How to Check if AI Answer Is Correct: Hallucinations Guide

How to check if AI answer is correct: treat every number, date, name, rule, section, quotation and reference in the answer as unverified, and confirm each one against an official or original source such as India Code, the Gazette of India, the regulator’s website or the document itself. AI chat tools generate likely-sounding text rather than looking facts up, so they sometimes “hallucinate”: they state false information confidently, with fake references. Judge the wording yourself, but never publish a fact you could not trace.

AI hallucination examples you will actually meet

• Invented references: a “Ministry of Labour report, 2024” that does not exist, or a court case with a real-sounding name and a fake citation.
• Wrong numbers: an old GST rate, a penalty amount from a repealed law, a made-up market size.
• Outdated law: an IPC section for an offence after 1 July 2024, or pre-Labour Code rules after 21 November 2025, when the four Labour Codes came into force.
• Summaries of unseen documents: ask for a summary of a report by name, without the text, and you may get guesswork.
• Fake quotations from famous people, and schemes or experts that do not exist.
• Stale instructions for a government portal or software menu that has changed.

The best-known courtroom example is from the United States: in Mata v. Avianca (2023), a federal court in New York sanctioned lawyers who filed a brief citing cases that ChatGPT had invented. In India, the Bombay High Court has imposed costs of ₹50,000 on a litigant whose written submissions relied on a non-existent, AI-generated judgment, which is why lawyers and students must open every authority they cite.

Why AI tools get facts wrong

A large language model learns patterns from huge amounts of text and produces a reply by predicting what a good answer looks like. Most of the time that is also a true answer. When its training is thin or out of date, it can blend similar patterns into something that sounds right and is not, without warning you. Tools are built to be helpful, so they tend to answer rather than say “I don’t know”. A leading question (“Give me the Supreme Court judgment that says X”) pushes them to invent something that fits. And some tools simply agree when you push back, even if you are wrong.

Tools with web search help, but the written answer is still generated. A link beside a sentence does not prove the page supports that sentence. Click it.

How to check if AI answer is correct: a six-step routine

1. Rate the risk. Low: rewording your own email. Medium: summaries, reports, lesson plans. High: anything on law, tax, money, health, safety, official figures, client deliverables or statements about real people.
2. Mark every factual claim. Numbers, dates, names, laws and sections, cases, quotations, “you must” rules and procedural steps.
3. Check each against the right source:
   • Central Acts and sections: India Code.
   • Notifications and rules: the Gazette, PIB, the ministry’s site.
   • Banking rules: rbi.org.in. Tax: the Income Tax and GST portals, or your chartered accountant.
   • Judgments: the court’s website, or reputable legal news such as LiveLaw and Bar and Bench.
   • Statistics: the original report. No original, no number.
4. Use the tool against itself, carefully. Ask: “List every factual claim in your answer and the official source I should check” and “Which parts might be outdated?” Then open those sources yourself.
5. Get human review for high-risk material: a manager, compliance, your CA or lawyer, or your teacher.
6. Keep a one-line note: “Drafted with AI; figures checked against Q2 sheet; sections checked on India Code.”

Example: Anjali, a compliance executive in Mumbai, got an AI draft saying new data protection duties applied “from 1 January 2026”. Checking the DPDP Rules, 2025, she found most duties apply from 13 May 2027. Ten minutes of checking saved her team from the wrong deadline.

Prompts that reduce hallucinations

• “Do not invent facts, figures, names, laws or references. If you need information I have not given, leave a placeholder in square brackets.”
• “Use only the text I have pasted below.”
• “List at the end every statement I should verify.”
• Ask neutral questions (“Is there a rule on X?”) rather than leading ones.

These lower the risk; they do not remove it. Checking remains your job, and you are responsible for anything you send, submit or publish under your name.

Frequently asked questions

How to check if AI answer is correct?

Mark every factual claim and verify each one against an official or original source: India Code for Acts, the Gazette or ministry site for notifications, the court’s website for judgments, and the original document for summaries. If you cannot trace a fact within a few minutes, leave it out.

What is an AI hallucination?

It is false or invented information that an AI tool presents as fact, often with confident wording and fake references. It happens because the tool predicts likely text rather than retrieving verified facts.

Can ChatGPT give wrong legal section numbers?

Yes, often. It may quote repealed IPC or CrPC sections, another country’s law, or a rule that was only proposed. Always check sections on India Code and remember that the BNS, BNSS and BSA apply from 1 July 2024.

Are AI answers with sources always correct?

No. A tool with web search can still misstate what a source says or attach a link that does not support the sentence. Open each link and confirm the exact wording.

Quick quiz

Sources

• Bharatiya Nyaya Sanhita, 2023 (s.318 cheating), in force from 1 July 2024
• Digital Personal Data Protection Rules, 2025 (phased commencement; main duties from 13 May 2027)
• Mata v. Avianca, Inc., US District Court for the Southern District of New York (2023)
• India Code (indiacode.nic.in); Gazette of India (egazette.gov.in)

Make verification a habit with the AI at Work certificate course.

This page is for education and general information, not legal advice for a specific case. Law as on 29 October 2026.