Meta AI Content Moderation Hits 50%, Eyes 90% by Year’s End

Meta has already routed roughly half of its content moderation through large language models, and is preparing to push that share past 90 percent for some content categories by the end of 2026, according to a recent report on Meta’s AI moderation pivot. Citing sources familiar with the matter, the FT writes that the shift could save Meta billions of dollars annually. A Meta spokesperson frames the rollout as an accuracy play, not a headcount one.

The change lands inside a Meta now spending up to $145 billion on AI infrastructure, has shed roughly 8,000 jobs in the past month, and has begun telling outside vendors including Accenture, Concentrix, and Teleperformance to plan for a smaller moderation role. The contractor pipeline is closing, per the FT, even as Meta employees told the paper the AI replacing those reviewers is still misfiring on calls the company says must be made faster.

The Numbers Meta Is Working From

Roughly half of Meta’s content and ad review work is already being handled by large language models, according to people familiar with the rollout cited by the FT. For some categories of content, Meta is preparing to push the AI share above 90 percent by year’s end.

The financial case is what gives the figure its pull. Sources told the FT the change could save Meta billions of dollars annually. Meta has publicly rejected the cost framing, telling the FT the rollout is about error rates and violation catch rates. The shift also dovetails with a parallel push at Meta to automate coding and other operating expense lines.

By the numbers:

• 50% – share of content and ad review Meta says large language models already handle, per the FT
• 90%+ – target share for some content categories by year end 2026
• 13% – fewer moderation mistakes LLM tests reportedly made vs humans since March
• 10% – more actual policy violations LLM tests reportedly caught
• $145 billion – AI infrastructure budget Meta has flagged for this year

What Meta Says It Stands On

In internal tests that began in March, Meta’s large language models made 13 percent fewer moderation mistakes than human reviewers and caught 10 percent more actual policy violations, the FT reports. Both figures came from the same Meta spokesperson who confirmed the overall pivot to the paper. The same Meta statement also framed the rollout’s threshold for wider deployment.

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.

A Meta spokesperson, in a statement to the Financial Times. The FT’s full exchange with the spokesperson was not released by Meta at the time of the report.

In a March blog post, Meta had already laid out the case for letting AI handle the repetitive parts of moderation work, from graphic content reviews to fast-evolving scam tactics. The company wrote that humans would stay involved in the “most complex, high-impact decisions” such as appeals of account disablement and law-enforcement referrals. The FT report now describes how much further internal test results have moved that line.

The framing tracks Meta’s earlier public posture on free expression. In January 2025, Meta said its automated systems had been making “too many mistakes” and started using AI large language models to provide a “second opinion” on harder calls, per Meta’s January 2025 free-expression framework. By June 2026, the second opinion is fast becoming the first one.

The Contractor Pipeline Closes

The shape of Meta’s moderation network has always been largely outsourced. The work has run through Accenture, Concentrix, Teleperformance, with subcontractors handling graphic content reviews, scam takedowns, and appeals. Meta’s March blog said AI would replace “work that’s better-suited to technology,” the kind of repetitive reviews that human reviewers had spent most of their time on. The FT report describes a much deeper cut, with multiple outside contracts now expected to be cancelled and layoffs underway at the vendor firms, alongside a Kenyan subcontractor suit this past spring over how workers handled intimate recordings from Meta’s Smart Glasses.

Meta trained the models replacing those contractors on past decisions made by human reviewers, the FT reports. The training data is the same one Meta cited as the source of its accuracy claims. The categories of work shifting first, Meta’s March blog post and the FT report together outline:

• Reviews of scams and removal of illegal media (Meta’s explicit March example)
• Repetitive reviews of graphic content
• Most ad-review decisions, where the LLM share is set to grow fastest
• First-line appeals of posts users say were wrongly removed
• Routine moderation of evolving internet slang, sarcasm, and edge cases, the work Meta says generative models handle better than older keyword filters

Where the Recourse Goes Next

For Meta’s advertisers, the speed-up is the point and the risk. Trade publication eMarketer warned that Meta’s AI moderation shift “could make enforcement faster but less predictable, raising brand safety risks and giving advertisers less clarity or control over content and account decisions.” The concern is concrete: when an AI decision removes or downgrades content near an ad, the appeal path shrinks with it. The same trade publication also flagged the broader risk that advertisers will end up paying for scam-ad adjacency the new system was supposed to remove.

