Quick answer: Normalize the question, extract testable claims, verify each against primary sources, and keep the answer whose claims and limits are supported.
Two confident answers are not independent evidence and can repeat the same wrong assumption. Work through one test at a time so the result identifies the layer that is failing.
Why this happens
The best clue is the boundary of the problem: one file or every file, one device or several, one network or every network. Note when it began and any update, hardware change, permission prompt, or account alert that happened just before it.
- The prompts contain different dates, countries, constraints, or conversation history.
- One answer interprets an ambiguous term differently.
- Both answers rely on the same unverified secondary claim.
- A product or policy changed between the two responses.
Fix it in the safest order
- Write one precise question with date, location, audience, and constraints.
- Ask each answer to separate facts, assumptions, estimates, and unknowns.
- List the claims that would change the decision if wrong.
- Open primary documentation for those claims and record the supporting passage or data.
- Combine only verified points and state remaining uncertainty.
Repeat the same small test after each change. Keep a note of the original setting so you can restore it if the change does not help.
How to confirm the fix
A reviewer should be able to trace every consequential claim to evidence that actually supports it.
Common mistakes to avoid
- Do not pick the longer or more confident answer automatically.
- Do not count two AI outputs as two sources.
- Do not hide disagreement in high-stakes medical, legal, financial, or safety questions.
FAQ
Does asking the same question many times reveal the correct answer?
No. Repetition measures output variation, not truth. Verification against independent primary evidence is what matters.
Should I reset or reinstall everything?
No. A factory reset destroys useful evidence and is rarely the first step. Back up important data, test a second file, account, network, or device, and use the narrowest reset that matches the confirmed cause.
Related PHWhy help
Browse more AI & ChatGPT guides or search PHWhy for the exact error message and device model.
