Meta increasingly relies on AI moderation instead of human reviewers
Meta aims to largely automate moderation on Facebook and Instagram by the end of 2026. LLMs are intended to replace human reviewers and increase efficiency.
(Image: Koshiro K / Shutterstock.com)
Meta aims to largely automate content moderation in its networks by the end of the year. According to the Financial Times, Large Language Models (LLMs) have already replaced half of all human review requests. By the end of 2026, this proportion is expected to decrease further, with AI handling nine out of ten moderation tasks for certain content types. Informants suggest Meta is primarily pushing this transformation to co-finance the company's enormous AI investments. The use of AI is said to save billions of US dollars. Meta itself denies this. As early as the beginning of 2026, Meta CEO Mark Zuckerberg announced that individual employees would be equipped with AI tools to take over tasks of entire departments.
The parent company of Facebook, Threads, and Instagram has so far used a mix of automated systems and human reviewers, with some of these tasks outsourced to contract companies. However, previous machine learning tools have sometimes struggled with satire and changes in everyday language. Meta itself justifies the AI push, according to the FT, with quality improvement. Internal tests have indicated that LLMs make 13 percent fewer errors on average and find 10 percent more violations than humans.
Switch from Google Gemini to Muse Spark
The transition will lead to job cuts. Initially, contract companies whose contracts are not renewed will be primarily affected. As early as spring, Meta had announced that it would lay off around ten percent of its workforce while simultaneously transferring thousands of employees to new AI departments. Ultimately, a high degree of automation is likely to lead to personnel savings at Meta itself as well.
Videos by heise
Previously, Meta is said to have used Google Gemini for AI moderation. The LLMs were trained, among other things, on previous decisions by human reviewers in appeal procedures. However, employees have now been instructed to switch to Meta's own model, Muse Spark, the first model in the Muse series from Meta's Superintelligence Labs. As early as June 2025, Meta had begun to replace risk assessments largely for new functions with AI – at that time, it concerned data protection and integrity assessments for Facebook, Instagram, and WhatsApp. According to FT information, there is internal criticism that AI moderation is being rolled out too quickly. The LLMs made mistakes such as blocking or making harmless content invisible. Furthermore, there are insufficient metrics for performance measurement. Meta rejects this.
(mki)
