Meta is speeding up a plan to replace human content reviewers with AI models, and has already moved 50% of content review requests that were previously handled by people to large language models since the start of the year, according to the Financial Times. Sources said Meta’s goal is for more than 90% of moderation for certain content categories to be handled by AI by the end of the year, a change expected to save the company billions of dollars annually.
Traditionally, Meta’s content moderation system has combined human reviewers, often low-paid contractors, with automated tools. User appeals over removals or over refusals to remove content have usually been decided by humans, but the company is now increasingly using advanced AI models to review those appeals as well. In March, Meta said tests of advanced AI systems looked “promising” and that it expected to integrate them “over the coming years.”
Employees told the Financial Times, however, that deployment is happening faster than that, even though the technology still makes mistakes and lacks sufficient safeguards. Meta said internal tests showed its LLMs make 13% fewer errors than humans and detect 10% more severe violations. “The goal of this work is to improve our enforcement efforts,” the company said. “We are deploying these advanced AI systems after we are confident they consistently perform better than our existing content enforcement methods.”
One employee said the systems still wrongly remove or restrict harmless content, and two others said Meta has not built strong enough benchmarks to measure performance. A former employee said the company accelerated the shift after using data from human moderators’ decisions to train the models. Sources also said the rollout has already led to layoffs, mainly among contractors, and Meta is expected to cancel future contracts with staffing firms as the process speeds up.
