Why AI chatbots 'hallucinate', and how to work around it
A language model predicts plausible text, not verified fact. That single design choice explains most confident-sounding errors.
Meridians Tech Desk
Published 2 July 2026 · Updated 2 July 2026 · 6 min read
AIThe short answer
- Large language models generate the most statistically likely next words, which is not the same as retrieving true statements.
- Fabricated citations, invented figures and confidently wrong answers are the predictable result, not rare glitches.
- Grounding a model in a specific source and asking it to quote from that source reduces, but does not remove, the problem.
- Treat any factual claim, name, number or citation from a chatbot as unverified until you check a primary source.
The word hallucination makes the behaviour sound like a malfunction. It is closer to the opposite: a system doing exactly what it was built to do, in a situation where that behaviour is unhelpful.
Prediction, not recall
A language model is trained to continue text with the most plausible next words given everything before them. It has no separate store of verified facts it consults, and no built-in sense of whether a fluent sentence is true. When it does not have the information, it does not stop — it produces the most likely-sounding continuation, which can be entirely invented.
Why it sounds so convincing
The same machinery that makes the writing fluent makes the errors fluent. A fabricated statistic or a non-existent citation arrives in the same confident register as a correct one, because the model is optimising for plausibility, not accuracy.
Practical ways to reduce it
- Give the model the source material and ask it to answer only from that text, quoting the relevant passage.
- Ask it to say explicitly when it is unsure rather than guessing.
- Cross-check every name, number, date and citation against a primary source before relying on it.
- Prefer it for drafting, structuring and explaining over retrieving specific facts.
Sources
Every factual claim above is traceable to these documents. Check them — that is why they are here.
- Artificial Intelligence at NISTUS National Institute of Standards and Technology
- AI Risk Management FrameworkUS National Institute of Standards and Technology
About this byline
Meridians Tech Desk is an editorial desk at Meridians, not an individual. A desk byline means the article was produced and fact-checked to that desk's published standards. Read our editorial standards and corrections policy.
Sponsored
Paid placement · not editorial
Verdant Home · Furniture & interiors
Made to be repaired.
Solid-wood furniture with replaceable parts, delivered flat and assembled by hand. Sponsored modules are produced outside our editorial desks and are always labelled.
Browse Verdant →
Verdant HomeWhy am I seeing this? · Demonstration creative from a fictional advertiser. Advertising never influences our journalism.
Passkeys, explained: what replaces the password, and what it doesn't
8 min read · 12 July 2026
EV charging speeds: why '150 kW' rarely means 150 kW at your car
6 min read · 11 July 2026
The Meridians Brief
One considered email a week
What changed, what it costs you, and what to do about it — from the Meridians desks. No sponsored picks disguised as recommendations.
Sign-up opens with our launch issue. Nothing is sent or stored yet.