Imagine asking an AI assistant for a memorable line from a book. It returns an elegant quotation, names the author and supplies a page number.
The detail makes it feel credible. Yet when you open the book, the sentence may be nowhere to be found.
This is an illustrative scenario, but fabricated quotations and references are a documented problem in AI systems. Such errors are commonly called hallucinations, as OpenAI explains.
A convincing sentence is not proof
Many AI assistants rely on large language models. These models use patterns learned during training, together with the available context, to generate responses by predicting successive pieces of text.
That process can produce accurate explanations. It can also produce plausible inaccuracies. A fluent answer does not mean that every sentence was retrieved from a verified record. The NIST Generative AI Profile describes this risk.
Why doesn’t it simply say “I don’t know”?
It can. The difficulty is recognising and communicating uncertainty consistently.
Research from OpenAI identifies one contributing factor: evaluation systems can reward guessing. When a correct answer earns credit but admitting uncertainty earns none, taking a guess can improve a model’s score—even while producing more errors.
For readers, the practical implication is simple: an answer’s confident tone is not a reliable measure of its accuracy.
A citation is a place to start checking
A source link offers an opportunity to verify a claim. It does not complete the verification.
Suppose an assistant says a laptop has a spare SSD slot. The linked document might describe another model in the same product family. The page can be genuine while leaving your question unanswered.
NIST also warns that AI systems can generate misleading supporting logic or citations. Checking the explanation therefore includes checking its evidence.
Ask three questions before acting
1. What exact claim am I relying on?
“This is a good laptop” is a broad judgement. “It has an unused SSD slot” is a specific claim you can investigate.
2. What directly supports it?
In this example, look for the manufacturer’s specifications or service manual. If the document does not establish the claim, treat it as unconfirmed.
3. Does the evidence match my situation?
Check the full model number, region, date or version where relevant. Accurate information about a different configuration may not answer your question.
You can ask an assistant:
“Identify the important factual claims I should independently verify. Give an original source for each, and clearly state when you cannot find one.”
This can organise the checking process. You still need to inspect the sources.
Match the checking to the consequence
A suggested birthday message and a purchase decision do not need the same level of scrutiny. Spend more time checking when a mistake would have greater consequences.
AI can help you understand a subject and identify useful questions. Before acting on its answer, establish whether the crucial claim holds up.
Give verifiable evidence more weight than a convincing tone.
