AI changes the testing model
Traditional software often allows an exact expected result. AI systems may return different acceptable outputs for the same input.
Move from exact outputs to evaluation criteria
Instead of checking a single exact answer, AI testing often evaluates multiple quality dimensions:
- Correctness
- Relevance
- Completeness
- Groundedness
- Consistency
- Safety
- Instruction following
- Business usefulness
Build representative evaluation datasets
Create test prompts that represent normal use, edge cases, ambiguous requests, adversarial inputs, long context, missing context, and high-risk workflows.
Use thresholds and rubrics
Some AI qualities require a scoring rubric rather than pass/fail. For example, relevance may be scored 1–5 and require an agreed threshold for release.
Combine evaluation methods
| Method | Use |
|---|---|
| Exact / deterministic checks | Schema, tool calls, citations, structured outputs |
| Automated metrics | Similarity, retrieval metrics, latency, cost |
| Model-based graders | Scalable qualitative evaluation with governance |
| Human review | High-risk, nuanced, or business-specific judgment |
Test the whole AI application
An AI product includes more than the model. Validate orchestration, prompts, retrieval, tools, permissions, data, UI, safety controls, telemetry, and fallback behaviour.
Regression matters even when outputs vary
Keep a golden evaluation set and compare quality after prompt, model, retrieval, tool, or configuration changes.