'AI bots trained to replace humans won't keep children safe online', says former Meta worker

Approximately 90% of Meta's content moderation will be done by bots by the end of 2026.

A former artificial-intelligence (AI) bot trainer has warned that Facebook and Instagram owner Meta's plans to rely on AI for content moderation will not keep children safe from explicit online material.

The woman, who spoke to the Irish Examiner anonymously, also raised concerns about skilled migrant workers being deported because robots will be trained to take their jobs.

The young woman, who spoke at the Labour Party think-in, worked as an AI annotator for Meta for 14 months. She had been contracted by outsourcing firm Covalen. An AI annotator helps train AI bots and models.

She was tasked with creating prompts and responses for AI bots, which she compared to "teaching a kid how to walk, how to eat, how to talk".

"All these things that you see in AI, it's someone training them," she said.

AI bots must be taught which content is and is not allowed, including topics such as suicide, self-harm, eating disorders, sex, and child exploitation.

The AI bot is taught what the content is and how to provide it. For example, AI trainers must "think like a predator", the woman said, and teach the bot what a child pornographic image is and where it can be found.

These workers are training bots that will "eventually take your job", the woman said.

The Financial Times previously reported that Meta plans to have 90% of its content moderation done by AI by the end of 2026.

However, the AI trainer insisted that bots cannot provide the same content moderation that humans can.

"At this point, based on my experience, I don't think relying on a machine to do this work will actually protect our children from watching this content," she said.

"We still have people behind computers evaluating this sensitive content, and there's still a lot of videos out there.

"Right now, we have people who are supposed to remove this, and they're working so hard on removing this sensitive content. Still, it's out there.

"I can't imagine a machine doing it by itself. We need someone who actually is analysing it."

Migrant workers are vulnerable and open to deportation when they are made redundant, the woman said.

Covalen announced in April that it was proposing to make 700 workers redundant at its office in Dublin.

If someone on a general employment permit loses their job, they have just six months to find another. They must then apply for a new employment permit, which costs €1,000 for two years.

The woman called for the Government to examine the inclusion of 'cooling-off periods' in contracts. These prohibit workers from being contracted for six months to companies that have made them redundant, even if another agency has jobs available with that company.

This could prevent some people from working, causing further visa issues.

A Meta spokesperson referred the Irish Examiner to an article it published in March arguing that AI systems can "catch more harmful content, stop more scams, and respond faster to real-world events".

"The results from our early testing have been promising. So far, across our tests, we're seeing these AI systems catch things that both review teams and current automation systems missed," it said.

Meta also said that it will "always have people involved in review of the most severe areas that involve reports to law enforcement, such as child safety, credible threats of violence, suicide or self-injury, and exploitation".