On June 15, 2025, two days into the twelve-day Israel-Iran war, an AI-generated video appeared on Facebook showing damaged buildings in Haifa, Israel. The account that posted it claimed to be a news outlet. It was actually operated from the Philippines. The video was completely fabricated. It accumulated over 700,000 views.
Six users reported the content to Meta. No company review occurred. No third-party fact-checker ever saw it. The video stayed up, racking up views during an active military conflict where real people were deciding whether to evacuate based on what they saw on social media.
Here's the detail buried in the Oversight Board's ruling that no one is talking about: the Facebook page posting this fabricated wartime footage was eligible for monetization through Meta's Stars program.
Meta was paying the creator. Then it ignored reports about the content. Then its own Oversight Board had to intervene to get the video labeled.
Today, the Oversight Board issued 10 recommendations demanding Meta overhaul its approach to AI-generated content. The recommendations include creating a new Community Standard for AI content, investing in better detection tools, and implementing Content Credentials at scale. All reasonable. All treating the symptom while ignoring the disease.
The Monetization Paradox
The fundamental problem isn't that Meta lacks AI detection tools or content policies. It's that Meta's business model creates a financial incentive to produce exactly the kind of content it claims to be fighting.
Engagement-driven platforms reward content that generates views, shares, and emotional responses. Fabricated wartime footage showing buildings being destroyed is, by design, exactly the kind of content that performs well in this system. A creator operating from the Philippines can generate an AI video of a war zone, post it as "breaking news," accumulate hundreds of thousands of views, and get paid through Meta's own monetization programs.
This is the same perverse incentive structure I wrote about with ChatGPT's ad-supported future: when the platform's revenue depends on engagement, every policy designed to remove engaging content works against the business model. You can't fix this with 10 recommendations. You fix it by changing what gets rewarded.
The Oversight Board's recommendations ask Meta to invest in detection, improve labeling, and create new standards. None of them address the fact that Meta's Stars program, its ad revenue sharing, and its algorithmic amplification all reward the behavior the recommendations are trying to prevent.
Detection Is Losing the Arms Race
According to Rest of World's reporting, Meta admitted to the Oversight Board that it can only automatically detect AI-generated content in static images. Not video. Not audio. The most dangerous deepfakes, the ones most likely to influence public opinion during a conflict, exist in the formats Meta can't detect at all.
This shouldn't surprise anyone who's followed the detection arms race playing out across AI platforms. The pattern is consistent: generative AI quality improves faster than detection capabilities. Creation is cheap and getting cheaper. Detection is expensive and perpetually behind. Every dollar Meta spends on detection buys less ground than the equivalent dollar spent on generation tools that are freely available to anyone.
The same dynamic is already playing out with voice cloning, where detection accuracy has dropped to near coin-flip levels. Microsoft published a report finding that current media authentication tools "aren't ready" for the AI content flood. The C2PA Content Credentials standard, which the Oversight Board specifically recommends Meta adopt, is facing systemic criticism because platforms, including Meta, strip metadata on upload.
Read that again: the Oversight Board is recommending that Meta implement a content provenance standard that Meta's own upload pipeline destroys.
The Self-Disclosure Absurdity
The Oversight Board's ruling includes a line that deserves more attention than it's getting. The board described Meta's current approach as "overly dependent on self-disclosure of AI usage" and stated it "cannot meet the challenges posed in the current environment."
This is diplomatic language for something blunter: Meta's primary mechanism for identifying AI-generated content is asking creators to voluntarily label it. The system depends on the good faith of the people most incentivized to lie.
In enterprise security, we have a name for controls that depend on adversary cooperation: theater. No security professional would design an authentication system that asks attackers to self-identify. No fraud detection system works by asking fraudsters to check a box. Yet this is essentially what Meta built for AI content moderation, and what it operated with through the first industrial-scale AI disinformation campaign in history.
