Meta Platforms is dismantling the human infrastructure that has policed its social networks for more than a decade. The company has begun cutting back on third-party content moderators and shifting that work to artificial intelligence systems, a move that will reshape how billions of posts, images, and videos are screened across Facebook, Instagram, and Threads. The scale is staggering. The implications are messy.
According to The Information, Meta has started reducing its reliance on the armies of contract workers — many employed through outsourcing firms like Accenture and Teleperformance — who have long served as the front line against hate speech, misinformation, graphic violence, and other harmful content. The plan, which accelerated in early 2025, calls for AI models to handle an increasing share of moderation decisions that were previously routed to human reviewers.
This isn’t a quiet operational tweak. It’s a philosophical reorientation of one of the most consequential functions in the modern information economy.
Mark Zuckerberg telegraphed the shift in January when he announced that Meta would end its fact-checking program in the United States and replace it with a community-driven system modeled on X’s Community Notes. At the time, Zuckerberg framed the decision as a response to overreach — too many false positives, too much suppression of legitimate speech, too many complaints from conservative users and politicians who accused the platform of censorship. “The recent elections also feel like a cultural tipping point towards, once again, prioritizing speech,” Zuckerberg said in a video statement posted to Facebook. But the fact-checking overhaul was only the most visible piece. Behind it sat a broader restructuring of Meta’s entire moderation apparatus, one in which AI would assume responsibilities that tens of thousands of human workers once carried out.
The numbers tell a striking story. Meta has employed, directly and through contractors, roughly 40,000 people working on safety and security across its platforms. That figure, which the company has cited in congressional testimony and public reports, includes content reviewers scattered across operations in countries like the Philippines, India, Kenya, and several European nations. These workers — often paid modest wages, sometimes as low as $1.50 per hour in developing countries — have been tasked with viewing and adjudicating some of the most disturbing content on the internet. Beheading videos. Child sexual abuse material. Terrorist recruitment propaganda. The psychological toll has been well-documented, including in lawsuits filed by former moderators who developed PTSD, anxiety disorders, and substance abuse problems after years of exposure.
So there’s a human cost to the current system that rarely gets discussed in the free-speech framing. And there’s a financial cost that Meta’s shareholders understand perfectly well.
Content moderation at scale is extraordinarily expensive. Meta spent more than $5 billion on safety and security in 2023 alone, according to figures the company shared with investors. A significant portion of that went to third-party moderation contracts. Replacing even a fraction of those workers with AI systems that can operate around the clock, across every language, without bathroom breaks or mental health leave, represents an enormous potential savings. Wall Street has noticed. Meta’s stock has performed well in 2025, and analysts at firms including Morgan Stanley and Bank of America have cited the company’s AI-driven efficiency gains — across moderation, ad targeting, and content recommendation — as key drivers of margin expansion.
But efficiency and accuracy aren’t the same thing.
Meta’s AI moderation systems, built on large language models and computer vision technology developed internally, have improved significantly over the past three years. The company has said its AI can now detect and remove the vast majority of violating content before any user reports it — a metric it calls “proactive detection rate.” For categories like spam and fake accounts, that rate exceeds 99%. For hate speech, it’s lower but still high, hovering around 95% according to Meta’s most recent Community Standards Enforcement Report. The company has pointed to these numbers as evidence that machines are ready to take over.
Critics aren’t so sure. And the critics have data too.
A report published by the Oversight Board — the quasi-independent body Meta itself created to review its most difficult content decisions — found persistent problems with the company’s automated enforcement. In multiple cases reviewed in 2024, the board identified content that was incorrectly removed by AI systems, including satire, news reporting, and political commentary that the models misclassified as hate speech or violence. The board also flagged cases where genuinely harmful content slipped through automated filters entirely, particularly in non-English languages where Meta’s AI models have less training data and lower accuracy. Arabic, Burmese, Amharic, and Tigrinya were singled out as languages where moderation failures had real-world consequences, including in contexts of armed conflict and ethnic violence.
This is the tension at the heart of Meta’s bet. AI works well in English. It works well on content that looks like content it’s been trained on. It struggles with context, nuance, slang, coded language, and the thousand small cultural signals that a human moderator — even an underpaid, overworked one — can sometimes catch.
