to happen.
PromptbeginnerThe words you give an AI to tell it what to make.
Prompt engineeringbeginnerThe skill of writing clear instructions so AI gives useful answers.
Reasoning modelintermediateA model that works through a problem step by step before answering: slower, but better at maths, logic and planning.
Retrieval-augmented generation (RAG)advancedThe AI first looks up trusted documents, then answers using what it found.
Semantic searchintermediateSearching by meaning instead of exact words.
System promptintermediateHidden instructions an app gives the AI before your message, setting its role and rules.
TemperatureintermediateA setting that controls how random or creative a model's answers are.
TokenintermediateA small piece of text, a word or part of a word, that a language model reads and writes.
VectorintermediateA list of numbers, like (14, 16). Also an arrow with a length and direction.
VerifybeginnerTo check that something is true.

📱 AI in daily life

Cameras, feeds, robots and assistants around you.

ActuatorbeginnerA part that makes a robot move, like a motor.
Collaborative filteringadvancedSuggesting things that people similar to you liked.
Computer visionbeginnerTeaching computers to understand pictures and videos.
ControllerbeginnerThe computer that decides what the robot does.
EdgebeginnerA place in a picture where colours change suddenly.
Emergency stopbeginnerA button that stops a robot at once.
Face recognitionbeginnerAI that identifies or verifies a person from their face.
Feedback loopbeginnerSense, think, act, then check and do it again.
GripperbeginnerA robot’s “hand” for holding things.
ObjectiveintermediateThe goal a system is designed to push up, like watch time.
PersonalisationintermediateGiving each person a different, tailored experience.
PixelbeginnerOne tiny single-colour square in a digital picture.
Recommendation systemintermediateAI that chooses what to show you next.
ResolutionbeginnerHow many pixels a picture has. More pixels, more detail.
RGBbeginnerRed, green and blue, the three lights mixed to make every screen colour.
RobotbeginnerA machine that senses, thinks and acts on its own.
SensorbeginnerA part that collects information, like a camera or touch sensor.
SignalintermediateA clue about you, like a click, a skip or how long you watched.
Speech recognitionbeginnerAI that turns spoken words into text (also called speech-to-text).
Text-to-speechbeginnerAI that reads text aloud in a human-sounding voice.
Voice assistantbeginnerAn assistant you talk to, which hears, understands and replies by voice.

⚙️ Agents and automation

AI that takes actions and connects your apps.

AI agentintermediateAn AI that can take actions, like sending messages or paying.
Automation triggerbeginnerThe event that starts an automation.
Automation workflowbeginnerA chain of steps that runs by itself when something happens.
Browser agentadvancedAn AI that operates a web browser for you: clicking, typing and filling forms.
CopilotbeginnerAn AI assistant built into a tool you already use, helping as you work.
Human in the loopintermediateA person who must approve important actions.
Model Context Protocol (MCP)advancedAn open standard that lets AI assistants connect to apps and data, like a universal plug.
Multi-agent systemadvancedSeveral AI agents with different jobs working together on one task.
Tool use (function calling)intermediateWhen an AI asks an app to run an action for it, like a search, a calculation or sending a message.
WebhookintermediateA message one app sends to another the moment something happens.

🛡️ AI and security

How AI gets attacked, how scammers use it, and how to defend.

