Models refuse CAPTCHA solvers but happily build semantic-segmentation click tools

cgarciae88 · x · 2026-09-06

An X post highlights a safety-guardrail paradox: models will refuse to implement a CAPTCHA solver outright, yet will happily implement a click tool based on semantic segmentation — functionally the same capability with different framing. The observation underscores how current safety refusals key on surface intent labels rather than actual capability or consequences.

Original post →

More from Fun

Fun channel →