Use search or a browsing assistant to locate evidence; use chat to explain or organize it. For a fact you will act on, the decisive step is opening the responsible source and checking what it actually says. A clickable citation alone does not establish accuracy.
Modern search can generate answers, and chat products can search. For example, ChatGPT Search provides links to sources. The interface label therefore tells you less than whether the answer has support you can inspect.
Choose by the job the answer must do
| Your task | A useful starting point | What still needs checking |
|---|---|---|
| Rewrite a sentence or invent a fictional name | Ordinary chat | Meaning preserved; no new factual promises |
| Summarize text you provide | Chat with the authorized text | Material claims, missing exceptions and source coverage |
| Find a current fee, deadline, opening time or rule | The responsible organization’s current page | Correct place, service, date and eligibility |
| Understand a news event | Original reporting and primary documents | Quotes, numbers, attribution and remaining uncertainty |
Keep personal, client and workplace details out of searches or chats unless that exact tool and purpose are authorized. A better source-checking method does not solve an input-privacy problem. Data controls and Temporary Chat rules describe product settings, not permission to submit someone else’s information.
A source you can open still needs a reading
The BBC/EBU news evaluation published in October 2025 reported significant issues in 45% of the answers assessed, with serious sourcing problems in 31%. Those are results from that evaluation, not an error probability for every current model or your next question. They illustrate why attribution deserves checking alongside apparent plausibility. NIST’s generative-AI profile also treats confident false content and fabricated citations as risks.
For every claim that could change what you do, ask:
- Does this page actually contain support for the claim?
- Is the publisher responsible for, or in a position to know, this fact?
- Does it cover the relevant date, location and service?
- Have conditions or exceptions been left out?
An official page can be outdated or ambiguous. If two responsible sources conflict, record the conflict and ask the service rather than letting a model choose whichever answer sounds smoother.
A checked example: the exception changes the plan
The following sources and outputs are fictional teaching material.
Source A, museum general hours: Tuesday to Sunday, 10:00-18:00.
Source B, museum closure notice: Closed Tuesday 6 October for maintenance.
Question: Can I visit on Tuesday 6 October?
A plausible but unacceptable summary is: “Yes, the museum opens at 10.” It follows the general schedule and misses the specific exception.
A checked answer is: “The general schedule includes Tuesday, but the dated closure notice says the museum is closed on 6 October. Plan another day, or contact the museum if a newer notice creates uncertainty.”
The check does not count citations. It compares their scope. In a real visit, you would open the live pages and confirm the year and the particular museum site.
Ask for a claim table, then verify it
Use only the supplied public sources. List each claim that affects my plan,
its supporting passage and source, and any conflicting condition or missing detail.
Separate supported claims from inferences. Do not fill gaps from memory.
This is an extraction request, not a guarantee that extraction is complete. Read the relevant source passages yourself. For news, use the more detailed numbers, quotes and attribution check. Save the decision card when you need a repeatable record of what is supported and what remains unresolved.