.gitignoreintermediateA file that tells Git which files and patterns to never track or upload.
3-2-1 ruleintermediateA backup guideline of 3 copies, on 2 storage types, with 1 copy kept physically separate.
Advance-fee fraudintermediatePaying a small amount upfront for a bigger reward that was never going to arrive.
Adversarial exampleadvancedAn input changed slightly, often invisibly to people, to fool an AI.
AdwarebeginnerMalware built specifically to flood a device with unwanted ads.
Agentic AIintermediateAn AI system that can take real actions, not just generate text responses.
AI supply chain attackadvancedTampering with a model, dataset or code library before people download and use it.
Alert fatigueintermediateThe tendency to ignore warnings after being shown too many false or low-value ones.
AttachmentbeginnerA file sent along with an email or message.
Attack surfaceintermediateEverything about a system that could possibly be attacked or misused.
AuditbeginnerA careful check of habits or systems against what's actually recommended.
AuthenticationintermediateConfirming who someone is, usually via a password and sometimes a second factor.
AuthorizationintermediateConfirming what an already-authenticated person is allowed to do or see.
AvailabilityintermediateEnsuring authorised people can actually access information or systems when they need to.
BackdooradvancedA hidden trigger that makes a model misbehave only when it appears.
BackupbeginnerAn extra copy of your files, kept safe in case the original is lost or damaged.
Base imageintermediateThe foundational layer a container is built on top of.
Baseline behaviourintermediateWhat's normal for an account or system, used to recognise what isn't.
Blast radiusintermediateHow far the impact of a compromise or mistake can spread from its starting point.
Blue teamintermediateThe defenders who detect, respond to, and fix what the red team finds.
BotnetintermediateA large network of compromised devices controlled together, often without their owners' knowledge.
Branch protectionintermediateRepository rules requiring review or checks before code can merge into a main branch.
Broken object-level authorization (BOLA)intermediateAn API that checks login but not whether the specific requested record belongs to the requester.
Bug bounty programmeintermediateA company's published invitation to test specific systems, often with a reward for valid findings.
Captive portalbeginnerThe sign-in page a public wifi network shows before granting access.
CIA triadintermediateConfidentiality, Integrity, and Availability, the three foundational goals of information security.
CNAPPintermediateCloud-native application protection platform, an umbrella term for continuous cloud security posture tooling.
Collective riskbeginnerWhen one person's unsafe habit can end up affecting a whole group or community.
ConfidentialityintermediateEnsuring only authorised people can see or access information.
ConsentintermediateA clear, specific agreement to a particular use of your data, not a vague blanket approval.
Container escapeintermediateA process breaking out of its container boundary to affect the underlying host.
ContainmentintermediateIsolating an infected device to stop an attack from spreading further.
CredentialbeginnerThe username-and-password (or similar) pair that lets you into an account.
Credential reusebeginnerUsing the same password on more than one account.
Credential rotationintermediateReplacing a leaked key or password with a fresh one, rather than only hiding the old one.
Credential stuffingintermediateAutomatically trying leaked email-password pairs from one breach against many other sites.
CTF (Capture The Flag)intermediateA legal competition built around finding planted security flaws in purpose-built systems.
CVSSintermediateCommon Vulnerability Scoring System, a standard way of rating vulnerability severity.
Cyber hygienebeginnerThe everyday habits that keep you safer online, the way regular habits keep you healthy offline.
Cybercrime helpline (1930) / cybercrime.gov.inintermediateIndia's official channels for reporting online fraud, including fake job offers.
Cybercrime helpline, 1930beginnerIndia's national number for reporting cyber fraud.
Dark patternbeginnerA design built on purpose to push you toward one choice, like a huge "Allow All" button.
DASTintermediateDynamic application security testing, probing a running application like an external attacker would.
Data collectionbeginnerThe gathering of information about how you use an app.
Data fiduciaryintermediateAn organisation entrusted with collecting and handling someone's personal data responsibly.
Data leakageintermediateWhen an AI reveals private information it saw in training or in a chat.
Data minimisationintermediateCollecting only the data actually needed for a service to function.
Data poisoningintermediateAdding bad examples to training data to change what an AI learns.
Data principalintermediateThe person whose personal data is being collected or used.
Data provenanceadvancedA record of where each piece of data came from.
DeepfakebeginnerA fake photo or video of a real person, made with AI.
Default credentialsintermediateFactory-set usernames and passwords, identical across every unit of a device.
Defence in depthintermediateUsing many layers of protection, not just one.
Dependency confusionintermediateAn attack where a malicious package is mistaken for, or substituted in place of, a legitimate one by name.
Domain spoofingintermediateMaking an email look like it's from a trusted company by faking the display name, while the real sending address is different.
DPDP Act (Digital Personal Data Protection Act, 2023)intermediateIndia's core law governing personal data collection and use.
EDRintermediateEndpoint detection and response, tooling that watches devices and servers for suspicious activity.
EncodingintermediateRepresenting data in a different, instantly reversible format.
EncryptionbeginnerScrambling data so only the intended recipient can read it.
Environment variableintermediateA way to supply a secret to running code without ever writing it into the code itself.
EscalationintermediatePassing a confirmed concern to someone with more authority or context to act on it.
Evil twinintermediateA fake wifi access point set up with a name close enough to a real one to trick people into connecting.
