If a chatbot gives wrong answers, first check whether the correct information exists in its sources. Then inspect the passage it used and how it interpreted that passage. Outdated sources, poor information selection and incorrect interpretation need different fixes. Locate the failure before changing the model.
Record the incorrect answer and its context
Record the question, relevant conversation context, expected answer and approved source supporting it. “The bot is unreliable” is less useful than “It explains this service using an outdated support policy.” A precise case lets technical and content owners investigate the same problem. Avoid introducing unnecessary personal details into the evaluation set.
NIST's generative AI profile addresses the risk of systems presenting inaccurate information confidently. Fluent, persuasive wording therefore cannot replace a correctness check. Assess tone and factual accuracy separately: a polite incorrect answer is still a failure to resolve.
Trace source → selected information → answer
Illustrative failure: a bot describes standard support response time as the delivery time for a custom project. Find the approved condition first. If the selected passage concerns standard support, investigate source selection or document separation. If it used the correct passage but omitted a condition, investigate answer generation.
Record the question, expected answer, selected source passage and actual reply side by side. Keeping only the final answer makes diagnosis harder. If documents conflict, have the content owner identify the valid version rather than expecting the bot to resolve the contradiction itself.
- No source: add approved information or route the topic to a person.
- Old source: update it and check whether obsolete versions remain in use.
- Wrong passage: inspect document titles and the retrieval process.
- Wrong summary of the right passage: add the expected answer to the test set and correct the relevant response rule.
Choose a fix for the error type
Populate the table with examples from your conversations. Give each correction an owner and a verification case. Avoid accumulating every issue in one general instruction; simplify the information and workflow boundaries first.
| Failure type | Question | Correction focus |
|---|---|---|
| Outdated or conflicting facts | Is the current source established? | Content ownership and version cleanup. |
| Wrong information retrieved | Is the relevant passage found? | Information structure and retrieval. |
| Correct facts misrepresented | Was the answer checked against its source? | Response rules and evaluation examples. |
| Confident out-of-scope answer | Is a handover condition defined? | Scope and human escalation. |
| An uncompleted action reported as done | Does a real system result exist? | Action-result and confirmation alignment. |
Retest the question and related cases
Repeat the problematic question, but also test another wording, a request missing essential information and a similar request outside the scope. A correction may improve one answer while creating unnecessary handovers or false certainty elsewhere. Record results against the same criteria.
For consequential actions such as pricing, appointments or order changes, use the actual system record to establish what happened. Answer-quality review and action verification are different jobs. Until a repeated failure is corrected, narrowing the relevant authority or routing that topic to a person can be practical. This manages the affected workflow without assuming the whole assistant must stop.