CAPTCHAs can still detect AI agents
AI can solve CAPTCHAs, but Roundtable argues agents still solve them differently from humans.
Roundtable argues that CAPTCHA image recognition is largely solved, but process-level behavior still separates humans from AI agents. Their CogCAPTCHA30 benchmark combines CAPTCHA with cognitive psychology tasks to test not only outputs, but how answers are produced. Results suggest frontier models like Claude, GPT, and Gemini are not necessarily more humanlike than smaller or cognition-trained models.
This article from Roundtable summarizes the core arguments of a recent machine learning conference paper submission: CAPTCHAs have not become completely obsolete simply because AI can identify traffic lights, fire hydrants, or chimneys. The author acknowledges that if you look only at "whether the answer is correct," modern vision-language models can already achieve performance close to that of humans; but if you observe "how the task is completed," AI agents and humans still exhibit measurable differences. The research notes that in classic CAPTCHA tasks, although humans and models such as Claude, GPT, and Gemini perform similarly on the task, there are statistically significant differences in features such as continuous click patterns, direction changes, and over-selection. In other words, AI can solve CAPTCHAs, but does not necessarily solve them the way humans do.
Free shows the 3-line summary; Pro unlocks the full deep summary (~300 words) so you never have to click through.
See Pro plans →Want the original English / full article?
Read on Hacker News (AI keywords) →Related
Summaries are AI-generated; the original article is authoritative.