Core reframe: fluency is not evidence of accuracy. A confident, well-formatted, grammatically clean paragraph can still be entirely wrong. The polish of the delivery says nothing about the correctness of the content — treating them as correlated is the single most common evaluation mistake.
Where hallucination risk concentrates:
- Specific figures, statistics, and dates
- Citations, quotes, and named sources
- Niche or obscure claims — precisely the details that feel most "factual" and get checked least, because they're too specific to be an obvious guess
Unsupported claims are statements presented as settled fact with no traceable source behind them — most often numbers, dates, names, and statistics dropped into an otherwise well-argued paragraph without attribution.
Bias is a distinct failure mode from factual error: a biased output can be built entirely from individually true statements, arranged in a skewed or cherry-picked way, or treating a genuinely contested topic as settled in one direction. Spotting bias means checking the arrangement and framing of true statements, not just their truth value.
Conflicting information shows up two ways: internal contradiction (the output disagrees with itself), or contradiction with the source documents the task was supposed to be grounded in.