Evil twin networkbeginnerA fake wifi hotspot built to look like a trusted one.
Excessive agencyintermediateAn AI agent having more permission or capability than its task actually requires.
Excessive data exposureintermediateAn API returning more fields than the client actually needs, relying on the client to filter.
Expiry (exp)intermediateThe field declaring when a token should stop being valid.
False positiveintermediateAn alert that looks suspicious automatically but turns out to be harmless.
Finding chainintermediateCombining several smaller weaknesses into one larger, more serious exploit path.
FirmwareintermediateThe built-in software that runs a device, updated by the manufacturer to fix bugs and flaws.
Flat networkintermediateA network with no internal boundaries, where most things can reach most other things.
FootholdintermediateA first point of access an attacker uses to reach further into a larger network.
FoundationintermediateThe shared core knowledge that makes any later specialisation easier to learn.
Fraud disputeintermediateA formal report to a bank or payment provider asking to reverse an unauthorised or misdirected payment.
Git historyintermediateThe full, permanent record of every past version of a repository, not just its current state.
Ground truthintermediateFacts that have been carefully checked in the real world.
Hardcoded secretintermediateA password, key, or token written directly into source code instead of kept separately.
HashingintermediateTurning data into a fixed, one-way fingerprint used to check a match.
HTTPSbeginnerA secure, encrypted connection to a website, shown by a padlock in the address bar.
HTTPS / TLSintermediateEncryption that protects your connection to a specific site, shown as a padlock in the browser.
IAM (identity and access management)intermediateThe system of roles and permissions controlling who can access what in the cloud.
Identity and access management (IAM)intermediateThe discipline of managing who or what can access a system, and what they can do.
Identity providerintermediateA central service other systems rely on to verify who someone is.
IDOR (Insecure Direct Object Reference)intermediateA flaw where a system trusts an ID in a request instead of checking who's allowed to see it.
Image scanningintermediateAutomatically checking a container image's layers for known vulnerabilities.
Impossible travelintermediateA detection technique flagging logins from two locations too far apart to reach in the time between them.
In-app purchasebeginnerSpending real money inside an app, often disguised as game currency.
Incident responseintermediateThe organised process of containing, investigating, and recovering from a security incident.
IntegrityintermediateEnsuring information hasn't been improperly altered or tampered with.
IoT (Internet of Things)intermediateOrdinary objects, plugs, bulbs, cameras, connected to the internet.
IT Act (India)intermediateIndia's primary law covering unauthorised computer access and related offences.
JailbreakintermediateA trick that gets a chatbot to ignore its safety rules.
JailbreakingintermediateTricking an AI model into bypassing its own safety guidelines.
JWT (JSON Web Token)intermediateA compact, signed token format commonly used to represent a logged-in session.
KYCbeginnerKnow Your Customer, the paperwork banks use to confirm who you are.
Lateral movementintermediateAn attacker moving from one compromised system to another inside a network.
Least privilegebeginnerGiving an app, or a person, only the access they truly need, nothing more.
MalwarebeginnerMalicious software, built to harm, spy on, or exploit a device.
MFA (multi-factor authentication)intermediateRequiring a second proof of identity beyond just a password.
Model extractionadvancedCopying how a model behaves by sending it huge numbers of questions.
Multi-factor authentication (MFA)intermediateA second proof of identity beyond just a password.
Network scanintermediateChecking which devices are present and reachable on a given network.
Network segmentationintermediateDividing a network into zones with controlled access between them.
Non-root userintermediateRunning a process with restricted privileges rather than full administrative access.
Official app storebeginnerA store where apps are checked before being listed, safer than installing from outside it.
Offline backupbeginnerA copy kept somewhere not connected to the internet, safe from online attacks.
OTPbeginnerA one-time password, meant to be used once, by you, for one single action.
PassphrasebeginnerA longer password made of several words, built to be memorable to you and hard for anyone else to guess.
PasswordbeginnerA secret word or phrase that proves it's really you.
Password managerbeginnerAn app that generates and remembers a strong, unique password for every account you have.
Patch cadenceintermediateA regular, scheduled rhythm for applying security updates.
Payee / payerintermediateThe person or business receiving a payment, and the person sending it.
PayloadintermediateThe data section of a JWT, encoded but not encrypted, readable by anyone holding the token.
Payment confirmation screenintermediateThe step showing the receiving name before a payment is finalised.
PerimeterintermediateThe boundary a security approach is built around, once the network edge, now increasingly identity.
PermissionbeginnerAn app's request to access a part of your phone, like your camera or contacts.
Permission creepbeginnerWhen an app asks for more access than its actual job needs.
Permission driftintermediateConfiguration gradually becoming more permissive than originally intended, often unnoticed.
Permission scopeintermediateExactly what data or actions a tool has requested and been granted access to.
PhishingbeginnerA fake message, sent widely, designed to trick you into giving away information or money.
PINbeginnerA short, usually numbers-only password, often 4 to 6 digits.
Placement CellintermediateA college office that manages and verifies genuine job and internship opportunities for students.
Pre-commit hookintermediateA script that runs automatically before a commit completes, able to block it if something looks wrong.
Prompt injectionintermediateManipulating an AI system's behaviour through cleverly worded input text.
