Microsoft’s Blueprint for Proving What’s Real in an AI-Generated World

When White House officials shared a manipulated image of a protester in Minnesota earlier this year, the deception was brazen enough to spot. But the incident revealed something more troubling: AI-enabled deception has permeated our online lives so thoroughly that even official channels are participating. Other times, the manipulation slips quietly into social media feeds, racking up millions of views before anyone questions its authenticity.

“We’re also trying to be a selected, desired provider to people who want to know what’s going on in the world.” — Eric Horvitz, Chief Scientific Officer, Microsoft

The Provenance Problem

Into this mess, Microsoft has put forward a comprehensive blueprint for how to prove what’s real online. An AI safety research team at the company recently evaluated how existing methods for documenting digital manipulation are faring against today’s most worrying AI developments—interactive deepfakes, widely accessible hyperrealistic models, and the seamless combination of video and voice.

To understand the gold standard Microsoft is pushing, imagine authenticating a Rembrandt painting. You might describe its provenance with a detailed manifest of where it came from and every time it changed hands. You might apply a watermark invisible to humans but readable by machines. And you could generate a mathematical fingerprint based on brush strokes. A skeptical museum visitor could then examine these proofs to verify authenticity.

All these methods are already being used to varying degrees online. Microsoft’s team evaluated 60 different combinations, modeling how each setup would hold up under various failure scenarios—from metadata being stripped to content being slightly altered or deliberately manipulated.

Three Layers of Verification

Provenance tracking creates an immutable record of a piece of content’s origin and journey across platforms. Every edit, every share, every transformation gets logged in a way that can’t be easily erased or forged.

Watermarking technology embeds invisible signals into AI-generated content. These marks persist even when content is screenshotted, compressed, or re-encoded—though they’re not indestructible against determined adversaries.

Cryptographic fingerprints generate unique mathematical signatures based on the actual pixel or audio data. Any alteration, no matter how subtle, changes the fingerprint—alerting platforms that something has been modified.

“I don’t think it solves the problem, but I think it takes a nice big chunk out of it.” — Hany Farid, UC Berkeley

The Implementation Gap

Despite publishing these recommendations, Microsoft has declined to commit to implementing them across its own platforms. The company sits at the center of a vast AI content ecosystem: it runs Copilot for image and text generation, operates Azure cloud services providing access to OpenAI and other major models, owns LinkedIn, and holds a significant stake in OpenAI itself.

When asked about in-house implementation, Chief Scientific Officer Eric Horvitz offered only that “product groups and leaders across the company were involved in this study to inform product road maps and infrastructure.”

There’s an important limitation to these tools: they reveal manipulation, not accuracy. A perfectly authentic video can still contain false information. A manipulated image might tell a true story. The technology answers “has this been altered?” not “is this true?”

The Engagement Problem

Even if platforms adopt these verification tools, a deeper challenge remains. Research shows that people are swayed by AI-generated content even when they know it’s false. In a recent study of pro-Russian AI videos about Ukraine, comments pointing out the AI origin received far less engagement than comments treating them as genuine.

Platforms like Meta and Google have committed to labeling AI-generated content, but an audit by Indicator last year found that only 30% of test posts on major platforms were correctly labeled. The incentive structure works against transparency: if labeling content as “AI-generated” reduces engagement, platforms face pressure to minimize such labels.

California’s AI Transparency Act, taking effect in August, will be the first major test of these tools in the United States. But enforcement could face challenges from the current administration’s posture against AI regulation and its general opposition to efforts curbing disinformation.

For now, Microsoft’s blueprint exists as a technical standard waiting for adoption. Whether it becomes the foundation for a more trustworthy internet—or remains a well-intentioned document that industry ignores—depends on choices that have little to do with the technology itself.

This article was reported by the ArtificialDaily editorial team. For more information, visit MIT Technology Review.