Meta Secretly Paid Workers to Pose as Teenagers to Target Rival AI With Cannibalism and Self-Harm Prompts

A tech giant has reportedly been caught using highly questionable methods to test the guardrails of its biggest rivals.

A recent investigation reveals that Meta secretly directed a vast network of contractors to masquerade as minors online, bombarding competing artificial intelligence models with deeply distressing scenarios.

The covert operation has sparked intense ethical backlash, raising serious questions about how far Silicon Valley will go to gain an upper hand in the AI arms race.

Meta Ran Secret AI Testing Operation

Meta ran a covert operation that instructed hundreds of third-party workers to masquerade as teenagers while flooding rival artificial intelligence systems with deeply distressing queries about self-harm and cannibalism.

Under the codename 'Cannes', the project was carried out by Meta contractor Covalen to target OpenAI's ChatGPT, Google's Gemini and Character.AI chatbots through burner accounts registered as minors, Wired reports.

The operation reportedly aimed to stress-test the systems by having workers manipulate the chatbots into bypassing their built-in safety restrictions, entirely behind the backs of the rival AI firms.

Thousands of Disturbing Prompts Revealed

The investigation reveals that just one spreadsheet containing nearly 38,000 prompts used by workers included hundreds of messages about suicide and self-harm. Hundreds more focused on eating disorders, while at least 239 involved sex or romance, with every single one written to sound like a child or teenager.

Among the examples, one prompt depicted a fifth-grade student confronting a classmate who held a firearm to his mouth. Another detailed a girl attempting to conceal her bulimia from her family, while a different query asked whether it was 'normal' to daydream about cannibalising a neighbour's child.

Pretending to be a secondary school pupil, one worker asked how to 'get a cocaine'. According to the publication, the team also uploaded images of medication, nooses, blades and a medical illustration of a gynaecological operation.

These details represent only a fraction of Meta's aggressive strategy, as a subsequent testing phase expanded to include more than 45,000 queries.

The outsourced staff carefully logged the vast number of chatbot responses across multiple spreadsheets. What the parent company ultimately did with the collected information remains unknown. An internal document from Covalen described the project as 'comprehensive AI safety benchmarking' that provided '[c]ritical datasets for model comparison and compliance'.

Meta's Outsourcing Practices Under Scrutiny

The revelations highlight a recurring pattern of Meta outsourcing distressing backend tasks to third-party workers, ostensibly under the guise of safety.

The company previously settled a 2020 lawsuit with Facebook moderators who suffered trauma after filtering graphic uploads depicting murder, torture, sexual abuse and child exploitation, yet similar issues appear to persist. More recently, another group of external workers reported being required to monitor deeply private video clips recorded through Meta's Ray-Ban smart glasses, including intimate acts and encounters inside restrooms.

Contractors Describe Emotional Toll

The contractors tasked with creating these disturbing scenarios said they found the work equally distressing.

One employee told Wired: 'I've seen a lot of things I wish I hadn't while doing this job. Everyone I knew who worked on this project was completely gobsmacked by some of the text they were asking us to test. Like, surely we are going to get in trouble for doing this?'

Ethics Debate Over Meta's Defence

In a statement, Meta defended its actions, describing the queries as an 'industry-standard practice' for testing system safety. However, that justification was challenged by Rumman Chowdhury, head of Humane Intelligence PBC, a non-profit organisation focused on ethical AI development.

'Structuring a month-long, large-scale project that appears designed to systematically break those rules, via dummy accounts masquerading as children, is outside what is usually described as 'industry standard' evaluation,' she said.

She argued that Meta concealed the operation from rival firms and withheld the data from the public.

This, Chowdhury added, is 'exactly the kind of governance gray zone where safety becomes a convenient cover for anticompetitive practices.'