Discord’s captchas are run by a company called hCaptcha. “The technology that generates these prompts is proprietary to our third-party partner and Discord does not directly determine what is presented to users,” Discord told Motherboard. “While most hCaptcha interactions do not result in a visual challenge, many variants are used at any given time,” a spokesperson for hCaptcha told Motherboard. “This particular question was a brief test seen by a small number of people, but the sheer scale of hCaptcha (hundreds of millions of users) means that when even a few folks are surprised by a challenge this often produces some tweets.”
The issue with hCaptcha’s strange AI generated prompts highlights two issues with machine learning systems. The first is that the AI systems require an enormous amount of human input to not be terrible. Typically image labeling is outsourced to foreign workers who do it for pennies on the dollar. The other is the issue of data drift. The longer these machine learning systems run, the more input they require. Inevitably, they begin to use data they’ve generated to train themselves. Systems that train on themselves long enough become AI Hapsburgs, churning out requests to identify incomprehensible objects like “Yokos.”
Read more of this story at Slashdot.
Source: Slashdot