On May 31, 2025, NPR published a report built on internal Meta documents, bylined by Bobby Allyn and Shannon Bond: the company planned to hand most of its product risk reviews to AI, with up to 90% of privacy and integrity risk assessments across its apps automated rather than reviewed case by case by human teams. TechCrunch covered the change the same day.
For years, Meta ran what it calls privacy and integrity reviews before launching new features for Instagram, WhatsApp, and Facebook. Could a change violate user privacy, harm minors, or accelerate the spread of misleading or toxic content? Those reviews were conducted almost entirely by human evaluators. They are being replaced by a combination of questionnaires and AI-generated instant decisions. For teams shipping updates daily, that means speed. For anyone counting on review as a safeguard, it is a gamble.
What the internal documents revealed
According to the documents NPR obtained, as much as 90% of risk assessments at Meta would soon be automated. The affected work includes critical updates to Meta’s algorithms, new safety features, and changes to how content can be shared across the company’s platforms. These were the kinds of changes that human teams once debated in detail, weighing how a switch might be misused or produce unforeseen consequences. NPR’s conclusion was blunt: sensitive changes such as algorithm updates and safety adjustments would no longer face careful scrutiny from dedicated staff, and would instead be approved by an AI-powered system.
How the questionnaire and instant decisions work
As TechCrunch summarized the new process, product teams fill out a questionnaire before shipping an update, and the AI system usually returns an “instant decision.” The system flags risks it identifies and attaches requirements that must be met before launch. The judgment about whether a product change carries privacy or societal risk therefore shifts from human deliberation to the team’s own answers plus the model’s reading of them. Releases can move much faster, but the whole mechanism rests on two links: how the questionnaire is designed, and whether the model reads the risks behind the answers correctly. If either fails, the system will efficiently approve problematic changes at scale. The questionnaire effectively becomes the audit trail.
Employee and ex-executive concerns
Current and former Meta employees who spoke to NPR worry that automation puts AI in charge of the hardest judgments: anticipating how a platform change could cause real-world harm, or how it might be misused, something humans previously worked through at length. One former executive said that insofar as the process means more things launching faster with less rigorous scrutiny and opposition, the risk is higher. Inside Meta, the change has been received as a win for product developers, who can now release updates and features more quickly without waiting in a review queue. For those watching the safeguard function, that quiet weakening is exactly the problem.
Regulatory backdrop and Meta’s response
The human review process did not exist only by choice. As both NPR and TechCrunch noted, Meta’s 2012 consent agreement with the US Federal Trade Commission requires privacy reviews, which makes this shift especially sensitive. In response, Meta said it has invested over $8 billion in its privacy program and that novel or complex issues will still be handled by people. The company’s position, in short, is that AI handles routine cases and humans handle exceptions. Critics ask who decides where that line sits, and the uncomfortable answer is the automated system of questionnaires and models itself.
What it means for product teams
For anyone building on large platforms, the story carried two signals worth keeping. First, platform gatekeeping was moving into automated pipelines: release cycles get faster, but the logic behind approvals gets less transparent, and when something goes wrong it becomes harder to trace which safeguard failed. Second, the quality of risk review increasingly depends on upstream artifacts such as questionnaire design and model judgment, which replace human deliberation. Looking back from 2026, Meta’s move stands as the landmark case of AI being inserted into a company’s internal governance process, and it remains a standard reference when deciding how much review your own product launches can afford to automate. Teams that depend on platform distribution should also note how quickly a safeguard role can shift from people to pipelines.
Sources
- NPR: Meta plans to replace humans with AI to assess risks
- TechCrunch: Meta plans to automate many of its product risk assessments
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.