AI Glossary: 120 AI Terms Explained in Plain English

This glossary explains 120 AI terms in plain English, without assuming a technical background. It is written for teachers, students, school leaders and working professionals — the people who need to understand what these words mean in a document or a meeting, not implement them. Terms that appear in CBSE’s Computational Thinking and AI curriculum for Classes 3–8 are marked [CBSE].

If you are working through the 30-day AI literacy roadmap, roughly ten of these terms are all you need in week one. The rest can wait until you meet them.

A

• Agent — an AI system that can take a sequence of actions towards a goal, rather than just answering one question. Booking a flight end to end, instead of telling you how to book one.
• AGI (Artificial General Intelligence) — a hypothetical AI that matches human capability across essentially all intellectual tasks. It does not exist; treat confident claims about its arrival date as opinion.
• Algorithm [CBSE] — a precise sequence of steps for solving a problem. A recipe is an algorithm. Not all algorithms involve AI.
• Alignment — the problem of making an AI system actually pursue what its users intend, rather than a literal or distorted version of it.
• Annotation — the human work of labelling data so a model can learn from it. Someone drew boxes around ten thousand cats.
• Anthropomorphism — attributing human feelings or intentions to a machine. A model that says “I’m happy to help” is producing text, not reporting a mood.
• API (Application Programming Interface) — a way for one piece of software to use another’s capabilities. How an app gets answers from an AI model without hosting it.
• Artificial Intelligence [CBSE] — the broad field of building systems that perform tasks normally requiring human intelligence: recognising, predicting, deciding, generating.
• Attention — the mechanism that lets a model weigh which parts of the input matter most for each part of the output. The core idea behind transformers.
• Augmented intelligence — framing AI as amplifying human decision-making rather than replacing it. Often a positioning term rather than a technical one.
• Autonomous system — a system that operates without step-by-step human control. Degree of autonomy matters far more than the label.

B

• Backpropagation — the training method that works backwards through a neural network adjusting weights to reduce error. How networks learn.
• Benchmark — a standard test used to compare models. Useful, but models can be tuned to score well on benchmarks without being better in practice.
• Bias [CBSE] — systematic unfairness in a model’s output, usually inherited from patterns in its training data rather than inserted deliberately.
• Black box — a system whose internal reasoning cannot be inspected in a human-readable way. Most large models are black boxes.
• Bounding box — the rectangle drawn around an object in an image for computer vision training or detection.

C

• Chain of thought — prompting a model to work through intermediate steps before answering. Often improves accuracy on reasoning problems.
• Chatbot — a conversational interface to a model. The interface, not the intelligence.
• Classification [CBSE] — predicting which category something belongs to. Spam or not spam; pass or fail.
• Clustering [CBSE] — grouping similar items without being told the groups in advance. Used to discover structure you did not know was there.
• Compute — the raw processing power used to train or run a model, usually measured in GPU hours. A major cost and a major constraint.
• Computational thinking [CBSE] — the problem-solving approach built on four pillars: decomposition, pattern recognition, abstraction and algorithmic thinking. The foundation CBSE teaches before AI.
• Computer vision [CBSE] — the field concerned with making machines interpret images and video.
• Context window — how much text a model can consider at once. Exceed it and the earliest material is effectively forgotten.
• Convolutional neural network (CNN) — a network architecture designed for images, which detects features like edges and shapes in layers.
• Copilot — a general term for an AI assistant embedded inside a tool you already use, suggesting rather than deciding.
• Corpus — a large body of text used for training or analysis.

D

• Data augmentation — artificially expanding a dataset by modifying existing examples, such as rotating or cropping images.
• Data drift — when the real-world data a deployed model sees gradually stops resembling its training data, and accuracy quietly degrades.
• Dataset [CBSE] — a structured collection of data used to train, validate or test a model.
• Decision tree — a model that makes predictions through a branching series of yes/no questions. Easy to interpret, which is why it is still widely used.
• Deep learning — machine learning using neural networks with many layers. The approach behind most modern AI.
• Deepfake [CBSE] — synthetic audio, image or video that convincingly depicts a real person doing or saying something they did not.
• Diffusion model — the architecture behind most image generators. It learns to reverse a process of adding noise, effectively sculpting an image out of static.
• Dimensionality reduction — simplifying data with many variables into fewer, while keeping most of the useful information.

E

• Edge AI — running a model on a local device — phone, camera, sensor — rather than in the cloud. Faster and more private.
• Embedding — a representation of text, images or other data as a list of numbers, arranged so similar things sit close together.
• Encoder — the part of a model that converts input into an internal numerical representation.
• Ensemble — combining several models and pooling their predictions, usually more accurate than any single one.
• Epoch — one complete pass through the training data during training.
• Explainable AI (XAI) — techniques for making a model’s decisions understandable to humans. Increasingly a regulatory requirement, not just good practice.

