AI-foutmodi

Examples include unsupported claims, missed cases, data leakage, unsafe tool actions, and failures under changed inputs. Classifying failure modes helps teams test and address causes rather than treating every mistake as the same problem.

Key takeaways

• Describe triggers and consequences precisely.
• Separate model, data, and workflow failures.
• Match mitigations to the observed cause.

Begin with the intended behavior and its boundaries. A wrong category, an invented citation, and a duplicated payment request require different responses. Record the trigger, observed result, affected component, and practical consequence for each failure. Distinguish model errors from system errors. A model may correctly interpret a request while a tool executes with the wrong account, a stale document supplies outdated policy, or a retry repeats a completed operation. End-to-end verification is essential when an output can change external state. Test ordinary variability and deliberate misuse separately. Formatting changes, dialects, missing data, long documents, and conflicting instructions can expose weaknesses without an adversary. Security tests add cases where an attacker tries to redirect behavior or access information. Choose controls matched to the cause: input contracts, evidence checks, permission limits, transaction identifiers, abstention, or human review. Keep failed cases for regression testing and record residual uncertainty. A mitigation that catches one example should not be described as eliminating an entire class of failures.

1. Imagine an assistant saying that a file was saved after its storage tool timed out.
2. Inspect the destination to determine whether a file exists and whether its contents match the request.
3. If the outcome is unknown, report that state and use a safe reconciliation step before retrying. Add the timeout case to the regression suite.

This hypothetical example tests observable completion rather than the assistant’s description of it.