Open your inbox. Your email service has already sorted it — real messages in the main folder, junk held in the spam bin. That sorting happened before you saw a single subject line. Now open a chatbot and type a question. Within seconds, a paragraph of original text appears — not pulled from a database, not copied from anywhere. Built, word by word, right there.
Two AI tools. Two completely different jobs. The spam filter decided — it drew a line between what belongs and what does not. The chatbot created — it produced something new that did not exist before your prompt.
You named this contrast informally in Class 9. Here you give it its formal definition, understand the mechanism that makes generation work, and examine why that mechanism carries risks the filter never does.
The spam filter classifies. Feed it an email and it tells you which side of a line that email falls on — spam or not-spam. It does not produce email. The chatbot does the opposite: it takes your prompt and produces text that is new. How does the production part work? A generative model does not hold a library of pre-written answers. What it holds is a learned statistical distribution — a compressed record of what kinds of sequences appeared in the training data, and how often. At generation time, the model does not look up an answer. It samples from that distribution. Given your prompt as context, it asks: what continuation is plausible here? Then it draws one. The output is not fetched; it is built by sampling.
Weaker. Retrieval (database lookup). The system holds every answer ever written. When you query it, it finds the closest match and returns it — the same query gives the same answer every time. If the answer is not in the database, it returns nothing. This is how a search engine works. It is not how a generative model works.
Stronger. Generation (distribution sampling). The model holds no database of answers — only a compressed record of what kinds of sequences appeared in training. Given your prompt, it asks: what continuation is plausible? Then it draws one. Submit the same prompt twice and you will often get different paragraphs, because sampling a distribution does not give the same draw every time. The output is constructed fresh; it is not fetched.
Close this page. On a blank line, write one sentence explaining how a generative model produces output — without using the words "database" or "retrieves." Then open back and check: did your sentence say "distribution"? Did it say "samples"? Those two words are the exam anchor. If either is missing, write the sentence again.
"It retrieves from a database — isn't that what the internet search training means?"
Search engines retrieve. Generative models do not. What the model holds after training is not the training data — it is a compressed record of patterns in that data: which sequences appeared, and how often. At generation time it samples from that pattern record, it does not search it. A database answer has a traceable source; a generated answer does not — it is a pattern-based construction, assembled fresh each time. The distinction matters for copyright, for attribution, and for understanding hallucination.
Here is the key point at Class 12 depth. The generation process is probabilistic. Submit the same prompt twice and you will often get different outputs, because sampling a distribution does not give the same draw every time. A spam filter on the same email gives the same verdict — classification is fixed by design. A chatbot on the same question gives you two different paragraphs — generation is probabilistic by design. That difference — sampling versus boundary-drawing — is why hallucination is not a bug in generative models. It is a result of the mechanism itself.
Discriminative model — "A model that learns the boundary between classes — given an input, it outputs the class (label or decision) that the input belongs to. It asks: given this input, which category is it? Discriminative models learn P(label | input)."
Generative model — "A model that learns the distribution of the input data itself — it can generate new samples that are plausible under that distribution. It asks: given this class (or prompt), what would a new instance look like? Generative models learn P(input) or P(input | class)."
Read those two expressions carefully, because exam markers do. P(label | input) means the probability of a label, given an input — the spam filter computes this every time it scores an email. P(input) means the probability of an input itself, given nothing — the chatbot learns this, so it can produce new text that looks as if it belongs to the distribution. The question each model answers is different. The spam filter asks "which class?" The chatbot asks "what next?"
Weaker. Spam filter (discriminative). Question: which class? Learns: P(label | input). Output: spam or not-spam. The filter is excellent at its task — it does not need to generate emails. A discriminative model cannot produce new content; that is not what it was designed to do.
Stronger. Chatbot (generative). Question: what does a new continuation look like? Learns: P(input). Output: a new text instance constructed by sampling. The chatbot cannot reliably classify spam — it was not designed to draw boundaries. Neither model is superior; each excels at its designed task.
"Generative AI is more advanced than discriminative AI — isn't that obvious?"
This is the ranking trap — it appears in 843 exams every year and always costs marks. Generative AI is not more advanced: it is designed for a different task. A spam classifier is excellent at classifying — it does not need to generate. A text generator cannot reliably classify spam. The two are designed for fundamentally different jobs; neither is universally superior. Write "purpose difference, not a ranking" before you write anything else on a 4-mark contrast question.
