Hallucinations and the Knowledge Cutoff

LLMs are powerful, but using them well means knowing exactly how they fail. The two failures that matter most for everyday use are hallucinations and the knowledge cutoff.

Hallucinations: fluent, confident, wrong

A hallucination is when a model produces something that sounds right but is false — an invented statistic, a fake citation, a library function that does not exist.

It happens because the model generates plausible text, not verified facts. OpenAI's research Why language models hallucinate (September 2025) adds a further reason: common training and evaluation methods reward a confident guess more than an honest "I don't know", so models learn to guess.

How to reduce hallucinations

• Give it the source. Paste the document and say "answer only from this text". (Doing this automatically from a database is called retrieval-augmented generation, RAG.)
• Allow uncertainty. Say "if you are not sure, say so".
• Ask for sources — then open them. A citation is only evidence once you have checked it exists and says what was claimed.
• Verify what matters. Facts, numbers, legal, medical and financial claims need a reliable second source.

The knowledge cutoff

A model learns only from data collected up to a certain date — its knowledge cutoff. Without tools, it does not know what happened after that, and it may not know how old its knowledge is.

• Recent news, prices, versions and laws may be out of date or simply missing.
• Assistants that can search the web or read files you give them can work around this — the fresh information comes from the tool, not from the model's training.

A good habit: for anything that changes over time, ask "as of when?" and check a current source.

Key takeaways

• Hallucinations are plausible but false outputs; confidence is not evidence.
• Ground answers in sources, allow "I don't know", and verify important claims.
• Without tools, a model knows nothing after its knowledge cutoff.