BlueAIBack to chat

BlueAI hub

How to check an AI answer

A fluent answer is not the same as a true one. This is the short version of how to tell them apart, using the sources sitting next to a BlueAI reply.

· BlueAI team · about 5 minutes

Do not trust it. Check the parts that would cost you.

An assistant is built to produce plausible sentences. That is a different job from producing true ones. Most of the time the two coincide. When they do not, the text still sounds like a person who has done the reading.

You do not need to verify every clause. You need to verify the two or three claims that would actually matter if they were wrong: a figure, a legal rule, a medical dose, a price, a date, a quotation, a citation. Everything else can be read the way you would read a well-informed colleague talking quickly.

If it looked something up, open the pages

On BlueAI a factual question can go to the web mid-answer. When that happens the sources sit next to the reply, not in a footnote you have to hunt for. That is the whole point of them: thirty seconds of clicking beats trusting a confident paragraph.

Open the official page
A government, standards body or primary document beats a blog that restates it. If the answer cites the official page, start there.
Check the date
A rule that changed last year is the classic failure. If the source is old, ask again and tell it to flag anything recent.
See whether the page actually says that
Models sometimes attach a real URL to a claim the page does not make. The link is not the proof. The sentence on the page is.

If there are no sources, ask for them

An answer without sources is a first draft. For anything factual, write back: search the web, cite the official pages, and flag anything that changed this year. That one follow-up is worth more than regenerating the whole reply.

A number you would act on

Weak

Is this figure right?

Better

Search the web for the current official figure. Cite the page it comes from, quote the sentence, and say if anything changed in the last twelve months.

It forces a live lookup, makes the claim checkable, and catches the case where the training data is simply out of date.

The safest questions are about something you supplied

A PDF, a spreadsheet, a screenshot, the text of an email. Questions about a document you attached are grounded in something real, which is why they fail less often than questions asked in the abstract.

The same is true of drafting and rewriting. Asking it to turn your notes into an outline is safer than asking it to invent the facts those notes should have contained.

Frequently asked questions

Keep reading: how to write prompts, what an AI chatbot is, coming from ChatGPT, the learning hub.