An AI once gave OpenAI researcher Adam Tauman Kalai three different titles for his own doctoral dissertation. None was correct. Asked for his birthday, it supplied three different dates, also wrong.

The interesting part wasn’t the error. Search engines return bad results every day, and colleagues misremember things all the time. What stood out was the presentation. The answers arrived cleanly, without hesitation, written in the same composed tone the model might use for something it actually knew.

That calm delivery is why AI hallucinations remain a serious problem. A clumsy mistake invites checking. A polished one gets copied into a document.

Fluency has been mistaken for knowledge

Language models are unusually good at producing sentences that belong together. They know which words commonly follow other words, which explanations sound convincing and how a professional response is normally structured.

None of this automatically means the facts underneath are correct.

That distinction gets lost because people are used to treating fluent language as evidence of competence. When someone explains a topic clearly, cites a paper and uses the right terminology, we tend to assume there is knowledge behind the words.

With an AI model, there may only be a strong prediction about what a knowledgeable answer should sound like.

OpenAI’s researchers argue that hallucinations begin partly because language-model training resembles a statistical prediction task. Rare or uncertain facts are difficult to learn reliably, so the model may produce a plausible answer instead. Later evaluations often make the behaviour worse by rewarding a correct guess and penalising an admission of uncertainty. When guessing offers a chance of success and saying “I don’t know” guarantees failure, guessing becomes the attractive option.

That explains quite a lot.

The citation looked completely normal

A familiar failure appears during research. Ask for academic papers on a narrow subject and the model may return a tidy list with authors, titles, journals and publication years.

One reference refuses to turn up anywhere.

The title sounds exactly like something researchers in that field would publish. The journal exists. Sometimes the named authors even work on related topics. The model has assembled several believable details into a paper that was never written.

This type of error can survive a quick review because nothing looks obviously ridiculous. Someone pastes the reference into an article, another person finds the article through search, and the fictional citation begins travelling.

The model did not secretly locate a bad database. It generated the shape of a citation.

That is a fairly useful way to think about hallucinations. AI can reproduce the shape of an answer without possessing the evidence that should support it.

Asking better questions helps, but only so much

Vague prompts create extra room for invention. “Tell me about this company” leaves open which company, which time period and what kind of information matters. A precise request narrows the possibilities.

Still, prompt quality is sometimes treated as a complete solution, which feels unfair to users. Someone should not need a minor qualification in prompt engineering to ask whether a product supports a particular feature.

Clearer instructions reduce ambiguity. They do not turn a predictive model into a verified database.

I’ve also noticed that adding “Do not hallucinate” has become the AI equivalent of putting a stern note on the office printer. It expresses a preference. It does not create a technical safeguard.

Fresh information creates another weak spot

A language model’s internal knowledge is limited by its training and whatever information it can retrieve during the conversation. Ask about a product update released yesterday, a revised company policy or a recently appointed executive, and the model may fill the missing space with older patterns.

Sometimes it gives an outdated answer. Occasionally it blends the old and new versions into something that never existed.

Retrieval can improve this by giving the model access to relevant documents before it writes. Anthropic recommends grounding answers in supplied sources, allowing the model to acknowledge uncertainty and requiring direct quotations or citations for claims that need verification. Even then, retrieval does not guarantee correctness. The wrong document may be selected, the source itself may be outdated, or the model may misread a relevant passage.

RAG is useful. It is not holy water.

Confidence is mostly a writing style

Users often ask an AI how confident it is, then treat the numerical answer as meaningful.

“I am 95% confident” looks scientific. Unless the system has been specifically designed and tested to estimate uncertainty, that number may be another generated sentence rather than a calibrated measurement.

The same problem appears in softer language. “Certainly,” “the evidence shows” and “research has established” can all be produced because those phrases fit the response, not because the model has checked the evidence.

This is where interface design matters. AI products should make sources visible, separate retrieved facts from generated interpretation and make uncertainty easy to express. A system that refuses occasionally may feel less impressive during a demo, but I would trust it more in actual work.

Some tasks can tolerate invention

Hallucination is not equally harmful everywhere.

During brainstorming, a slightly odd suggestion may lead somewhere useful. Fiction writing does not require every invented company or character to exist. Drafting ten possible names for a podcast is low risk.

The situation changes when the output contains legal clauses, medical guidance, financial numbers, software dependencies or research citations. A wrong answer can still look perfectly ordinary, which makes review necessary.

The sensible response is not to avoid AI. It is to match the checking process to the cost of being wrong.

I will happily use AI to reorganise notes or suggest alternative headlines. I would not let an uncited AI summary settle what a contract requires. Those tasks do not deserve the same level of trust.

“Smarter” does not mean finished

Newer models have become more capable, and developers have found better ways to ground responses, test factuality and encourage honest uncertainty. Hallucinations have not vanished. Research surveys still describe them as an open problem involving training data, model behaviour, retrieval, evaluation and the difficulty of identifying the boundaries of a model’s knowledge.

I don’t think waiting for a perfectly factual model is a useful plan. The more practical approach is to design AI use around the possibility of error.

Ask for sources. Open them. Check names, figures and quotations. Give the model current documents when current information matters. Let it say that the evidence is missing. Keep humans involved where a polished mistake would cause real damage.

The awkward truth is that AI does not need to be consistently wrong to create problems. It only needs to be wrong occasionally, in ways that look finished.