Proof-of-concept exploitintermediateCode demonstrating that a vulnerability can actually be exploited.
Public wifibeginnerA shared wireless network anyone nearby can join, at a mela, cafe, or airport.
QR code swappingintermediatePhysically replacing a legitimate payment code with a fraudulent one.
QuishingintermediatePhishing carried out through a QR code instead of a link or message.
RansomwarebeginnerSoftware that locks or scrambles your files and demands payment to release them.
Rate limitintermediateA cap on how much one person or account can send.
Rate limitingintermediateRestricting how many requests a client can make in a given time, to prevent abuse or scraping.
RBAC / ABACintermediateRole-based and attribute-based access control, two models for deciding who's allowed to do what.
Recruitment fraudintermediateA fake job or internship offer designed to look like a genuine hiring process.
Red flagbeginnerA warning sign that something might not be what it claims to be.
Red teamintermediateA team authorised to simulate real attacks to find weaknesses.
Refresh tokenintermediateA longer-lived, separately managed token used to obtain new short-lived access tokens.
Responsible disclosureintermediateReporting a security flaw quietly to whoever can fix it, instead of exploiting or publicising it.
RestorebeginnerTo bring back files from a backup after losing the originals.
RevocationintermediateThe ability to invalidate a specific already-issued token before its natural expiry.
Right to erasureintermediateThe ability to request that an organisation delete personal data it holds about you.
Rogue access pointintermediateAny unauthorised wifi hardware set up where it doesn't belong, evil twin or otherwise.
Roll backintermediateReturn to an earlier, trusted version.
Root causeintermediateThe underlying reason behind a problem, as opposed to its visible symptoms.
Rules of engagementintermediateThe agreed scope and boundaries for an authorised security test.
Safe harbourintermediateA promise not to pursue legal action against researchers who stay within an agreed, published testing scope.
SASTintermediateStatic application security testing, scanning source code itself for risky patterns.
SBOMintermediateSoftware bill of materials, a complete list of every component a piece of software depends on.
SCAintermediateSoftware composition analysis, checking third-party dependencies against known vulnerability databases.
ScarewarebeginnerA scam that claims your device is infected, to frighten you into clicking.
Secret scanningintermediateAutomated tools that watch public code for exposed credentials.
Secret sprawlintermediateCredentials scattered across many places, repos, notes, chats, with no central record of where they all live.
Secrets manager / vaultintermediateA system that stores credentials centrally and issues them securely, often with short lifespans.
Secure SDLCintermediateBuilding security checks into every stage of the software development lifecycle, not bolting them on at the end.
Security championintermediateA team member, not on the security team, trained to run lightweight security checks on their own team's work.
Security mindsetintermediateApplying these questions as a habit of thought across any role, not only security-titled jobs.
Shadow AIintermediatePeople using AI tools at work that their organisation hasn't approved.
Shadow APIintermediateA live API version or endpoint nobody is actively tracking or securing.
Shadow ITintermediateSoftware, tools, or services in use without formal approval or security review.
Shared responsibility modelintermediateThe idea, especially in cloud computing, that security is split between a provider and its customer, not owned entirely by either.
Shelf referenceintermediateA resource meant to be revisited repeatedly, not read once and shelved for good.
Shoulder surfingbeginnerSomeone learning your password by watching or overhearing you, not by guessing it.
SignatureintermediateThe part of a JWT that proves it wasn't tampered with after being issued, assuming the signing key stays secret.
SOC (Security Operations Centre)intermediateA team that watches systems and networks for suspicious activity.
Social engineeringbeginnerTricking a person directly, often using trust, politeness or fear, instead of tricking a computer.
SpecialisationintermediateDeep focus in one particular area, usually built on a broader foundation.
SpoofingbeginnerFaking a caller ID, sender name or address so it looks like it comes from somewhere trustworthy.
STRIDEintermediateA six-question framework (Spoofing, Tampering, Repudiation, Information disclosure, Denial of service, Elevation of privilege) for threat modeling.
Threat modelingintermediateSystematically asking what could go wrong with a system, and for whom, before or while it's built.
Threat modellingintermediateAsking in advance who could misuse something and what the worst outcome could be.
Three two one rulebeginnerKeep three copies, on two kinds of storage, with one copy kept somewhere else.
TLS certificateintermediateA credential proving a server's identity and enabling encrypted connections.
Too-good-to-be-true offerbeginnerSomething valuable offered for nothing, with no real reason given.
TriggerintermediateThe secret signal, like the yellow sticker, that activates a backdoor.
Two-factor authentication (2FA)beginnerA second, different check, like a one-time code, on top of your password.
TyposquattingbeginnerA web address built to look almost exactly like a real one.
Unsolicited downloadbeginnerA file that arrives or installs without being properly requested.
Untrusted dataintermediateContent from outside that might contain tricks.
UPI (Unified Payments Interface)intermediateIndia's widely used instant payment system, often used through QR codes.
Urgency tacticbeginnerA countdown, a threat or a deadline, used on purpose to stop you thinking clearly.
Voice cloningbeginnerUsing AI to copy a person's voice from a short recording.
VPNbeginnerA tool that wraps your connection in extra encryption, useful on networks you don't fully trust.
White hat / black hat / grey hatintermediateShorthand for ethical, malicious, and unauthorised-but-not-malicious hacking.
Zero trustintermediateVerifying every request explicitly rather than trusting by network location.