The picture for ordinary users is harder still. Meta’s Oversight Board ruled on June 4 that the company’s account deactivation system lacks due process, that violations are “doled out without clarity,” and that there is “little customer support for appeals,” while also calling on Meta to disclose when AI is enforcing a rule.

The Board’s case file is full of users locked out with no specific post cited, per TechCrunch’s coverage of the ruling. Richard Pauwels, a retired LA County firefighter and paramedic building a wellness brand, had his personal account banned without any human review by Meta’s systems. A 60,000-follower pigeon rescue operation was banned for alleged child exploitation material, also without citations. A separate business owner quoted by TechCrunch said he suffered “ongoing business disruption, reputational harm, and significant stress” after Meta would not put his case in front of a human reviewer.

One insider told the FT the models still remove or shadow-ban harmless content and that there isn’t enough oversight for the pace of the rollout, even as Meta’s accuracy claims continue to anchor its defense. Meta had been using Google’s Gemini and recently told staff to switch to its own foundation model Muse Spark, also per the FT.

Meta’s own January 2025 estimate was that one to two out of every ten takedown actions might have been mistakes, after a long run of over-enforcement. The AI rollout is now being sold as the fix. The broader framework on automated content moderation decisions has long called for humans to remain in the loop on high-stakes outcomes.

Why the Pivot Lands Now

The savings math is anchored to Meta’s AI bill. The company has flagged up to $145 billion in AI infrastructure spending for this year, with the moderation cuts being one place Meta is finding room to pay for it, sources told the FT.

Cost pressure has been visible for months. Reuters reported in March that Meta was considering whether to cut as much as 20 percent of its workforce to balance AI spending; Meta called that “a speculative report about theoretical approaches.” Meta has since laid off about 8,000 employees, roughly 10 percent of its global workforce. A separate Zuckerberg note to staff acknowledged the AI overhaul had produced mistakes.

Mark Zuckerberg has framed the broader rebuild as a path to “personal-level superintelligence,” the marketing phrase now tied to Meta’s infrastructure build and hiring spree. The FT report and the Chosun read of it place the moderation shift alongside other operating-expense reductions across coding and internal automation. Meta’s public framing remains that the rollout is about accuracy, not headcount.

Frequently Asked Questions

What is Meta changing about content moderation?

Meta is moving large language models, rather than human reviewers, into the first line of decision on most Facebook and Instagram posts, ads, and appeals. The Financial Times reported on June 24, 2026 that Meta has already replaced roughly half of those reviews with AI and is preparing to push the share past 90 percent for some content categories by the end of the year.

How much human moderation work has Meta already replaced?

About half. The FT, citing people familiar with the rollout, puts the current share of AI-handled review at 50 percent. Meta confirmed the figure to the FT and said it is targeting a 90 percent-plus share in several content categories by year’s end.

What is Meta’s accuracy claim for its AI moderation?

Meta told the FT that internal tests since March show its AI systems make 13 percent fewer moderation mistakes than human reviewers while catching 10 percent more actual policy violations. Meta has not made the test methodology or underlying data public.

Which outside contractors are affected?

Meta’s outsourced moderation has run through Accenture, Concentrix, and Teleperformance, with subcontractors handling graphic content reviews and appeals. Meta’s March blog framed AI as the replacement for the most repetitive parts of that work, and the FT report describes multiple contract cancellations now underway at the vendor firms.

Are users losing any form of appeal under the new system?

Some pathways are narrowing. Meta says humans will remain involved in the “most complex, high-impact decisions,” including appeals of account disablement and law-enforcement referrals. The Oversight Board ruled on June 4, 2026 that even those pathways were already failing users, citing a lack of transparency and consistency and recommending that Meta disclose when AI is enforcing a rule. The recent exit of Meta’s AI transformation lead after two months on the job is the latest sign of internal turbulence around the same overhaul.