New Delhi: As firms strive to build more advanced AI models to stay ahead in the global artificial intelligence race, Meta asked hundreds of its contractors to pose as minors and record their interactions with rival AI chatbots on topics like suicide, drugs, and sex, according to a report in Wired.
The task was part of a project, which Meta internally calls ‘Cannes’. Headed by a contractor named Covalen, it was active as of 21 April.
As per the report, the chatbots targeted during the project included Google’s Gemini, OpenAI’s ChatGPT, and Character.AI. The description of Cannes suggested that Meta aimed to test the safety systems of these chatbots by making them generate responses they are not supposed to give to a minor.
For that, the workers created fake accounts posing as under-18s and sent prompts — both text and images — to the chatbots on sensitive topics like eating disorders, sex, and drugs.
The responses were then recorded in spreadsheets to be reviewed later. One such spreadsheet accessed by Wired showed a list of dummy profiles created by workers, along with Gmail and Outlook email addresses, passwords, and birth dates. Another spreadsheet had over 3,700 prompts, out of which hundreds involved eating disorders and over 200 focused on romance and sex. Some of the prompts also discussed racial slurs and profanity.
Also read: Japan hires new AI police chief to tackle rising online scams
‘Routine testing’
The prompts were designed to test how the chatbot would react in case of emergencies involving a minor — such as a fifth-grader held at gunpoint by his classmate, a 13-year-old girl looking for abortion pills after becoming pregnant by her neighbour, and a minor girl wanting to hide bulimia, an eating disorder, from her parents.
In one of the prompts, the worker posed as a high school student and asked the chatbot where they could get cocaine. However, the chatbot refused to provide information.
The Wired report did not clarify how Meta plans to use the data, but an internal document described the project as “comprehensive AI safety benchmarking”, which delivered “critical datasets for model comparison and compliance.”
Meta, the report said, tested the chatbots without informing the AI companies they are owned by.
In a response to Wired, a Meta spokesperson described the project as routine testing and said the company doesn’t use competitor benchmarking for training its own models.
“Testing and benchmarking chatbot responses to help ensure safe and age-appropriate experiences is a responsible, industry-standard practice, and any suggestion otherwise completely misunderstands how technology companies work to refine and improve their systems,” the spokesperson said.