AI hallucination is when a model generates plausible-sounding but factually incorrect content — and does so without any signal that something is wrong. The model doesn't know it's wrong. It predicts text that fits the pattern of correct answers, even when it isn't.
Real examples of hallucinated AI output: - Fabricated legal case citations that look real (correct court name, plausible case name, believable date) but don't exist - Invented research paper abstracts with fake author names and DOIs - Incorrect statistics presented with false precision - Biographical details about real people that contain invented facts
The root cause is the training objective: predict the next most likely token. When an AI is asked about a specific case study, citation, or fact that: - Didn't appear often in training data - Falls after the training data cutoff - Requires precise detail the model doesn't have stored
...the model fills in with what sounds right, not what is right. It's the same mechanism that produces a fluent essay — but applied to fact-retrieval, it produces confident fabrications.
| High hallucination risk | Lower hallucination risk |
|---|---|
| Specific citations and case numbers | General explanations of concepts |
| Statistics and precise numbers | Brainstorming and ideation |
| Recent news (after training cutoff) | Writing assistance and editing |
| People's exact quotes | Code in common languages |
| Niche or specialist knowledge | Widely-documented historical facts |
Hallucination is especially dangerous because AI doesn't hedge. A human expert who doesn't know something says 'I'm not sure.' An AI model predicts the most plausible-sounding response — which is delivered in the same confident tone whether it's absolutely correct or completely invented.
This is called calibration failure: the model's expressed confidence doesn't match its actual accuracy.
AI hallucination is when a model generates plausible-sounding but factually incorrect content — and does so without any signal that something is wrong. The model doesn't know it's wrong. It predicts text that fits the pattern of correct answers, even when it isn't.
Real examples of hallucinated AI output: - Fabricated legal case citations that look real (correct court name, plausible case name, believable date) but don't exist - Invented research paper abstracts with fake author names and DOIs - Incorrect statistics presented with false precision - Biographical details about real people that contain invented facts
The root cause is the training objective: predict the next most likely token. When an AI is asked about a specific case study, citation, or fact that: - Didn't appear often in training data - Falls after the training data cutoff - Requires precise detail the model doesn't have stored
...the model fills in with what sounds right, not what is right. It's the same mechanism that produces a fluent essay — but applied to fact-retrieval, it produces confident fabrications.
| High hallucination risk | Lower hallucination risk |
|---|---|
| Specific citations and case numbers | General explanations of concepts |
| Statistics and precise numbers | Brainstorming and ideation |
| Recent news (after training cutoff) | Writing assistance and editing |
| People's exact quotes | Code in common languages |
| Niche or specialist knowledge | Widely-documented historical facts |
Hallucination is especially dangerous because AI doesn't hedge. A human expert who doesn't know something says 'I'm not sure.' An AI model predicts the most plausible-sounding response — which is delivered in the same confident tone whether it's absolutely correct or completely invented.
This is called calibration failure: the model's expressed confidence doesn't match its actual accuracy.