AI Trainers Admit Using ChatGPT to Do Their Jobs
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Whistleblowers told New Scientist that paid AI trainers commonly use chatbots like ChatGPT to complete their assigned tasks, with one worker calling the practice "very widespread" across multiple companies despite explicit anti-AI policies
- Outlier, a training platform owned by Scale AI, used Hubstaff screenshot monitoring to catch workers running AI models; worker Bob* says people hid ChatGPT in tabs or minimized windows but were often caught via desktop folders or task bars
- Scale AI claims on its website to do work for Meta, Cisco, and Google (the latter per Bob's account); none of the three companies responded to New Scientist's request for comment
- Worker Carol* began using LLMs to check her work against lengthy guidelines out of fear of expulsion and income loss, then escalated to running entire tasks through multiple LLMs to create scenarios and supporting files
- Mark Lee at the University of Birmingham said recursively training AI on AI-generated content causes model "collapse," but noted roughly 10% human data mitigates this — the real-world effect would be degraded performance rather than catastrophic failure
Why it matters: The third-party gig model powering AI training pipelines — short contracts, low pay, minimal job security — creates direct incentives for workers to cheat, and major clients like Meta, Google, and Cisco could end up with diminished model quality if Scale AI's training data is contaminated by AI-generated content rather than human-crafted exchanges.
Ask SkimNews
