“Hallucination,” in the context of AI tools, refers to a model generating a confident, fluent, plausible-sounding statement that's factually wrong or entirely fabricated — a citation to a paper that doesn't exist, a specific statistic invented rather than recalled, a biographical detail attributed to the wrong person. The term is evocative but slightly misleading, since it suggests something closer to a malfunction or an unusual error state, when the actual mechanism is a direct, expected consequence of how these models fundamentally work, discussed in the what-llms-actually-do guide elsewhere on this site.

For another practical perspective on prompt structure and iteration, Google AI prompting strategies is useful further reading.

Why this isn't simply a bug that gets patched out over time

A model generates text by predicting statistically plausible continuations, and a fabricated fact can be just as statistically plausible, in the specific sense of matching the patterns of genuine facts, as a real one — particularly for a narrow, specific, or obscure claim that's underrepresented in training data. Newer models have genuinely reduced hallucination rates through various training improvements, and the underlying mechanism that makes hallucination possible in the first place hasn't fundamentally changed, which is why it's more accurate to expect a continuing, gradually improving risk than a fully solved problem with a specific model release.

The specific conditions that tend to raise hallucination risk

Obscure or narrow topics, underrepresented in training data, carry higher risk than well-known, extensively documented topics, since the model has less genuine signal to draw on and more room for statistically plausible invention to fill the gap. Very specific, precise claims — an exact statistic, a precise date, a specific citation — carry higher risk than general, qualitative claims, since precision leaves less room for a vague, defensibly-approximate answer and more temptation for the model to generate a specific-sounding but unverified number. And requests pushing toward a confident, definitive answer on a genuinely uncertain or unsettled topic tend to produce a falsely confident response more often than a request that explicitly allows for uncertainty or an “I'm not sure” answer.

This topic also has a human attention and collaboration dimension; https://www.monitask.com/business-glossary/interpersonal-synchrony/ provides a useful related explanation.

Why understanding the mechanism is more useful than just knowing the word

Knowing the term “hallucination” without understanding why it happens tends to produce either excessive, blanket distrust of every AI output, or an underestimate of the risk once a specific model seems to “handle it well” in casual testing. Understanding the actual mechanism — statistically plausible generation without a built-in fact-checking process — supports a more calibrated response: trust scales with how well-documented a topic is and how precise a claim is, rather than treating hallucination risk as either uniformly high or a solved, no-longer-relevant problem.

Hallucination isn't a bug waiting to be patched — it's a direct, structural consequence of how these models generate text, which is why the practical response isn't waiting for it to be fixed but calibrating trust to the specific conditions (topic obscurity, claim precision, available retrieval) that predict when it's more or less likely.

This closes the loop on the mechanism-first understanding introduced in the what-llms-actually-do guide on this site's Writing section — hallucination is that guide's central caution, named specifically and explained in enough mechanical detail to actually predict when it's most likely to matter for a given task.