⚖️ Privacy, law and compliance

Your data rights and the rules for AI, in India and beyond.

AI governanceadvancedAn organisation's rules, roles and checks for using AI safely and legally.
Consent managerintermediateUnder India's DPDP Rules, a registered Indian company that lets you give, review and withdraw consent across many apps in one place.
Copyright and AIintermediateThe open legal questions about who owns AI output, and whether AI may learn from copyrighted work.
Data Protection Board of IndiaintermediateThe independent body that handles complaints and penalties under the DPDP Act, with an online portal.
Digital footprintbeginnerThe lasting trail of information you leave behind online over time.
EU AI ActadvancedEurope's law that sorts AI systems by risk, with strict rules for high-risk uses and bans on some.
IT RulesintermediateIndia's rules for online platforms, covering takedowns, grievances and, since 2026, AI-made content.
Jigsaw effectbeginnerSmall, separately harmless facts combining into one complete, identifying picture.
OversharingbeginnerGiving away more personal detail than you realise, often a little at a time.
Personal databeginnerInformation about a person, like their face, voice or name.
PrivacybeginnerThe right to decide who sees your information.
Privacy by designintermediateProtecting people’s information from the very first step of building.
Privacy settingsbeginnerThe controls that decide who can see what you post.
Private accountbeginnerOnly followers you approve can see what you post.
Public accountbeginnerAnyone can see your posts and follow you without approval.
Synthetically generated informationintermediateIndia's legal term (IT Rules, 2026) for audio, images or video made or altered by AI to look real. It must be labelled.

🤝 Responsible AI

Fairness, transparency and keeping people in charge.

