Feature suggestion: Context-Adaptive Fact Verification
I’d like to suggest a reasoning feature I informally call “semi-analysis.”
The idea is not to verify every statement or turn normal conversations into Deep Research. Instead, ChatGPT should monitor when it moves from general reasoning to a specific factual claim.
For example:
reasoning → specific factual claim → confidence check → targeted verification → continue reasoning
General knowledge would not require verification. But if the model is discussing a specific scene, quotation, episode, chronology, canon detail, technical specification, or author statement and is not sufficiently confident, it should verify that particular claim before using it as a premise for further reasoning.
If verification is unavailable, the model should explicitly mark the claim as uncertain or hypothetical rather than presenting it as fact.
When verification is performed, the relevant source should ideally be shown to the user so they can independently check the factual basis of the analysis.
The key idea is:
Don’t verify everything. Verify uncertain factual premises when they become necessary for the reasoning.
This could reduce a particularly difficult type of hallucination: an otherwise excellent and convincing analysis being built around one plausible but unverified detail.