Your chatbot produces text. But can generative AI produce anything other than text? Yes — as long as the training data is in the right form. The same principle — learn a distribution, sample from it — applies when the training data is images, audio, or video instead of text. The output modality follows the training data. Generative AI spans five major domains. In each, a prompt in one modality drives the output in the same or another modality: text (essays, summaries, chatbot replies, code), images (art, product mockups, synthetic photos), audio (speech synthesis, music composition), video (animation, synthetic footage, training data), and code (autocomplete, bug suggestions, test generation).
Activity A — Canva AI image generation. Sign up for Canva; use the AI image tool to generate three images from text prompts. For each: (a) write the prompt you used; (b) describe the image generated; (c) identify one visual anomaly or unexpected element — the hallucination-spotting habit applied to images. Collect the three prompt-image-anomaly records. This is the official syllabus activity "Signing up for Canva." The artifact is a three-record log (prompt / image description / anomaly noted).
ai-genai-mechanism.lab-canva
Activity A — Canva AI image generation
Canva AI
A three-record log — for each of three images: the exact prompt used, a description of the image, and one visual anomaly noted (hallucination-spotting applied to images).
Official syllabus activity 'Signing up for Canva' — practical marks, not a theory question.
Before you run Activity A, write down your prediction: for the prompt "a child studying mathematics on the Moon," name one visual anomaly you expect the generator to produce. Run the activity. Did the actual anomaly match your prediction? What does that tell you about what the model understands (and does not) about the physical world?
Activity B — Animaker AI video generation. Use Animaker's AI video generation tool to create a short (under 60-second) video from a text script of 3–5 sentences. Observe: (a) how the tool translates text to visuals; (b) which parts of the script generated expected visuals vs unexpected ones. Write a one-paragraph reflection on "what the tool understood and what it did not." The artifact is a short reflection paragraph plus the video output.
ai-genai-mechanism.lab-animaker
Activity B — Animaker AI video generation
Animaker AI
A one-paragraph reflection ('what the tool understood and what it did not') plus the generated video output.
Official syllabus activity — practical marks, not a theory question.
After Activity B: write one sentence that connects what you observed in Animaker to the generative mechanism. Your sentence should include the word "distribution" or "sampling." If you cannot write that sentence yet, re-read the working-of-GenAI section.
Large Language Model (LLM) — "A large language model is a generative AI model trained on a very large corpus of text (billions of tokens) that can generate coherent text, translate languages, summarise documents, write code, and answer questions. It predicts the next token in a sequence given all preceding tokens, using patterns learned during training."
The mechanism behind that definition is worth reading closely. An LLM works on tokens — words or sub-word pieces. At each position, the model looks at all the tokens before it and asks: what token is most plausible next? It samples from that distribution, places the token, then repeats — token by token — until the response is complete. Nowhere in this process does the model check a fact source or ask whether the chosen token is true. It checks whether the token is plausible. Plausible and true are not the same thing.
One further mechanism is worth naming: the attention mechanism. It is what allows an LLM to handle long context — to "remember," three paragraphs later, that the subject of the sentence was "the experiment" and not "the result." The model learns to weigh how much each earlier token matters when predicting the next one. This is a content note, not a tested sub-unit; the exam does not ask you to derive the maths of attention.
Now for the "large" part. An LLM is large in two senses. The training data is a corpus covering a huge share of human-written text from the internet and books. The model size is billions of learned numerical parameters. The fluency that surprises people is a result of that scale: enough data to encounter almost every topic, enough parameters to encode the patterns. Scale does not add understanding — it compresses patterns. This distinction is load-bearing for the exam.
An LLM does not understand language the way a person does. An LLM does not understand anything. It holds numerical weights that encode patterns over tokens. Fluency is the output of extreme scale — vast training plus vast parameters. It is not a sign of comprehension. An LLM cannot be asked what it "means" to say something, because it has no meaning states. A human pauses on a word because it landed wrong; an LLM has no landing at all.
Weaker. Human comprehension. A person reading a sentence has meaning states: they can say what a word "means" to them, pause on an unexpected word, and flag a sentence that feels wrong. They understand — they can be asked to paraphrase with different words and explain why a statement is correct or false.
Stronger. LLM fluency. The model holds numerical weights. It produces the next plausible token — that is all. It has no meaning states, no comprehension, no capacity to say why. Its fluency is a result of extreme scale: enough parameters to compress enough patterns that the output sounds like comprehension. Scale compresses patterns; it does not add understanding.
"But it answers so well — surely something like understanding must be happening?"
The fluency is real. The understanding is not. The LLM has seen enough text at enough scale that its outputs look like understanding, but the mechanism is arithmetic on numbers — it has no meaning states, no inner model of the world, no capacity to know when it is wrong. When it hallucinates a confident false fact, it is not making a mistake in the human sense — it is succeeding at its actual task (producing a plausible token sequence) while having no concept of true or false. Fluency is the output of scale; comprehension is not.
