Intermediate level · Module 5: Evaluating AI Outputs · Lesson 3 of 3
Human Evaluation and Error Analysis
Review outputs with people, categorise failures and decide what to fix.
In this lesson you will
- Run structured human evaluation
- Categorise and prioritise errors
Key concepts
human reviewerror categoriesseverityevaluation reporting
The lesson
Ask reviewers who know the subject — school staff, clinic staff — to grade a sample using the rubric. They notice errors that outsiders miss.
Categorise failures: wrong retrieval, ignored passage, hallucinated detail, unclear language, failed refusal. Rate severity by potential harm.
Report results with the numbers that matter: correctness rate, grounding rate, refusal behaviour and the most serious failures, with planned fixes.
Step by step
- Run a human review of 20 answers.
- Categorise all failures.
- Rate their severity.
- Write the evaluation summary.
Evaluating a school-registration FAQ assistant
Results of a 40-question evaluation by question type, with what to fix next.
Correct answers
85%
34 of 40 questions
Illustrative evaluation data
Correctly declined
5 of 6
Out-of-scope questions
Citations pointing to the right passage
90%
36 of 40 answers
Tricky / multi-part questions
2 of 4
Weakest area
Show the data behind the chart
| Question type | Questions | Correct |
|---|---|---|
| Registration | 12 | 11 |
| Fees | 10 | 9 |
| Documents | 8 | 7 |
| Out-of-scope | 6 | 5 |
| Tricky / multi-part | 4 | 2 |
Interpretation
Straightforward questions do well (27 of 30 correct across registration, fees and documents). Most failures are in multi-part questions, where the assistant answered only the first part.
One out-of-scope question — about a specific child's case — was answered instead of being referred, which is the most serious failure because it involves personal decisions.
Recommendations
- Add an instruction and examples for answering every part of multi-part questions, then re-run the full evaluation.
- Strengthen the rule that personal case questions are always referred to the school office, and add five such questions to the evaluation set.
Practice activity
Complete the FAQ assistant evaluation report from the practical activity.
Expected result: A report with metrics, categorised failures and prioritised fixes.
Check your understanding
-
Why involve subject experts in evaluation?
Show answer
They notice errors that non-experts miss.
-
How should failures be prioritised?
Show answer
By how often they happen and how much harm they could cause.
Taught live by an instructor — online or in person. Free course lessons are for self-study and do not lead to a certificate — see registration options.