How to verify AI responses and find hallucinations

Why AI can sound convincing and still be wrong

The language model selects a plausible continuation of the text. It does not consult a truth directory within itself before every sentence. Therefore, an answer can be smooth and logical, yet simultaneously contain an incorrect date, a non-existent quote, or a link that leads to the wrong document.

OpenAI explicitly warns that ChatGPT may hallucinate facts, dates, quotes, and sources. A separate study by the company explains: common evaluations often reward guessing instead of an honest "I don't know." NIST classifies false or inaccurate content from generative systems as information integrity risks and recommends documented verification.

This does not mean that AI is useless. It is excellent at accelerating structuring, finding directions, creating first drafts, and formulating questions. The model's role is simply to— prepare material for verification, rather than issue a certificate of truth.

What to check first

Not every sentence has the same cost of error. Start with statements that are easy to verify and that could change a decision.

Specifically flag words like "always," "never," "proven," "guaranteed," "all experts." They often mask a missing condition or an overly broad conclusion.

15-minute verification algorithm

1. Record the prompt and the answer. Save the exact text. Otherwise, after several refinements, it will be difficult to understand exactly what you were verifying.
2. Break the answer into atomic statements. One statement — one verifiable thought. "Temporary Chat is anonymous and deletes everything immediately" should be split into at least three parts.
3. Identify the risk. Highlight numbers, dates, quotes, relevant rules, and everything upon which an action depends.
4. Formulate a verification query. Don't look for the whole answer, but for a specific thesis: a function name, a number, a phrase from a document, or an exception condition.
5. Open the primary source. For a service function, this is the official documentation; for a law, the official text; for research, the article itself, not a summary in a blog.
6. Verify the context. Check the date, version, country, tariff, sample, definition, and the paragraphs before and after the quote.
7. Write down the verdict. Confirmed, partially, not confirmed, or outdated. If sources contradict each other, do not hide it.

If the answer is long, do not try to verify it "by eye." First, select 3–7 statements with the highest risk. This provides more value than a superficial scan of twenty links.

Statement log: a simple table that builds discipline

A log is not needed for every creative paragraph. However, for an article, a presentation to management, advice to a client, or automation, it shows where data ends and interpretation begins.

Hierarchy of sources: what is stronger than what

1. Original document or data: regulation, product documentation, repository, dataset, scientific article.
2. Official explanation from an author or institution: help center, release, technical report, regulator page.
3. Independent professional analysis: needed for critiquing methodology, comparison, and context that the manufacturer lacks.
4. Quality media or review: useful as a topic map, but it is better to trace key theses back to the primary source.
5. Social media, forum, other AI response: this is a hint of where to look, not proof.

Quantity of links does not replace quality. Ten sites may repeat the same mistake. Check if they are truly independent and if they lead to primary data.

Example: "Temporary Chat deletes everything immediately"

Imagine the AI replied: "Temporary Chat is completely anonymous, is not stored, and is immediately deleted from servers." The phrase sounds reassuring but contains several different promises.

• Does not appear in history: this can be confirmed in the official feature description.
• Not used to improve models: also confirmed for temporary chat.
• Deleted immediately: incorrect; official documentation allows a copy to be stored for up to 30 days for security reasons.
• Completely anonymous: the official source does not provide such a guarantee. This is an extra conclusion that should be removed.

Thus, the correct version is shorter and more precise: "Temporary Chat does not appear in history, does not create memory, and is not used to improve models, but a copy may be stored for up to 30 days for security." This is how fact-checking turns a beautiful promise into a limited, verifiable statement.

Three prompts to prepare for verification

These prompts do not prove truth. They help transform continuous text into a convenient list for independent work.

Why "checking yourself" is not enough

Self-correction research shows: without external feedback, a language model may fail to correct its reasoning or even worsen a correct answer. It sees its own text but does not automatically receive new evidence. Therefore, a second message like "are you sure?" is not equivalent to an independent source.

For medical, legal, financial, and security decisions, stop if there is no up-to-date official source or competent person. AI can help prepare a list of questions, but it does not replace responsible decision-making.

Checklist before using an answer

• The answer is broken down into separate verifiable claims.
• Numbers, dates, names, quotes, and URLs are given top priority.
• For key points, primary sources are opened, not just search snippets.
• Verified date, version, country, rate, and context.
• Calculations are repeated in a table, calculator, or code.
• Assumptions and estimates are not presented as proven facts.
• AI self-checking is not used as the sole evidence.
• For high risk, a competent specialist is consulted or the decision is halted.

Quick answers

Why does AI make up links?

It generates plausible text and may compose a name or address that looks real. Open the URL, check the author, date, and whether the page supports that specific claim.

Can you verify ChatGPT with another AI?

You can use a second model to look for contradictions, but this is not independent proof. Both models can repeat a common error. The final anchor is a source, a test, a calculation, or an expert.

How many sources are needed?

For a simple fact, one strong primary source is often enough. For a controversial conclusion, look for primary data and independent analysis. It is not the number of tabs that matters, but the relevance of the evidence to the claim.

What to do if verification is impossible?

Mark the claim as unverified, do not use it for important decisions, and rephrase the text without feigned certainty.

Safe use begins even before fact-checking: do not share unnecessary work data. There is a separate guide for this ChatGPT for work: how to use it safely. Read about editorial rules and corrections on the page "About the project" page.