Know when to slow down
A precise date, legal requirement, price, statistic, or medical claim deserves more scrutiny than a draft headline. New and changing topics also need a current source. Treat citations in the answer as leads to inspect, not proof by themselves.
- A deadline, law, grant, eligibility rule, or product requirement.
- A percentage, price, figure, or named study.
- Advice affecting health, money, employment, or a contract.
- A confident conclusion about a recently changed product or policy.
Separate facts, interpretations, and suggestions
A fact should be traceable to a source. An interpretation needs an argument. A suggestion must fit your situation and risk. Ask the model to label each kind, then verify the important facts independently.
Review your previous answer critically.
1. Which statements are verifiable facts?
2. Which are interpretations or recommendations?
3. What may have changed recently?
4. Which primary sources should I inspect?
5. What could you have got wrong?
If you cannot verify a statement, say so clearly.Trace claims to the source that can settle them
For law or public programmes, use the relevant authority or current legal text. For a product, check the maker’s documentation. For prices and conditions, inspect the current official page or signed offer. Several blogs repeating the same claim do not make it independent evidence.
Claim: [for example, an application closes September 30]
Primary source: [official URL or document]
Source publication/update date: [date]
Result: [confirmed / contradicted / unverified]
Exceptions or conditions: [region, eligibility, extension]
Next action: [what to do or whom to ask]Match the check to the stakes
- Low stakes
- Ideas and internal drafts: check the request was followed and the tone works.
- Medium stakes
- Purchases and external messages: check figures, terms, dates, and commitments.
- High stakes
- Health, law, tax, work, security, or investments: use current primary sources and a qualified professional.
Source and limits
OpenAI acknowledges that ChatGPT can produce inaccurate facts or references even when it sounds certain. This guide turns that warning into a record you can repeat.