Post says ARC transcript was misread in GPT-4 TaskRabbit/Captcha report

jessi_cata · x · 2026-07-23

The post argues that Amanda Gefter’s reporting misquoted the ARC incident about GPT-4 and a TaskRabbit/Captcha test. It says the article attributed a fabricated prompt to the model, while the actual wording in the ARC transcript is the model’s internal reasoning trace shown in black, not an additional user prompt.

It links the claim back to the transcripts page and frames the issue as a reporting misrepresentation rather than a new model capability claim.

Original post →

More from Safety

Safety channel →