General AI

Why AI fails

Reference only. Project progress and credentials come from assessed work.

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  • Stale context: the provided policy or record is no longer current.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.

Why AI fails | Mariete Academy