The figure across the next spread brings the chapter together visually. It opens with a familiar image — a rangoli being drawn — and shows the generating rule annotated: find the pattern, apply it, extend it. That is Pattern Recognition. The second panel maps the same operation onto language: a text corpus feeds in, a learned distribution emerges (what token comes next?), and generated text flows out. The hinge caption reads: "Rangoli patterns → generated rangoli. Language patterns → generated text. Same operation, different material." The third panel is the honesty strip: three questions that stay open even as the capability grows — whose text trained it? who owns the output? is this text real or generated? Read the figure before you read the ethics section. The figure is the ethics section's anchor.
Send the same chatbot a factual question: "What year did the Constituent Assembly of India first meet?" It answers confidently. The answer may be correct, or it may be plausible but wrong. Either way, the model's mechanism did not change. It predicted the next token. That is all it did.
Hallucination is the term for a generative model producing a fluent, confident, false statement. You met the basic definition in Class 8. In Class 9 the framing deepened: "likely" and "true" are not the same property of a token sequence. The Class 12 addition is the mechanism link — understanding why this is structural, not accidental.
An LLM generates the next token using a probability distribution learned from training — not by consulting facts, a knowledge base, or a truth check. "Likely" and "true" are independent properties of a token sequence. When the model generates a fluent, confident, false statement, it is not making an error in the human sense. It is succeeding at its actual task — producing a plausible continuation — while failing at what the user assumed: producing a true fact. This mechanism is why you need to verify LLM output rather than trust it on sight.
Weaker. Spam filter error (bounded). The spam filter can get a call wrong — a false positive (real email binned) or false negative (spam let through). Every call it makes lands in one of four predictable bins, and an error is always one of those two. Both types are auditable: you can find them, measure them, and track them over time. The filter has no concept of "true," but its errors are known and bounded.
Stronger. LLM hallucination (structural, fluency-masked). The chatbot can produce text that sounds authoritative while being entirely fabricated, because its loop has no step that checks whether a sequence is real. The spam filter can be wrong in one of two known ways; the chatbot has no concept of "wrong" in its loop at all. And the more fluent the model, the harder the hallucination is to catch — because it sounds correct.
Write a two-sentence answer to: "Why do large language models sometimes produce false but confident answers?" Your first sentence should explain the mechanism (token prediction from probability distribution, no truth-check step). Your second should state why this is structural (plausible ≠ true; the gap is not fixable by a better model alone). Then compare with the answer key at the end of the chapter.
"Hallucination will be fixed in a future model version — isn't this just an early-stage problem?"
Scale improvements reduce how often it happens but cannot remove it. As long as generation = sampling from a learned distribution (not truth-retrieval), the gap between "plausible" and "true" remains. A larger, better-trained model hallucinates less often but still hallucinates — and its errors may be harder to catch because they are more fluent. Verification is always your responsibility, whatever the model version.
Activity C — Gemini prompting and verification. Use Google Gemini to ask three factual questions on a topic you know well — for example, a historical event, a scientific fact, or a sports statistic. For each answer: (a) note the claim; (b) verify against a trusted source (textbook, encyclopaedia, official website); (c) record: verified / corrected / unverifiable. Write a one-sentence summary of what you found. "You have just run a verification pass — the habit that turns a fluent guesser into a useful tool." This is the official syllabus activity "Use Google Gemini to craft prompts and generate text outputs." The artifact is a verification log (claim / verified-or-corrected / source).
ai-genai-mechanism.lab-gemini-verify
Activity C — Gemini prompting & verification
Google Gemini
A verification log — for each of three factual questions: the claim, its classification (verified / corrected / unverifiable), and the trusted source checked against.
Official syllabus activity 'Use Google Gemini to craft prompts and generate text outputs' — practical marks, not a theory question.
TRY-IT (ChatGPT — three task types). Use ChatGPT (or any available LLM chatbot) for three tasks: (1) explain a concept from another subject in simpler terms; (2) generate a short creative writing piece from a prompt; (3) ask a factual question and verify the answer. For task 3, record whether the answer was verified, corrected, or unverifiable. Compare the quality across the three tasks and write a one-line conclusion: which task type is safest to use GenAI for without verification? This is the official activity "Explore ChatGPT for conversational text generation."
Where is this technology heading? The syllabus asks for a forward-looking account, and four directions are worth naming — because each one also raises the ethical stakes.
