Researchers at Vicarious AI have developed a new brain-inspired AI technique that can solve CAPTCHAs with less training data than other AI techniques
Context & Ripple Effects
This lands mid-arms-race: months after Google rolled out invisible reCAPTCHAs that score users behaviorally instead of showing puzzles, Vicarious AI reports a brain-inspired technique that cracks CAPTCHAs with far less training data than conventional deep learning — attacking the assumption that bot defense wins by making puzzles machines can't learn cheaply.
The throughline runs forward too: seven years later, researchers would claim a model solving reCAPTCHAv2 at 100% accuracy within human-level attempt counts, and by then Google's escalating challenges were already straining human users. Vicarious's data-efficiency result is an early proof that sample-hungry models were never the only threat.
First-order effects
- Sites using image-based CAPTCHAs lose their core defense premise: if a brain-inspired model needs little training data, defenders can't rely on puzzle novelty outpacing what attackers can afford to train on.
Second-order effects
- Google is pushed further down the path it had already started with invisible reCAPTCHA — behavioral scoring rather than solvable puzzles — which per the later reporting makes challenges progressively harder for legitimate humans while bots keep pace.
Third-order effects
- If sample-efficient architectures keep closing the gap, human-verification migrates from 'prove you can solve this' to opaque trust signals, and the CAPTCHA as a human-vs-machine test becomes structurally obsolete — with the cost borne by users who fail the heuristics.
The trend: Bot-detection is shifting from puzzles designed to be hard for machines toward behavioral and trust-based signals, as data-efficient AI erodes the training-cost barrier that puzzles relied on.