Consider a simple example. The phrase “burn it all down” could be an expression of political frustration, a reference to a television show, or a genuine incitement to arson. A human reviewer, given context about the post, the poster’s history, and the surrounding conversation, can usually make the right call. An AI model operating at scale, processing millions of posts per hour, has to make probabilistic judgments. It will get most of them right. The ones it gets wrong can matter enormously.
The timing of Meta’s moderation overhaul has drawn scrutiny from researchers and advocacy groups who see it as politically motivated. The announcement came weeks before Donald Trump’s inauguration in January 2025, and Zuckerberg made a series of public gestures that appeared designed to align Meta more closely with the incoming administration. He visited Mar-a-Lago. He appointed Dana White, the UFC president and Trump ally, to Meta’s board of directors. He donated $1 million to Trump’s inaugural fund. And he restructured Meta’s content policies in ways that relaxed restrictions on topics politically sensitive to conservatives, including immigration and gender identity.
The New York Times reported that Meta’s decision to end fact-checking was made after extensive internal debate, with some senior executives warning that the move could increase the spread of health misinformation and election-related falsehoods. Those warnings were overruled. Joel Kaplan, Meta’s chief global affairs officer and a former George W. Bush administration official, was described as a key architect of the policy shift.
The political dimension matters because content moderation has never been a purely technical problem. It’s a governance problem. Every decision about what stays up and what comes down reflects a set of values, priorities, and risk tolerances. When those decisions are made by humans operating under written policies, there’s at least the possibility of accountability — appeals, audits, public reporting. When they’re made by AI systems operating inside a black box, accountability gets harder. Much harder.
Meta has said it will continue to employ human reviewers for the most sensitive content categories, including child exploitation and terrorism. The company has also said it will maintain human oversight of AI moderation decisions through sampling and quality assurance processes. But the details of how many humans will remain, where they’ll be located, and how much authority they’ll have over AI decisions remain unclear. The Information’s reporting suggests the cuts to third-party moderation staff are already underway and will continue through the rest of 2025.
Other tech companies are watching closely. Google’s YouTube has invested heavily in AI moderation but has maintained a large human review workforce, particularly for ads and content aimed at children. TikTok, facing its own regulatory pressures, has expanded both its AI moderation capabilities and its human moderation teams, particularly in Europe where the Digital Services Act imposes specific transparency and staffing requirements. X, under Elon Musk’s ownership, went in the opposite direction — gutting its trust and safety team in late 2022 and early 2023, a move that researchers at Stanford’s Internet Observatory and other institutions linked to increases in hate speech and misinformation on the platform.
Meta appears to be charting a middle path. More AI than YouTube. More humans than X. But the direction is unmistakable.
The financial logic is compelling. Meta reported revenue of $164.7 billion in 2024 and is projecting capital expenditures of $60 to $65 billion in 2025, most of it directed toward AI infrastructure — data centers, chips, and the engineering talent to build and deploy models. Every dollar saved on human moderation is a dollar that can be redirected toward these investments. And the AI systems being built for moderation aren’t standalone products; they’re extensions of the same large language models Meta is deploying across its consumer products, its advertising platform, and its enterprise AI offerings. Training a model to detect hate speech also makes it better at understanding language generally, which makes it better at recommending content, targeting ads, and powering Meta’s AI assistant.
It’s all one system now. That’s the point.
But there are risks Meta can’t model away. Regulatory risk is chief among them. The European Union’s Digital Services Act, which took full effect in 2024, requires very large online platforms to conduct systemic risk assessments and maintain adequate resources — including human resources — to mitigate those risks. If Meta’s AI-first moderation approach leads to measurable increases in harmful content in EU member states, the company could face fines of up to 6% of global revenue. That’s nearly $10 billion based on 2024 figures. European regulators have already opened investigations into Meta’s handling of political advertising and child safety, and the moderation cuts could provide additional grounds for enforcement action.