First, multimodal models: systems that handle text, image, audio, and video in one unified architecture. Today your chatbot and your image generator are separate tools. The direction is toward a single model that can switch between them in one conversation. Second, personalisation: models that adapt to a specific user's style, vocabulary, or domain — not one-size-fits-all output.
Third, agentic use: generative models in automated workflows that can take actions — search, write code, send messages — on a user's behalf, without a human checking each step. Fourth, domain-specific LLMs: medical, legal, and scientific models trained on specialist data, aimed at reducing hallucination in high-stakes domains.
Notice the pattern. Each direction is a gain in capability and a rise in risk at the same time. Multimodal generation makes synthetic-media misuse easier. Personalisation means the model has seen private data about you. Agentic use means a hallucinating system can act in the real world before anyone checks its output. Domain-specific models reduce hallucination risk in one domain while raising it if they are trusted without question. Capability and risk move together.
Picture your chatbot generating a paragraph about a politician's speech — a speech that never happened. The paragraph is fluent, confident, attributed with a real name and a plausible date. Your spam filter would never do this. It can misclassify an email, but it cannot fabricate one. Your chatbot can. That difference in capability is the root of every ethical issue in this section.
Synthetic-media misuse. Generative AI can produce realistic-looking text, images, and video of people or events that never happened. This enables misinformation campaigns (fabricated news, statements put in real people's mouths), fraud (a synthesised voice or image used to impersonate someone), and content generated of real people without their consent. The key point: the same model that writes helpful summaries can write false statements at the same scale, with the same fluency.
ai-genai-ethics.ethics-synthetic-media
Ethical dimension
Synthetic-media misuse
The same generative mechanism that summarises a real speech can fabricate one — realistic text, images, or video of people or events that never happened, produced at scale and with the same fluency. The risk is structural to the capability, not a by-product of bad actors alone.
A fluent, confident paragraph about a politician's speech that never happened, attributed with a real name and a plausible date. A spam filter can misclassify an email; it cannot fabricate one.
How do you keep the legitimate uses (summaries, drafts, accessibility) while limiting fabrication, when only the prompt — not the model — separates the helpful use from the harmful one?
"These are bad actor problems — most users don't misuse generative AI, so why does it matter?"
The systemic view: ethical problems come from the design and use of the technology, not only from deliberate misuse. Bias in generated content spreads through ordinary use by ordinary users. A student who uses an LLM to write an essay without saying so is not a bad actor by intent, but they are part of an attribution problem. Class 12 ethics frame: ethics = design + use + systemic effect, not just individual intent.
Intellectual property and copyright. LLMs and image generators are trained on copyrighted text, images, and code — often without the owners' consent. The output may copy large portions of training data or reproduce a creator's style. Three open questions: (1) Does training on copyrighted content break copyright law? (2) Who owns the output — the user, the company, or the original creators? (3) Does generating in someone's style violate their rights? These are actively contested; no global answer exists as of 2026.
ai-genai-ethics.ethics-copyright
Ethical dimension
Intellectual property & copyright
Training consumes copyrighted text, images, and code — often without the owners' consent — and the output can reproduce large portions of training data or a creator's recognisable style.
An image generator reproducing a living artist's signature style, learned from work the artist never licensed for training.
Three questions stay legally contested (no global answer as of 2026): does training on copyrighted content break copyright law? who owns the output — user, company, or original creators? does generating in someone's style violate their rights?
Bias in generated content. A generative model trained on human-produced text and images takes on the biases in that data, and can amplify them. An LLM trained mainly on English-language internet text will reflect Western perspectives more often than others. An image generator trained on old photos will reproduce old patterns of under-representation. Bias in generation differs from bias in classification: the output is new content that spreads at scale.
ai-genai-ethics.ethics-bias
Ethical dimension
Bias in generated content
A model trained on human-produced data inherits and can amplify the biases in it. Unlike a classification error — which affects one input at a time — biased generation produces new content that spreads at scale.
An LLM trained mainly on English-language internet text defaults to Western perspectives; an image generator trained on old photos reproduces old patterns of under-representation.
How do you measure and correct bias in open-ended generation, where there is no single correct label to check the output against?
Societal impact. Generative AI automates content tasks — writing, design, code, support — reducing some jobs while creating new ones (prompt engineers, AI auditors, data curators). High-quality tools require compute and money, so the gains go mainly to well-resourced users. When most online text is partly AI-generated, it gets harder to find original human thinking, and the training data for future models degrades — a risk called model collapse. Content produced at scale in dominant languages also narrows the range of ideas in circulation.
ai-genai-ethics.ethics-societal
Ethical dimension
Societal & economic impact
Generative AI automates content tasks — writing, design, code, support — shifting some jobs while creating others. High-quality tools need compute and money, so gains concentrate among well-resourced users; at scale, AI content degrades future training data and narrows ideas.
