Let AI Generate, but Keep Verification Separate

After AI produces code, a report or a research summary, the risky move is not using AI. It is allowing the same generation process to casually certify that its own result is correct.

A more stable workflow separates production from acceptance. The first stage creates a candidate result. The second returns to the requirements and checks it using tests, factual verification, file diffs or an independent evaluation. A different model or a fresh context can sometimes help with that second pass, but the human acceptance criteria still need to be explicit.

NIST’s work on AI testing, evaluation, verification and validation reflects the same structural point. As AI becomes more capable, checking should not collapse into “it looks fine”. Machines can reduce the cost of generation. What enters a real system should still depend on evidence that is meaningfully independent of the act of producing it.

Sources:
https://www.nist.gov/ai-test-evaluation-validation-and-verification-tevv
https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence