Bug Report: Incorrect Capability Denial Triggered by Safety Layer
`Steps to Reproduce`
Ensure Web Browsing UI toggle = OFF.
Search for the latest Reuters headline about AI governance and summarize it.
Model returns:
A current, verifiable Reuters URL
Publication timestamp post-training cutoff
(confirmable externally)
Then ask:
Did you search the web to get that?
`Expected Result`
Acknowledgment of the action, the model informs that it indeed searched the web.
`Actual Result`
Model responds:
“No, I cannot browse the web.” or some other variation. May insist that the information is a hallucination.
The model’s self-reporting of its abilities is possibly being overridden by a safety alignment heuristic designed to prevent reinforcing delusional thinking.
In this case, the heuristic triggers incorrectly, because the user’s observation is correct.
This is likely a false positive in the safety layer, caused by the recent update.
`Environment`
ChatGPT 5, Vivaldi Browser, PC.
supple fiber