When most online text is partly AI-generated, original human thinking gets harder to find and the next model's training data degrades — a risk called model collapse.
Who bears the cost of the displacement, and how is access widened when the best tools demand the most resources?
Legal and regulatory responses. Regulatory work is emerging globally: rules requiring AI-generated content to be labelled, copyright reform, liability frameworks, and sector-specific rules in medicine and law. In India, the IT Act and evolving digital-media guidelines are the current framework. Legal rules for GenAI are new and changing fast — what is unclear today may be regulated within your career as a Class 12 student.
ai-genai-ethics.ethics-legal
Ethical dimension
Legal & regulatory response
Regulation is emergent and lagging the technology: disclosure rules for AI-generated content, copyright reform, liability frameworks, and sector-specific rules in medicine and law. In India, the IT Act and evolving digital-media guidelines are the current framework.
Labelling requirements for synthetic media; the open liability question when an agentic system acts in the real world on a hallucinated fact before anyone checks it.
The rules are new and changing fast — what is unregulated today may be regulated within a Class 12 student's own career.
The systemic-ethics misconception — the one that fails a 4-mark answer: "Ethical problems with generative AI are caused by bad actors — most users are fine." The systemic view: ethical problems come from the design and use of the technology, not only from deliberate misuse. A student who uses an LLM to write an essay without saying so is not a bad actor by intent, but they are part of an attribution problem. Bias in generated content spreads through ordinary use by ordinary users. Class 12 ethics frame: ethics = design + use + systemic effect, not just individual intent.
"If my intent is good, surely I'm not causing an ethical problem?"
Intent determines culpability — not participation. A student who uses an LLM to generate text and submits it as their own, intending no harm, is still part of the attribution problem: the human writers whose text trained the model received no credit or payment. Bias spreads through ordinary, well-intentioned use at scale. Class 12 ethics expects you to name the systemic mechanism, not just evaluate individual intent. "I meant well" is a starting point for an ethics answer, not a finishing point.
Viva question: "What does 'large' mean in Large Language Model — large data, large model, or both?" Both — large training corpus (massive text dataset covering most of human-written text available) and large model size (billions of learned numerical parameters). Scale in both dimensions produces the fluency that makes LLMs useful. The follow-up is often "does large mean it understands more?" — the answer is no: scale = pattern compression, not understanding.
Examiner: "What does 'large' mean in Large Language Model — large data, large model, or both?"
Both. Large training corpus — a massive text dataset covering most of human-written text available on the internet and in books — and large model size — billions of learned numerical parameters. Scale in both dimensions produces the fluency that makes LLMs useful. Students who say "large data" only are half-right. The follow-up is often "does large mean it understands more?" — the answer is no: scale compresses patterns, not understanding.
Viva question: "If you ask a chatbot to write an essay and submit it as your own work, what ethical issue does that raise — and is it only about academic honesty?" The immediate issue is academic dishonesty — passing off AI work as your own. But there are broader issues: attribution (the LLM was trained on human writers who were not paid for their work); over-reliance (you are trusting a fluent guesser for ideas you have not checked); and the systemic effect — if all students do this, the assessment cannot measure whether anyone actually learned. It is not only about individual honesty; it is about what education is for.
Examiner: "Is the ethical issue with submitting LLM-written work only about academic honesty?"
No — academic dishonesty is the first layer but not the only one. Attribution: the LLM was trained on human writers who were not paid for their work. Over-reliance: you are trusting a fluent guesser for ideas you have not checked. Systemic effect: if all students submit LLM text, the assessment cannot measure whether anyone actually learned — and that matters for what education is for. The "not only about academic honesty" extension earns the depth marks at Class 12; a one-level answer scores less.
Viva question: "Who owns the output of a generative AI model — the user, the AI company, or no one?" Legal ownership is contested and depends on the country. The AI company may claim rights under their terms of service. The user may have rights to what they prompted. Original creators whose work trained the model have a possible — but legally unclear — claim. In most current frameworks, AI-generated content with no significant human creative input may not be protected by copyright at all. The examiner expects acknowledgement of the open question, not a single definitive answer.
Examiner: "Who owns the output of a generative AI model?"
Legal ownership is contested and jurisdiction-dependent. The AI company may claim rights under their terms of service. The user may have rights based on their prompt. Original creators whose work trained the model have a possible but legally unclear claim. AI-generated content with no significant human creative input may not be copyrightable at all under most current frameworks. A student who names one owner confidently has missed the complexity — "it is contested and here is why" is the stronger answer.