AI Verification Theater
Status: ready
Symptom
The team asks AI to review a screenshot, code or a transcript and treats the answer as proof.
AI: the screen looks accessible
Decision: ship
The answer sounds confident. The app wasn't verified.
Where it shows up
PR review, QA triage, release summaries, design review, backlog grooming, scanner report cleanup, App Store label preparation.
Why it blocks the task
AI is useful for finding risks, grouping bugs and suggesting checks. It doesn't run VoiceOver on the device, doesn't feel where focus lands and can't tell whether the user finished the task. A confident "looks accessible" without a task, a profile and evidence is an opinion. When it's recorded as a pass, real blockers ship.
Root cause
AI output is confused with evidence.
- The prompt has no task and no user profile.
- A screenshot is reviewed without the VoiceOver transcript.
- AI findings have no status: hypothesis and verified look the same.
- The release summary drops the uncertainty.
- Production data is pasted into the prompt.
Before
Prompt: Is this screen accessible?
Answer: Yes, it appears accessible.
Decision: ship
No task, no profile, no evidence.
Fix
Verdict: AI produces hypotheses with a status. A hypothesis becomes a finding only after someone checks it on a device. Only checked findings count in the release decision.
Why: AI makes the search wider and faster. The device check makes the result true. Mixing the two gives a confident wrong answer.
A bounded prompt:
Task: recover from a declined card
Profile: VoiceOver
Inputs: screenshot, VoiceOver transcript, SwiftUI code for the error state
List likely blockers. For each: the check that confirms it.
Mark every item as a hypothesis.
Every AI item has a status:
AI finding: the payment error may not receive focus
Check: VoiceOver walkthrough of the declined card flow
Status: hypothesis
Owner: checkout QA
After the check:
Status: confirmed
Evidence: transcript attached
Decision: blocker
Input is synthetic. No real names, emails, addresses, card numbers, health data or tokens.
Check
Profile: release review
Task: review the AI findings in the release summary
Expected: every item has a status
confirmed items link to evidence
hypotheses aren't counted as pass or fail
Profile: privacy review
Task: review the prompts
Expected: no production personal data, no secrets
WCAG mapping
AI can suggest a WCAG mapping. A mapping isn't verification.
4.1.2 Name, Role, Value: AI can flag likely missing semantics in code.2.4.3 Focus Order: AI can compare a transcript with the expected order.3.3.1 Error Identification: AI can flag a missing error path.