Summaries like this, in your inbox every morning.
Sign up free →What happened
Whistleblowers told New Scientist that people employed by third-party contractors to provide high-quality training data for AI models are using chatbots like ChatGPT to complete tasks instead, despite company policies forbidding it. One worker said the practice is "very widespread" and difficult to catch if workers mask telltale signs of AI output. A training platform called Outlier, owned by Scale AI, monitored workers with screenshot tracking software called Hubstaff but workers found ways to hide their use of AI models.
Why it matters
AI models trained on AI-generated content rather than human-produced data can experience "collapse," where their abilities drop dramatically, according to research cited by Mark Lee at the University of Birmingham. Even if full collapse does not occur, the cheating degrades model performance—models become "less good at doing human-like tasks." The problem is compounded by working conditions: these contractors typically lack full-time contracts and work for low pay, which incentivizes cutting corners.
What to watch
Lee notes that if human-generated data makes up roughly 10 percent of training material, it can mitigate model collapse. The risk scales with how much AI-generated content enters future training pipelines—companies including Meta, Cisco, and Google are relying on these contractors, though none responded to requests for comment.
No discussion yet for this article
Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.
Get Started FreeFree · takes 30 seconds · unsubscribe anytime
5 minutes a day. The AI essentials.
200+ sources · Email / LINE / Slack