In the United States, the regulatory picture is murkier. Section 230 of the Communications Decency Act continues to shield platforms from liability for most user-generated content, though the law’s future is uncertain. Both Republicans and Democrats have proposed reforms, albeit for different reasons. Republicans want to prevent platforms from removing content they consider protected speech. Democrats want to hold platforms accountable for amplifying harmful content. An AI-driven moderation system could become a target for both camps — too aggressive for one side, too permissive for the other.
Then there’s the reputational risk. Meta has spent years trying to recover from the fallout of the 2016 election interference scandal, the Cambridge Analytica data breach, and the 2021 Facebook Papers — internal documents leaked by whistleblower Frances Haugen that showed the company was aware its platforms caused harm and, in some cases, chose growth over safety. A high-profile moderation failure — a mass shooting livestreamed without detection, a viral misinformation campaign that influences an election, a surge in child exploitation material that slips past AI filters — could reignite those controversies and draw the kind of political and public backlash that no amount of AI efficiency can offset.
Meta knows this. The company’s own researchers have published papers acknowledging the limitations of automated content moderation, particularly in what they call “adversarial” contexts where bad actors deliberately craft content to evade detection. Terrorist organizations, state-sponsored disinformation operations, and child exploitation networks have all demonstrated sophisticated understanding of how platform moderation works and how to circumvent it. These actors adapt faster than models can be retrained. Humans, for all their limitations, can sometimes recognize novel threats that AI systems miss entirely.
And yet the economics are relentless. Meta isn’t alone in cutting costs through automation. Across the tech industry, the pattern is the same: invest heavily in AI, reduce headcount, expand margins, and tell shareholders the machines are better anyway. Sometimes they are. Sometimes they aren’t. The difference, in the case of content moderation, is that the failures don’t just show up on a balance sheet. They show up in people’s lives.
For the contract workers being displaced, the impact is immediate and concrete. Many of Meta’s third-party moderators in countries like the Philippines and Kenya have few comparable employment options. The outsourcing firms that employ them — Accenture, Teleperformance, Majorel (now part of Teleperformance after a 2023 merger), and others — have already begun shifting their own business strategies toward AI-assisted services, which require fewer workers. Reuters reported earlier this year that several major outsourcing firms were reducing their content moderation workforces in anticipation of declining demand from big tech clients.
The workers themselves have had little say in the matter. Efforts to organize content moderators — including a high-profile campaign by workers at an Accenture facility in Kenya who partnered with the Foxglove legal nonprofit — have produced some concessions but no fundamental shift in power dynamics. The workers remain contractors, not employees. They have limited legal protections, limited bargaining power, and limited recourse when their jobs are automated away.
There’s a bitter irony here. The same AI systems that Meta built partly on the labor of these workers — who generated training data by labeling millions of pieces of content as violating or not violating — are now replacing them. It’s a pattern as old as industrialization, accelerated to internet speed.
So where does this leave Meta? In the near term, probably in a stronger financial position. The cost savings from reduced moderation staffing will flow to the bottom line, and the AI systems will handle the vast majority of routine content decisions competently. Zuckerberg will point to metrics showing faster response times, higher proactive detection rates, and fewer user complaints about over-enforcement. Investors will approve.
In the medium term, the picture is less certain. The first major AI moderation failure — and there will be one, because there always is — will test whether Meta has maintained enough human capacity to respond effectively. It will test whether the company’s AI systems can adapt quickly to novel threats. And it will test whether regulators, particularly in Europe, view Meta’s staffing decisions as adequate under the law.
In the long term, Meta’s bet is that AI will get good enough to make these concerns moot. That the models will learn context, understand nuance, and make judgments as well as or better than human reviewers. That’s possible. It’s also possible that the hardest 5% of moderation decisions — the ones involving satire, political speech, cultural context, and genuine ambiguity — will remain stubbornly resistant to automation for years or decades to come.
Meta is building the plane while flying it. The passengers — all 3.3 billion of them — don’t get a choice about whether to board.
Dave Ritchie is a technology writer based in the Midwest. He’s been covering the intersection of AI, platform governance, and corporate strategy since before it was fashionable. He also has two dogs who are better at detecting bad content than most algorithms.
WebProNews is an iEntry Publication