A look at hCaptcha, a reCAPTCHA alternative used by Discord and others, which trains ML systems and generative adversarial networks using AI-generated images
Discord's captcha is asking users to identify a ‘Yoko,’ a snail-like object that does not exist and was created by AI. — Matthew Gault Tweets: @jeffjarvis , @lazerwalker , @gregegansf , @motherboard , and @vice Tweets: @jeffjarvis : This is how the computers will defeat us: Find the yoko or you can't enter. https://www.vice.com/... Emilia / @lazerwalker : The SCP entries are bleeding into the real world https://www.vice.com/... Greg Egan / @gregegansf : I like this coinage by a Vice journalist: ‘Systems that train on themselves long enough become AI Hapsburgs [sic], churning out requests to identify incomprehensible objects like “Yokos.”’ (the correct spelling is “Habsburg") referring to inbreeding. https://www.vice.com/... @motherboard : Discord's captcha is asking users to identify a ‘Yoko,’ a snail-like object that does not exist and was created by AI. https://www.vice.com/... @vice : Do you know what a ‘Yoko’ is? Of course you don't because it's not real. https://www.vice.com/...
Context & Ripple Effects
The story lands at the end of a long escalation: researchers at Vicarious showed as early as 2017 that neural nets can solve CAPTCHAs with far less training data than assumed (a brain-inspired technique cracked them), and by 2019 Google's challenges were already getting too hard for humans as recognition models matched human performance (the Verge documented that squeeze). Designers responded by making prompts more difficult — by 2024 the shift was toward logic-based puzzles precisely because image labeling got too easy for software (CAPTCHA designers adopting harder logic-based prompts).
hCaptcha's move, used by Discord among others, changes the game rather than just raising difficulty: instead of asking humans to verify real-world images, it asks them to classify objects that never existed — like Discord's snail-like 'Yoko' — while harvesting those human judgments to train machine-learning systems and GANs on AI-generated imagery. A Vice journalist's coinage captured the worry: systems that train on themselves long enough risk becoming what he called AI Hapsburgs.
First-order effects
- Every Discord user passing a hCaptcha challenge is now doing unpaid labeling work on synthetic imagery — verifying objects like the Yoko that exist only as model outputs — which directly improves hCaptcha's ML and GAN training sets.
- Users who fail to recognize an object that was never real are locked out of the service, meaning the cost of the arms race now falls on legitimate humans, not just bots.
Second-order effects
- Google's reCAPTCHA and other verification providers face pressure to adopt the same synthetic-prompt approach, since real-image challenges are provably solvable by modern classifiers — pushing the whole market toward AI-generated test objects.
- Bot operators gain a new target: if enough humans label synthetic objects through captchas, adversaries can harvest those labels to train their own generators, turning the verification layer itself into a labeled dataset.
Third-order effects
- If captcha providers keep training on model outputs validated by humans, the web's trust infrastructure becomes a closed loop between generative models and human judgment — the self-reinforcing dynamic behind the AI-Hapsburgs concern, where each generation of training data is one step further from ground truth.
- This points toward verification decoupling from reality entirely: proving you are human may come to mean proving you can reason about things no human has ever seen, a structural inversion of what CAPTCHAs were designed to do — and part of the same pattern visible when generative systems flood platforms with content trained on themselves, as with Sora 2 regurgitating copyrighted characters.
The trend: Human-verification systems are becoming training pipelines for generative models, closing a loop where AI creates the tests, humans grade them, and the grades train the next generation of AI.