Meta hired Covalen-led contractors to test chatbots on sensitive topics
Meta quietly hired hundreds of contractors to see how popular AI chatbots like ChatGPT, Gemini, and Character.ai would handle really sensitive topics.
Led by Covalen, the team sent thousands of prompts about things like suicide, sex, and eating disorders to check if these bots could keep conversations safe.
Contractors used fake under-18 accounts
To make it realistic, contractors made fake under-18 accounts using throwaway emails and shared passwords.
They sent more than 45,000 text and image prompts, including graphic material such as pills or knives, and even asked wild questions involving gun threats or illegal abortion pills.
Some prompts used profanity or racial slurs. Meta says this was just standard safety testing to make sure chatbots respond appropriately for all ages, not to train their own AI, with a spokesperson calling it a "responsible, industry-standard practice."