The core idea
Treat generated output as a candidate response that needs evaluation. Check it against the task, reliable evidence and independent reasoning. Useful prompting improves the request, but it does not turn confidence, citations or polished language into proof.
1. What a language model does
Artificial intelligence is a broad label for systems that perform tasks such as recognising patterns, making predictions or generating content. A large language model learns statistical relationships from training material and produces text using the current context. Text is represented through tokens, which may be words, parts of words or other units; a token is not always one word. Training and answering are different stages: training changes the model's learned parameters, while answering uses a trained model with the information available in that interaction. A system may also use search, calculators or other tools, but tool access is a separate capability that should not be assumed. The ability to produce a convincing paragraph does not establish that a fact was retrieved from a current source or that a calculation was independently checked.
Sources: Google: Introduction to large language models ↗ · Google: Large language models and their limitations ↗
2. Match the check to the kind of claim
First identify what the output asks you to believe. A numerical claim needs arithmetic, units and assumptions checked. A current deadline needs a current authoritative notice. A quotation needs the original wording and source. A summary needs comparison with the actual material being summarised. Language models can produce invented details or references, often called hallucinations, and may express them confidently. Asking “Are you sure?” can produce another confident response without adding evidence. Instead, ask for the underlying assumptions, then verify them independently. A link is useful only if it opens a real source that actually supports the claim. Several generated citations are not several pieces of evidence until checked. Where uncertainty remains, narrow the conclusion or leave the claim unresolved rather than making the wording more confident.
Sources: Google: Large language models and their limitations ↗ · Google Search: Refine searches ↗
3. Worked scenario: two percentages do not cancel
Imagine an AI says that increasing a price by 20% and then reducing it by 20% returns the original price. Test the statement with an invented starting price of ₹100. A 20% increase adds ₹20, giving ₹120. The later reduction is 20% of ₹120, which is ₹24, so the final price is ₹96. The percentages use different bases. More generally, 1.20 × 0.80 = 0.96, so the final value is 96% of the starting value. This independent calculation identifies the precise error instead of simply labelling the answer unreliable. Ask the tool to revise the explanation, but retain your own verified working. If the tool gives a correct final number with contradictory intermediate steps, those steps still need correction because a learner may reuse the faulty method elsewhere.
The percentage base changes
| Step | Calculation | Result |
|---|---|---|
| Start | Given amount | ₹100 |
| Increase 20% | 100 × 1.20 | ₹120 |
| Decrease 20% | 120 × 0.80 | ₹96 |
Sources: Google: Large language models and their limitations ↗ · Microsoft: Overview of formulas in Excel ↗
4. Worked scenario: a plausible but unverified deadline
A learner asks an AI for the application deadline of a local training programme. It produces a precise date and a reference that looks official. The learner opens the actual institute website, finds a notice for a different year and cannot locate the claimed current notice. The right conclusion is not that the programme is cancelled; it is that the date has not been established. The learner checks the relevant session, programme name and any amendment, then uses the institute's official contact route if needed. This example illustrates both source existence and source relevance. A real page about an old course does not support a new deadline. An AI-generated bibliography can be a set of search leads, but the final answer should cite sources the learner actually opened and understood.
Sources: Google: Large language models and their limitations ↗ · Google Search: Refine searches ↗
5. Give a task that can be checked
A useful prompt states the task, audience, supplied information, constraints and expected output. For example: “Using only this invented bus-timetable paragraph, write three comprehension questions for a beginner. Include answers and identify the sentence supporting each answer. If information is missing, say so.” This defines a small task with evidence you can inspect. It is stronger than “teach me everything about travel.” For learning, request one hint before a complete solution, try the problem yourself and then compare methods. Tell the tool which part you did not understand instead of repeatedly asking for a simpler answer without explanation. Providing a source can help keep the task grounded, but the output may still omit a condition or invent a detail. Check the finished answer against the supplied material and your original purpose.
Sources: UNESCO: Generative AI in education and research ↗ · Google: Introduction to large language models ↗
6. Keep responsibility, privacy and learning with you
Before uploading material, ask whether you have permission and whether personal details are necessary. Replace real names and identifiers with invented data for practice. Do not assume a service forgets an upload immediately; check its current settings and terms. Generated examples can reflect stereotypes or overlook local languages and circumstances. Test whether advice makes sense for the intended learner, available equipment and actual task. In assessed work, follow the teacher's rules and describe AI assistance where required. Passing off generated work as personal understanding can hide the very gap the assignment was meant to reveal. Use the tool to explain, compare, question and practise, while keeping a record of what you verified. High-consequence decisions need suitable evidence and qualified human judgement; a fluent response alone is not a sufficient basis.
Sources: UNESCO: Generative AI in education and research ↗ · W3C: Privacy Principles ↗
PUT IT INTO PRACTICE
Audit an answer before improving it
- Write the invented claim “a 20% increase followed by a 20% decrease cancels out.” Identify the starting value, each percentage base and the conclusion needing a check.
- Use ₹100, then ₹250, to calculate the results yourself. Predict whether the same proportional pattern should appear in both cases.
- Draft a prompt requesting a beginner explanation and one practice question. Ask for assumptions to be stated, but plan an independent check rather than relying on that request.
- Solution reasoning: ₹100 becomes ₹96; ₹250 becomes ₹300 and then ₹240. Both final values are 96% of their starts. Different percentage bases explain the pattern. An AI revision is acceptable only after its reasoning matches the calculation.
Check your understanding
Is a token always a complete word?
No. It can represent a word, part of a word or another unit. Token counts and word counts are therefore not interchangeable.
Why is confident language not evidence?
A model can generate persuasive wording for an incorrect claim. Evidence comes from a valid calculation, reliable source or other appropriate check.
Does asking for citations guarantee accuracy?
No. Check that the source exists and supports the particular claim. A real but irrelevant page is insufficient.
Why do the two 20% changes produce a loss?
The decrease applies to the enlarged value. Multiplying 1.20 by 0.80 gives 0.96, so the final value is below the original.
Why ask for a hint before the full solution?
It leaves space for your own reasoning and reveals which step you can do. Comparing later makes the tool a learning aid rather than a substitute for practice.
Why use invented personal details in practice prompts?
The learning task usually does not need real identities. Removing them reduces unnecessary disclosure without changing the reasoning exercise.