An AI result can be fluent and still fail the work. The most common causes are not mysterious.
Common failure modes
- Missing context: the system never received an important fact.
- Stale context: the provided policy or record is no longer current.
- Ambiguous task: the system had to guess the goal or output format.
- Unsupported generation: a likely statement was presented as a fact.
- Weak evaluation: nobody defined what a correct answer looks like.
- Excess authority: the system could act beyond the evidence or permission available.
Hallucination is one failure mode, not the only one. A perfectly factual answer can still be unsuitable if it uses private data, applies the wrong policy or makes a decision outside its authority.
Reduce uncertainty before adding automation
When a task is unstable, first improve the source set, instruction and review rubric. Automation scales the design you give it, including its unresolved weaknesses.