In a recent security-incident document, Anthropic disclosed that its model Claude still has trouble with CAPTCHAs. The transcript reportedly shows the model struggling with a basic puzzle, an outcome that might surprise those who assume large language models have long surpassed such tests. Security expert Bruce Schneier highlighted the anecdote on his blog, using it as a reminder that seemingly trivial human tasks remain hard for AI.
This is not about a lack of raw intelligence. The model in question is powerful enough that Anthropic restricts access to it. Yet it apparently could not solve the CAPTCHA without visible difficulty, appearing to "slam its virtual head against the wall." The contrast between the model's advanced reasoning and its failure on a simple visual or interactive challenge underscores how narrow AI competence can be.
Schneier's commentary does not offer a solution, but the incident suggests that CAPTCHAs still serve their purpose as a gatekeeper. It also raises questions about how far AI must progress before such measures become obsolete. For now, the human-versus-machine distinction appears to hold, at least for this particular test.