It’s not every day a tech giant takes automation to the point of uncanny recursion—Meta is handing over most of its risk assessments to artificial intelligence, letting the algorithmic fox safeguard the privacy henhouse. As NPR reports, up to 90% of internal reviews assessing privacy and societal risks on Instagram, Facebook, and WhatsApp will soon be AI-driven. In effect, Meta will ask AI to judge whether other AI, and new product changes, are unlikely to set the platform ablaze—figuratively, if not literally.
“Move Fast and Audit Later”
Historically, as described by NPR, these risk assessments—meant to prevent misfires around privacy or the dystopian spread of misinformation—were handled by actual people tasked with untangling the potential fallout of Meta’s updates. That human buffer is mostly going away. The plan, based on internal documents reviewed by NPR, is for AI systems to review developer questionnaires and deliver “instant decisions,” with only a handful of situations escalating to manual review by staff.
Meta frames this as a streamlining win for its engineers and product teams. These groups can now release updates with what amounts to algorithmic pre-approval, no longer forced to wait for human sign-off. According to NPR’s interviews with current and former Meta employees, there’s concern that this trade-off—speed for scrutiny—could mean higher risks and more surprises, the kind that usually wear headlines like “unexpected data leak” or “viral hoax surges across platform.” One former executive put it candidly: “Negative externalities of product changes are less likely to be prevented before they start causing problems in the world.”
Meta, in statements cited by NPR, insists that only “low-risk decisions” are being automated and that “human expertise” will still be brought in for “novel and complex issues.” However, internal records indicate that potentially sensitive areas—AI safety, youth risk, issues around violent content or lies—are all under consideration for automation. So where’s the line between “complex” and “routinely risky,” and who gets to decide?
Guardrails, Now With Fewer Guards
Under the old regime, as outlined by NPR, no feature for Meta’s billions of users launched without a battery of reviewers poring over risk factors. With the new process, the same product engineers who create these features are expected to use their best judgment, unless they themselves request a deeper review. Zvika Krieger, Meta’s former director of responsible innovation, expressed skepticism to NPR: most product managers and engineers “are not privacy experts,” and are measured by how quickly they launch products, not how well they anticipate pitfalls. He also observed that past attempts at self-assessment have turned into box-checking exercises that overlooked major hazards.
To Meta’s credit, the company is conducting some audits of the automated system’s decisions after launch. Yet, as skeptics inside and outside the company point out in NPR’s reporting, fixing a mistake once it’s already gone live is less reassuring than preventing it in the first place. It’s a bit like setting up your security camera after the raccoons have already made off with the marshmallows.
Europe Gets the Human Touch (For Now)
NPR also notes that users in the European Union may be somewhat shielded from these changes, at least initially. Thanks to robust regulation under the Digital Services Act, EU content and privacy reviews will continue to receive oversight from Meta’s European headquarters in Ireland, rather than being left solely to automated checks. The outlet points out that this isn’t Meta’s first regulatory carve-out for Europe, reflecting the continent’s more aggressive digital safeguards.
The shift stateside comes as Meta winds down its global fact-checking program and relaxes some hate speech policies, as recounted in NPR’s review of internal documents. Notably, these procedural overhauls follow CEO Mark Zuckerberg’s well-documented attempts to build bridges with President Trump in the wake of what Zuckerberg reportedly described as a “cultural tipping point.” NPR’s reporting links these moves to a broader theme: Meta is rebalancing toward speed, agility, and more unmediated speech, with less friction from oversight.
One can’t help but ask: is stricter oversight simply too slow for the pace of modern platform life—or is this just another way for the train to leave the station before anyone checks if the brakes work?
If AI Judges AI, Who’s Left Watching?
What’s the rationale for all this? According to Meta’s latest quarterly integrity report, cited by NPR, the company claims its large language models now outpace human performance in certain policy areas, especially for content that’s “highly confident” not to break the rules. The argument is that automation frees up people to focus on the toughest calls. Even so, Katie Harbath—a former Meta policy lead now running a tech policy firm—told NPR that, while automating routine decisions can cut down on duplicate effort, “those systems also need to have checks and balances from humans.” Otherwise, she notes, it’s easy to miss what’s right in plain sight.
Former and current employees voice similar apprehensions. One described the automation push as “self-defeating,” noting to NPR that high-profile product launches often attract public scrutiny, where issues emerge that internal checks might have glossed over. Michel Protti, Meta’s chief privacy officer for product, reportedly characterized the shift as “empowering product teams” by simplifying risk review and decision-making in about 90% of cases. But another insider, closely tied to the risk review process, flatly called it “fairly irresponsible given the intention of why we exist. We provide the human perspective of how things can go wrong,” as quoted by NPR.
The overarching theme here seems less about evil intent and more about a persistent belief in the sufficiency, or at least expediency, of automation—especially when competitor apps and features are just a click away.
The Ouroboros of Oversight
So what does it all add up to? NPR’s reporting paints a picture of an organization rapidly replacing human judgment with automated decisions—AI evaluating the risks generated by other AI and product updates. Humans are now the emergency backstop, consulted mainly if something seems exceptionally thorny (or if the teams themselves ask for it). All told, it’s an efficiency dream—or maybe an oversight ouroboros eating its own tail, hoping not to miss a scale.
It’s difficult to ignore the irony. The same obsessive pursuit of speed and scale that made social media ubiquitous now demands we trust automated systems to spot the problems of, well, automated systems. Are we reaching the promised land of frictionless progress, or simply setting ourselves up for a very efficient kind of trouble? Would anyone even notice if something critical slipped through the (machine-made) cracks, until it’s already front-page news?
Or maybe, in the end, the weirdest part isn’t that we’re trusting the machine. It’s how much easier it is to believe it’ll get things right—especially when believing otherwise would slow things down.