F

• Fairness [CBSE] — whether a system’s outcomes are equitable across groups. There are several mathematical definitions of fairness and they conflict with each other, which is why this is a values question, not only a technical one.
• Feature — an individual measurable property used as input. In predicting exam results, attendance would be a feature.
• Few-shot learning — getting a model to perform a task by showing it a handful of examples in the prompt.
• Fine-tuning — further training a pre-trained model on a narrower dataset so it specialises.
• Foundation model — a large model trained on broad data that can be adapted to many downstream tasks. GPT-class models are foundation models.

G

• GAN (Generative Adversarial Network) — two networks trained against each other, one generating and one detecting fakes. An earlier approach to image generation.
• Generative AI [CBSE] — AI that produces new content — text, images, audio, code — rather than only classifying or predicting.
• GPU (Graphics Processing Unit) — the hardware that makes training large models practical, because it performs many calculations in parallel.
• Gradient descent — the optimisation method that nudges a model’s parameters step by step towards lower error.
• Grounding — tying a model’s output to verifiable sources or real data, to reduce invention.
• Guardrails — the constraints placed around a model’s behaviour to prevent unsafe or off-policy output.

H

• Hallucination — when a model produces confident, fluent output that is factually wrong or entirely invented. Not a bug to be patched; a consequence of how prediction works.
• Human-in-the-loop — a design where a person reviews or approves AI decisions before they take effect.
• Hyperparameter — a training setting chosen by humans rather than learned by the model, such as learning rate.

I

• Inference — running a trained model to get an answer. Training is learning; inference is using.
• Instruction tuning — training a model specifically to follow instructions well, which is what turned raw language models into usable assistants.
• Interpretability — how well humans can understand why a model produced a given output.

J–K

• Jailbreak — a prompt crafted to bypass a model’s safety restrictions.
• K-means — a common clustering algorithm that groups data into k clusters by proximity.
• Knowledge graph — a structured network of entities and the relationships between them.

L

• Label [CBSE] — the correct answer attached to a training example, so the model has something to learn from.
• Large language model (LLM) — a model trained on vast amounts of text to predict likely continuations. ChatGPT, Gemini and Claude are interfaces to LLMs.
• Latency — the delay between asking and receiving an answer.
• Learning rate — how large a step the model takes when adjusting during training. Too high and it overshoots; too low and it never gets there.
• LoRA (Low-Rank Adaptation) — an efficient fine-tuning technique that adapts a model by training a small number of extra parameters.

M

• Machine learning [CBSE] — systems that improve at a task by learning patterns from data, rather than following rules written by a programmer.
• Machine translation — automatic translation between languages.
• Model [CBSE] — the trained artefact that makes predictions. The thing that learned.
• Model card — a short document describing a model’s intended use, training data, limitations and known risks. Worth asking for before adopting any tool in a school.
• MLOps — the practices for deploying, monitoring and maintaining models in production.
• Multimodal — a model that handles more than one type of input or output — text and images and audio together.

N–O

• Natural language processing (NLP) [CBSE] — the field concerned with making machines work with human language.
• Neural network [CBSE] — a model made of layers of connected units, loosely inspired by biological neurons. The inspiration is loose; do not over-read the metaphor.
• Neuron — a single unit in a neural network that takes inputs, weights them and passes on a result.
• Node — a point in a network or graph; often used interchangeably with neuron.
• Open-source model — a model whose weights are publicly released, so it can be run and modified independently.
• Optimisation — the process of adjusting a model to minimise error.
• Output layer — the final layer of a network, which produces the prediction.
• Overfitting [CBSE] — when a model memorises its training data and performs badly on anything new. The equivalent of a student who memorised last year’s paper.

P–Q

• Parameter — one of the internal numbers a model adjusts during training. Large models have billions.
• Pattern recognition [CBSE] — identifying regularities in data. One of the four computational thinking pillars.
• Perceptron — the simplest form of neural unit; historically the starting point of the field.
• Pre-training — the initial, broad training phase on large general data, before any specialisation.
• Prompt — the instruction or question given to a model.
• Prompt engineering — the practice of designing prompts to get better output. In practice, mostly clear task framing.
• Prompt injection — an attack where hidden instructions in content the model reads cause it to ignore its original instructions. A real security risk for any AI tool that browses or reads files.
• Quantisation — reducing the numerical precision of a model’s parameters so it runs faster and smaller, at some cost to quality.
• Query — a request made to a system or database.

