AI-Human Quality Assurance Protocol
A practical framework for trustworthy AI-assisted quality engineering
Key takeaways
- AI is already entering testing workflows — the real question is whether teams can trust its output enough to act on it.
- Fluent AI output is not verified truth; the paper argues trust must be engineered through controls, not assumed.
- Trust is built from four properties an organization can build and audit: evidence, controls, accountability and recoverability.
- The protocol is anchored by one principle: "AI may recommend. Evidence must prove. Humans authorize."
- The protocol defines an eight-step control loop — Govern, Ground, Generate, Verify, Approve, Execute, Observe, Learn — that is continuous, not linear.
In the paper's own terms
The problem
Most AI failures in testing are quiet: a quality signal can still look healthy after its meaning has changed — for example, a model silently rewriting an assertion so a failing test passes. The paper names the root cause as "blind trust": treating fluent output as correct output.
What the paper proposes
The AI-Human Quality Assurance Protocol (AHQAP) — an eight-step control loop (Govern, Ground, Generate, Verify, Approve, Execute, Observe, Learn) anchored by the principle "AI may recommend. Evidence must prove. Humans authorize."
Who should read it
The paper says it is "designed from the perspective of teams that must ship software, explain risk and stand behind release decisions" — QE and engineering leaders responsible for AI-assisted testing.
How this paper labels its claims
“AI may recommend. Evidence must prove. Humans authorize.”
