Meta wants generative AI to take over moderation because, apparently, human judgment was too pricey and squishy.
According to the Financial Times the $1.4tn tech outfit is racing to replace human content and advert reviewers with large language models.
Four people familiar with the matter said Meta has sped up plans to use LLMs to review posts and advertising across its platforms.
The shift could save billions of dollars a year. Meta has already replaced about half of human review requests with LLMs this year, several people said. The company wants that figure cut further by the end of the year, potentially by more than 90 per cent for some kinds of content.
For years, Meta has used automated systems and human reviewers, including third-party contractors, to decide whether posts or adverts broke its rules. User appeals have usually been handled by humans, which at least gave the system some chance of noticing sarcasm, context or plain old nonsense.
Meta insisted the move towards AI moderation was about using fast-developing technology better, rather than saving cash. Since March, Meta said its early tests showed LLMs made 13 per cent fewer mistakes than humans when enforcing rules against violating content.
It said the systems found 10 per cent more genuine violations. This will be news to those who have laid complaints against people spreading hate or had bogus complaints laid against them and kicked off the site.
“The point of this work is to improve our enforcement efforts, and we’re deploying these more advanced AI systems once we’re sure they’re consistently performing better than our current methods of content enforcement,” it added.
Zuckerberg is spending billions trying to build what he calls “personal superintelligence”, with hyper-personalised AI products and agents for users.
Meta is trying to cut operating costs elsewhere by using AI to automate internal work such as coding, several people said.
Meta has been using Google’s Gemini large language model for most moderation and customer support. Staff have recently been told to switch to Meta’s new foundational model, Muse Spark, the people said.
Traditional AI moderation usually relies on machine-learning classifiers to flag rule-breaking material. That kit can struggle with nuance, satire and evolving language, which is awkward when the internet is mostly nuance, satire and evolving language.
Some staff warn the rollout is now moving much faster and that the technology still makes errors.
One Meta insider said LLMs moderating content kept wrongly removing or “shadow-banning” harmless material. Two people said Meta had not properly worked out how to measure the technology’s performance, a claim the company denied.
Another former employee said Meta accelerated the shift after gathering plenty of data from earlier human decisions on appeals. That data helped train and improve the LLMs, the person said.
Advertising could prove especially touchy, since Meta is already under pressure over scam promotions on its platforms. Last year, Reuters reported that Meta had internally forecast it would earn about 10 per cent of its 2024 annual revenue, or $16bn, from adverts for scams and banned goods.
The company faces legal challenges over scam ads from California’s Santa Clara County and Australian mining magnate Andrew Forrest.