The Iran-Israel conflict wasn't the first war with online disinformation. Previous conflicts in Armenia, Ukraine, and Gaza featured recycled images, fake livestreams, and game footage passed off as real combat. But researchers at Witness described the Iran-Israel conflict as something qualitatively different: "AI-generated content related to the Iran-Israel conflict has taken disinformation to an industrial level." Both state actors used deepfakes systematically, and smaller creators used AI tools to fabricate and monetize fake war imagery.
Meta's self-disclosure model was built for a world where AI-generated content was rare and easy to spot. That world is gone.
The Liar's Dividend
There's a second-order effect of the deepfake flood that the Oversight Board references but that no coverage has explored deeply: the "liar's dividend."
When deepfakes become common enough, people stop trusting authentic content too. During the Iran-Israel conflict, both governments exploited this in both directions. They created fake content to spread disinformation, and they claimed authentic footage was fake to discredit real reporting. The prevalence of deepfakes gives bad actors a new tool: plausible deniability for anything caught on camera.
This is the deeper damage that engagement metrics can't capture. Every fabricated video that goes viral on Meta's platform doesn't just misinform the people who see it. It degrades the information ecosystem for everyone, making it harder to trust any content, even content that's real. The same pattern is emerging in AI recommendation systems, where the fix isn't technical but structural.
For a platform that positions itself as a place people go for news and information, this erosion of trust is an existential problem. But it's an existential problem that doesn't show up in quarterly engagement numbers. If anything, the confusion and conflict generated by deepfakes drives more engagement, more time on platform, more ad impressions. The liar's dividend is, perversely, also Meta's dividend.
The Enforcement Gap
The Oversight Board has given Meta 60 days to respond to its 10 recommendations. This is the part where the governance structure reveals its limitations.
The Oversight Board has no enforcement power. It can overturn individual content decisions and issue recommendations. It cannot compel Meta to change its business practices, restructure its monetization programs, or invest in specific technologies. Previous recommendations have been partially implemented or ignored entirely. The board itself acknowledged that Meta "is less responsive to outreach and concerns, in part due to a significant reduction in capacities for Meta's internal teams."
This tracks with the broader enforcement paradox in tech regulation: impressive-sounding regulatory actions that produce headlines but fail to change behavior. Meta's own €1.2 billion GDPR fine remains in appeals years later. The company has explicitly refused to sign the EU AI Code of Practice. When the cost of compliance exceeds the cost of non-compliance, rational actors choose non-compliance.
The Oversight Board's recommendations are well-crafted and technically sound. Creating a separate Community Standard for AI content makes sense. Implementing Content Credentials at scale would help. Better detection tools are needed. But none of it matters if the entity responsible for implementation has no incentive to implement it, no consequence for ignoring it, and a business model that benefits from the status quo.
What Would Actually Work
The AI safety implementation gap, the chasm between stated commitments and operational reality, is wider at Meta than almost anywhere else in tech. Closing it requires something the Oversight Board can't mandate: changing the incentive structure.
Three things would actually move the needle:
First, decouple monetization from virality for unverified content. Pages and accounts that haven't been verified as legitimate news sources should not be eligible for monetization on content related to active conflicts. This is not a technical challenge. It's a business decision Meta chooses not to make because it would reduce revenue.
Second, invert the burden of detection. Instead of trying to detect AI-generated content after upload (a losing game), require provenance verification before content can be amplified by recommendation algorithms. Unverified content can still be posted, but it doesn't get algorithmic distribution. This creates a natural throttle on the reach of fabricated content without requiring perfect detection.
Third, make the Oversight Board's recommendations binding, or stop pretending it's a governance mechanism. An advisory body with no enforcement power is a PR strategy, not a governance structure. If Meta is serious about AI content moderation, it needs a mechanism with actual teeth.
None of these are technically impossible. They're commercially inconvenient. And that's the gap the Oversight Board's 10 recommendations can't close: the distance between what Meta could do and what it will do when doing the right thing costs money.
The next industrial-scale AI disinformation campaign won't wait 60 days for Meta's response.