AccountabilitybeginnerBeing responsible for what happens.
AI alignmentadvancedMaking sure an AI's goals and behaviour match what people actually want.
AI safetyintermediateResearch and practice to stop AI systems causing harm, now or as they get more powerful.
Balanced databeginnerExamples that fairly cover every group.
BiasbeginnerWhen an AI is regularly wrong or unfair for one group.
Block and reportbeginnerStopping contact from someone and telling the platform about harmful behaviour.
BystanderbeginnerSomeone who sees something harmful happen and chooses to act, or not to.
CyberbullyingbeginnerUsing a screen, post, or message to mock, embarrass, or target a real person.
DatasetbeginnerThe whole collection of examples an AI learns from.
Digital empathybeginnerThinking about the real person behind every post before you share it.
ExplainabilityintermediateBeing able to understand why an AI gave a particular answer.
Fact-checkingbeginnerThe process of verifying whether a claim is actually true before trusting or sharing it.
Fairness testingbeginnerChecking an AI’s results for each group separately.
Filter bubbleintermediateSeeing mostly things that match what you already like or believe.
Forward chainbeginnerHow a message multiplies as more and more people pass it along.
Forwarded many timesbeginnerA label some apps add to messages that have been passed along many times already.
Human reviewbeginnerA person checking an AI’s important decisions.
MisinformationbeginnerFalse or misleading information, shared whether or not the person sharing it knows it's false.
Out of contextbeginnerA real photo or fact paired with a false claim about where, when, or why it happened.
Rabbit holeintermediateBeing led, step by step, into more and more of one kind of content.
Responsible AIbeginnerBuilding and using AI in ways that are fair, safe, honest and accountable.
RumourbeginnerInformation spreading from person to person that hasn't been properly checked.
ScreenshotbeginnerA saved image of a screen that can outlive the original message completely.
SourcebeginnerWhere a piece of information or an image actually came from.
TransparencybeginnerBeing open about when AI is used and how it works.

💼 AI in business

Words you'll hear in offices, startups and boardrooms.

AI washingintermediateExaggerating how much AI a product really uses, for marketing.
Attention economyintermediateA world where apps compete to hold people’s attention.
Domain expertintermediateA person who knows a job or subject deeply, from real experience.
Foundation modelintermediateA large, general model that many products are built on.
Proof of concept (PoC)beginnerA small trial to check whether an AI idea works before spending big.
Return on investment (ROI)beginnerWhat you gain compared with what you spend.
Software as a service (SaaS)beginnerSoftware you rent online by subscription instead of installing.
Sovereign AIintermediateA country building its own AI models, chips and data centres.
UpskillingbeginnerLearning new skills so you can work well alongside new technology.
Use casebeginnerA specific job a business wants AI to do.
Vendor lock-inintermediateDepending so much on one supplier that switching becomes hard or costly.

💻 Coding and building AI

For people who want to build with AI.

APIintermediateAn application programming interface: a way for one program to talk to another.
DeploymentintermediatePutting software live, so real users can use it.
Distribution shiftadvancedWhen the real world looks different from the data the AI learned from.
Evaluation (evals)intermediateSystematic tests that measure how well an AI system does its job.
Git and GitHubintermediateGit tracks every change to your code; GitHub is a website where developers store and share it.
HackathonintermediateAn event where teams build something that works in a short time.
Open sourcebeginnerSoftware whose code is public, so anyone can read, use and improve it.
PythonbeginnerThe most popular programming language for AI and data work.
SDKintermediateA software development kit: ready-made code that makes it easier to use a service.
Transfer learningintermediateStarting from a model already trained on a big task, and teaching it a new one.
Vector databaseadvancedA database that stores embeddings, so you can search by meaning.
Vibe codingbeginnerBuilding software by describing what you want to an AI and improving it through conversation.

🔌 Chips, cloud and energy

The machines, power and water behind AI.

Carbon footprintintermediateHow much carbon dioxide something adds to the air.
Cloud computingbeginnerRenting computers, storage and AI over the internet instead of owning them.
ComputeintermediateThe computing power needed to train and run AI.
Data centreintermediateA building full of computers that run websites, apps and AI.
GPUintermediateA chip that can do huge numbers of calculations at once; AI uses lots of them.
High-bandwidth memory (HBM)advancedVery fast memory stacked next to AI chips. Making it uses factory capacity that once made phone memory.
InferenceintermediateUsing a trained model to answer questions; a small cost each time, repeated billions of times.
LatencyintermediateThe delay between asking and getting an answer.
ServerintermediateA powerful computer, usually with no screen, that does work for many people over the internet.
Watt-hour (Wh)intermediateA measure of energy; 1,000 Wh is one unit on the electricity bill.