Understanding how Large Language Models think — and why they sometimes go wrong

 

 

Hallucinations — confident but incorrect outputs produced by Large Language Models (LLMs) and related AI systems — are one of the most pressing challenges in the responsible use of AI.

 

Most research on hallucinations is machine-centred: it evaluates LLM outputs and labels errors as failures of reasoning or logic.

While this view may help technical specialists, it can mislead others by suggesting that LLMs are “dangerous” because they are expected to reason like humans — yet inevitably fail.

 

We take a different perspective. We explain hallucinations through the mechanics of tokenisation — how LLMs split input into subword units (tokens) and predict tokens rather than words.

 

This view makes errors intelligible: for example, the tokenisation of numbers clarifies why models often struggle with counting and calculations.

 

By showing how hallucinations emerge from token-based processing, we offer a transparent and practical account of their sources.

 

You can contribute to our research program on AI hallucinations by providing data -- cases of hallucinations you have experienced -- click the blue button.