Mashable reported that Google is building AI study features into Search. Whether or not you use them, AI-generated answers now sit above the results on a large share of searches, and the skill of checking one has become a basic literacy.
Why these systems are confidently wrong
A language model predicts plausible continuations of text. Fluency and accuracy are separate properties, and the model optimises for the first.
So a wrong answer arrives in exactly the same confident register as a right one. There is no hesitation, no hedging, and no internal signal that distinguishes the two. This is the crucial difference from a human expert, who usually sounds less certain when they are less certain.
Retrieval-based systems that cite sources are better, because they are summarising retrieved documents rather than generating from memory. But they can still summarise a source incorrectly, cite something that does not support the claim, or retrieve a poor source in the first place.
Where errors concentrate
- Numbers. Dates, prices, measurements, statistics and dosages. A plausible-looking figure is the easiest thing to generate and the hardest to spot.
- Anything recent. Models have a training cutoff, and without live retrieval they will answer about the past as though it were the present.
- Citations. References, case names, studies and page numbers are frequently fabricated in a well-formed shape.
- Niche topics. Thin training data produces confident nonsense more often than common topics do.
- Negatives. “There is no evidence that…” is a claim about the entire literature and is rarely verifiable by the model.
- Local specifics. Law, tax, opening hours and regulation vary by jurisdiction, and answers tend to default to the largest market.
A checking routine that takes thirty seconds
- Open the citation, do not just note that one exists. The single highest-value check. A surprising proportion of cited sources do not contain the claim attributed to them.
- Search the specific fact independently, in its own words, rather than searching the question again.
- Check the date on whatever you land on. Much of the internet is confidently out of date.
- Look for a second, unrelated source. Two sites repeating the same sentence are one source, not two.
- Go to the primary source for anything that matters: the statute, the paper, the manufacturer, the official guidance.
Scale the effort to the stakes. A restaurant recommendation does not need this. A medication interaction, a legal deadline or a tax figure does.
Ask in a way that fails loudly
Some phrasings make errors easier to catch.
Ask for sources explicitly, and ask which parts the model is least confident about. Ask it to argue the opposite case, which surfaces the counter-evidence it skipped. Give it your jurisdiction rather than assuming it inferred one.
And be careful with leading questions. Ask why something is true and you will often get reasons, whether or not it is true. Ask whether it is true and you get a more honest answer.
Where AI answers genuinely help
They are strong at things where you can verify the output yourself: explaining a concept you can then test against a worked example, summarising a document you have in front of you, drafting something you will edit, or giving you the vocabulary to search properly for a subject you know nothing about.
That last use is underrated. Not knowing the right term is often the real obstacle, and getting the term is genuinely valuable even if everything else in the answer needs checking.
The failure mode to avoid is treating a fluent paragraph as a finished answer. Treat it as a well-informed first draft from someone who never says they are unsure, and check it accordingly.
AI answers can also be manipulated deliberately. Our guide to prompt injection explains how.
Join the discussion