R

• RAG (Retrieval-Augmented Generation) — retrieving relevant documents first and having the model answer from them. The standard technique for reducing hallucination.
• Recommendation system [CBSE] — a system that predicts what a user will want next. The most common AI most students encounter daily.
• Recurrent neural network (RNN) — an architecture for sequential data, largely superseded by transformers.
• Red teaming — deliberately attacking or stress-testing a model to find failures before users do.
• Regression [CBSE] — predicting a continuous number rather than a category. Predicting marks, not pass/fail.
• Reinforcement learning — training through trial, error and reward signals rather than labelled examples.
• RLHF (Reinforcement Learning from Human Feedback) — using human preference ratings to shape a model’s behaviour. A major reason modern assistants feel helpful.
• Robotics [CBSE] — the field of building machines that sense, decide and act in the physical world.

S

• Semi-supervised learning — training on a small amount of labelled data plus a large amount of unlabelled data.
• Sentiment analysis — automatically judging whether text expresses a positive, negative or neutral attitude.
• Speech recognition — converting spoken audio into text.
• Supervised learning [CBSE] — learning from labelled examples where the correct answer is provided.
• Synthetic data — artificially generated data used for training when real data is scarce or sensitive.
• System prompt — the hidden instruction that sets a model’s role and rules before a user’s conversation begins.

T

• Temperature — a setting controlling randomness in output. Lower is more predictable, higher is more varied.
• Test set — data held back entirely from training, used to measure real performance.
• Token — the unit a model actually processes: roughly a word or word-fragment. Costs and limits are usually counted in tokens.
• Tokenisation — the process of splitting text into tokens.
• Training data [CBSE] — the data a model learns from. Its biases, gaps and errors become the model’s.
• Transfer learning — reusing a model trained on one task as the starting point for another.
• Transformer — the architecture behind virtually all modern language models, built on the attention mechanism.
• Turing test — a 1950 proposal that a machine could be considered intelligent if a human could not distinguish it from another human in conversation. Historically important, no longer a useful benchmark.

U–Z

• Underfitting — when a model is too simple to capture the real pattern, and performs badly even on training data.
• Unsupervised learning [CBSE] — finding structure in data without labels.
• Validation set — data used during development to tune a model, kept separate from the final test set.
• Vector database — a database that stores embeddings and retrieves items by similarity rather than exact match. The storage layer behind most RAG systems.
• Vision language model (VLM) — a model that handles images and text together, such as answering questions about a photograph.
• Weights — the learned numerical values inside a model that determine its behaviour. “Open weights” means these have been released publicly.
• Word embedding — a numerical representation of a word capturing its meaning relative to other words.
• XGBoost — a widely used algorithm for structured, tabular data. Still outperforms neural networks on many business datasets.
• Zero-shot learning — performing a task with no examples provided, purely from the instruction.

The ten terms to learn first

If 120 terms feels like a lot, it is. These ten carry most of the practical weight, and understanding them properly is the whole of week one of the 30-day roadmap:

1. Model
2. Training data
3. Large language model
4. Hallucination
5. Token
6. Context window
7. Prompt
8. Bias
9. Fine-tuning
10. Inference

Frequently asked questions

Which AI terms do students need to know for CBSE CT & AI?

The terms marked [CBSE] above map to the Classes 3–8 curriculum: computational thinking and its four pillars in the earlier classes, then classification, regression, clustering, computer vision, NLP, bias and the AI project lifecycle from Classes 6 to 8. The full curriculum breakdown is in our CBSE CT & AI guide.

What is the difference between AI, machine learning and deep learning?

They nest. AI is the broad field. Machine learning is the subset where systems learn from data instead of following written rules. Deep learning is the subset of machine learning using many-layered neural networks. Every deep learning system is machine learning; not every AI system is.

Why do people say AI “hallucinates” rather than “makes mistakes”?

Because the failure is distinctive. A mistake implies the system tried to retrieve something and got it wrong. A language model was never retrieving — it was generating plausible text. The invented citation is not a lookup failure; it is the system working exactly as designed on a question where plausible and true diverge.

Do I need to know all 120 terms to be AI literate?

No. Vocabulary is the smallest part of AI literacy. Ten terms plus the judgement to verify output beats 120 terms without it.

Can I use this glossary in my classroom?

Yes, freely, with attribution to this page. If you would like a printable version aligned to a specific class level, our CT & AI curriculum resources include class-wise vocabulary lists.

---

Compiled by Piyush Wairale (IIT Madras), founder of PiyushAI Edtech. Related reading: What is AI literacy? · The 8 core AI literacy skills · AI literacy statistics 2026

Sources: CBSE Computational Thinking & AI Teacher Handbook 2026–27 (for terms marked [CBSE]); OECD/EC AILit Framework; UNESCO AI Competency Framework for Students.