How to check ChatGPT answers and avoid AI errors
Introduction
ChatGPT and generative AI tools can help you write, summarise, explain, structure information and prepare drafts. But they can also make mistakes.
An answer can be fluent and convincing yet inaccurate. This is one of the major pitfalls of generative artificial intelligence: it produces language with considerable confidence, even when the information needs checking.
For personal use, an error may be inconvenient. In professional, educational or technical work, it can become a problem: incorrect information, imprecise instructions, an unchecked figure, an invented reference or an oversimplified explanation.
The right question is therefore not: “Can we trust AI?”
It is: “How can we use AI with a method for checking its output?”
Article outline
- Why AI can produce errors
- The most common types of error
- The DocPilot IA method for checking an answer
- Practical examples of verification
- Useful prompts for checking an answer
- What you should never delegate entirely to AI
- Key takeaways
- Conclusion
Why AI can produce errors
A generative AI tool does not work like a conventional database. It generates answers from statistical models of language.
It can therefore produce a plausible answer without automatically guaranteeing that every detail is accurate, current or appropriate to the context.
This does not mean AI is useless. It means it should be used as an assistant rather than a final authority.
AI is very useful for:
- rephrasing;
- structuring information;
- generating ideas;
- explaining concepts progressively;
- transforming notes;
- suggesting outlines;
- improving a text.
But its output needs checking when it covers:
- numerical data;
- recent facts;
- legal references;
- medical information;
- technical standards;
- proper names;
- quotations;
- dates;
- institutional information.
The most common errors
1. Factual errors
AI may provide an incomplete definition, confuse two concepts or present outdated information as current.
For example, it may correctly explain the general principle of a sensor but get the precise specifications of a particular model wrong.
2. Invented references
AI can sometimes produce article titles, authors or links that look credible but do not correspond to real sources.
A simple rule: verify every reference.
3. Unsourced figures
A percentage without a source should be treated as a hypothesis, not as reliable data.
4. Oversimplification
To make an explanation accessible, AI may remove important nuances. This can help with an initial understanding, but it is not sufficient for a course, report or technical procedure.
5. Overly generic answers
AI may provide a correct answer that is of little use because it does not take your actual context into account.
The DocPilot IA method for checking an answer
Step 1 — Identify the level of risk
Before checking, ask: “If this information is wrong, what are the consequences?”
Low risk: rephrasing, brainstorming, titles and ideas.
Medium risk: internal documents, learning materials, summaries and simple procedures.
High risk: legal, medical, financial, safety or regulatory matters, and critical technical data.
The higher the risk, the more rigorous the checking needs to be.
Step 2 — Separate style from substance
A text can be well written and wrong.
Check these separately:
- form: clarity, tone and readability;
- substance: facts, figures, logic and sources.
Do not be impressed by polished wording alone.
Step 3 — Identify checkable claims
When reading an AI answer, mentally highlight everything that can be checked:
- dates;
- figures;
- names;
- laws;
- standards;
- quotations;
- references;
- technical specifications;
- rankings;
- comparisons.
Every checkable detail should be verified if the document will be published or used officially.
Step 4 — Ask the AI to critique its own answer
A useful prompt:
“Critically analyse your previous answer. List the claims that should be verified, potentially uncertain passages, risks of oversimplification and points that require a reliable source.”
This prompt does not replace human checking, but it helps identify sensitive areas.
Step 5 — Compare with reliable sources
For important information, compare the answer with:
- official documentation;
- an institutional website;
- a scientific publication;
- a technical manual;
- a manufacturer’s page;
- a regulatory text;
- an academic source.
Do not check one AI answer against another unsourced AI answer.
Step 6 — Rewrite cautiously
When information is uncertain, avoid absolute statements.
Instead of saying:
“This method is the best.”
Prefer:
“This method is often appropriate when the aim is to structure a first draft quickly, provided that the result is checked.”
Caution strengthens credibility.
Practical example: checking a technical explanation
You ask ChatGPT: “Explain how an ultrasonic sensor works.”
The answer may be broadly correct. But if you are preparing teaching materials, you need to check:
- the transmission and reception principle;
- the distance calculation;
- measurement limitations;
- the influence of temperature;
- possible errors;
- the component examples.
You can then ask:
“Rephrase this explanation for beginner students, but retain the important technical limitations.”
This gives you an explanation that is both accessible and more rigorous.
Practical example: checking a professional document
You ask AI to turn notes into a meeting report.
Points to check:
- participants’ names;
- the decisions actually made;
- deadlines;
- responsibilities;
- sensitive wording;
- implicit commitments;
- confidential information.
AI can structure the report, but it must not invent a decision or assign an action to the wrong person.
Useful prompts for checking an answer
Prompt 1 — Identify risks
“Analyse this text and identify information that must be checked before publication. Classify it by risk level: low, medium or high.”
Prompt 2 — Spot unsourced claims
“List every factual claim in this text. Indicate which ones require a source or external verification.”
Prompt 3 — Improve caution
“Rewrite this text in a more cautious, rigorous and professional style, without exaggeration or unverified claims.”
Prompt 4 — Check internal consistency
“Analyse the consistency of the following document. Identify contradictions, repetition, imprecise wording and passages that could be misinterpreted.”
What you should never delegate entirely to AI
Never delegate fully:
- final approval of an official document;
- a regulatory decision;
- a legal analysis;
- a medical decision;
- a financial conclusion;
- a safety check;
- critical data verification;
- a sensitive communication.
AI can help with preparation. It should not give approval on your behalf.
Key takeaways
- A fluent answer is not necessarily a reliable one.
- AI should be used as an assistant, not as the final source.
- Facts, figures, dates, references and sensitive information must be checked.
- The safest method is to separate form from substance.
- Good use of AI always includes a checking stage.
To learn how to scope, check and improve your AI-assisted work, download the free DocPilot IA pack and explore the resources on prompts, guided methods and professional documents.
Receive 25 AI prompts to frame your requests and check your professional, educational and technical materials.
Conclusion
Checking ChatGPT is not a waste of time. It is what makes serious use of AI possible.
The aim is not to distrust everything, but to know where to focus your attention.
AI is highly effective for ideas, structures and rephrasing. For facts, figures and sensitive decisions, verification remains essential.
DocPilot IA stands for a simple idea: AI becomes powerful when used with method, high standards and critical thinking.
