May 29, 2026
Botched by the bot test
CAPTCHAs can still detect AI agents
Bots may pass the test, but commenters say the internet war is far from over
TLDR: Researchers say AI can now get CAPTCHA answers right but still reveals itself by acting differently from humans during the task. Commenters weren’t wowed: many called it another endless bot-vs-filter arms race, while others said real users get punished more than the spammers.
The big claim from Roundtable is deliciously simple: today’s AI can solve CAPTCHAs, but it still doesn’t behave like a human while doing it. In plain English, the bot might click the right boxes, but it clicks them in a weirdly bot-ish way. The researchers say those patterns — like how it moves through choices and what mistakes it makes — can still give machines away.
But in the comments? Nobody is ready to declare victory. The loudest reaction was basically: cool paper, same old arms race. One camp shrugged that this is just another round of cat-and-mouse, with bots eventually learning the “human” style too. Another group brought receipts from the trenches, saying real-world spam filters are already failing. One commenter dragged Cloudflare’s Turnstile by saying their forms are still getting hammered with junk, which turned the thread into a mini roast of modern anti-bot tools.
Then came the most relatable outrage: regular humans say they’re the ones getting punished. Privacy-minded users complained that ad blockers, VPNs, and locked-down browsers get treated like criminal masterminds, while actual bad actors slide by with shady tools and borrowed home internet addresses. That sparked the thread’s most bitter joke: CAPTCHAs aren’t catching bots — they’re just making life miserable for normal people. So yes, the researchers say bots still have tells. The commenters say the bigger story is the chaos, the workarounds, and the internet’s ongoing identity crisis.
Key Points
- •The article says AI systems can solve classic CAPTCHA tasks at near-human levels, but their task-completion behavior differs from that of humans.
- •The cited paper reports statistically significant human-versus-AI differences in CAPTCHA process features including sequential click patterns, direction changes, and overselection.
- •The authors introduce CogCAPTCHA30, a benchmark combining one CAPTCHA task with 29 cognitive psychology tasks.
- •The article reports that output equivalence and process equivalence across the 30-task battery were uncorrelated.
- •Based on these findings, the article proposes a Process Turing Test focused on whether machine processes, not just outputs, are human-like.