Microsoft proposes blueprint to verify online content and detect AI manipulation

Microsoft proposes a blueprint for verifying online content

Microsoft has published a blueprint for how to authenticate digital content and detect AI manipulation. The plan, shared with MIT Technology Review, comes as AI-generated deepfakes and disinformation become pervasive online.

The company's AI safety research team evaluated 60 different combinations of existing verification methods. It then recommended technical standards for AI companies and social media platforms to adopt.

The three pillars of Microsoft's verification system

Microsoft's framework relies on three core techniques working in concert. The first is provenance, which is a detailed record of a piece of content's origin and edit history.

The second is a machine-readable watermark embedded directly into the content. The third is a unique digital fingerprint based on the content's mathematical characteristics.

"You might call this self-regulation," Microsoft's chief scientific officer, Eric Horvitz, told MIT Technology Review. He said the work was prompted by new laws and the rapid advancement of hyperrealistic AI.

Microsoft stops short of a full commitment

Despite creating the blueprint, Horvitz declined to commit to implementing it across Microsoft's own vast ecosystem. The company operates several major platforms that generate or host AI content.

• It runs the AI assistant Copilot, which can generate images and text.
• It operates Azure, the cloud service providing access to OpenAI and other models.
• It owns the professional network LinkedIn.
• It holds a significant stake in OpenAI.

When asked about in-house use, Horvitz stated that product teams were informed by the study and engineering teams "are taking action on the report’s findings."

The inherent limits of technical verification

Experts note these tools have clear boundaries. They can only signal if content has been manipulated, not whether it is accurate or true.

"It’s not about making any decisions about what’s true and not true," Horvitz said. "It’s about coming up with labels that just tell folks where stuff came from."

Hany Farid, a UC Berkeley digital forensics professor, says widespread adoption would make public deception harder. "I don’t think it solves the problem, but I think it takes a nice big chunk out of it," he said.

Why platforms might resist clear labeling

There is skepticism that tech platforms will fully embrace such systems. Their business models often prioritize engagement, which clear "AI-generated" labels could reduce.

"If the Mark Zuckerbergs and the Elon Musks of the world think that putting ‘AI generated’ labels on something will reduce engagement, then of course they’re incentivized not to do it," Farid says.

Evidence supports this concern. An audit last year found that only 30% of test AI posts on Instagram, LinkedIn, Pinterest, TikTok, and YouTube were correctly labeled.

The regulatory push and potential pitfalls

External regulation may force the issue. Laws like the EU's AI Act and California's AI Transparency Act, which takes effect in August, will require disclosure of AI-generated content.

Microsoft has actively lobbied on such rules. Horvitz said its efforts made California's requirements "a bit more realistic."

The researchers warn that a poorly executed rollout could backfire. If labeling systems are inconsistent or often wrong, the public could lose trust in them entirely.

The political landscape complicates enforcement

Enforcement faces political headwinds. President Trump issued an executive order in late 2025 seeking to curtail state AI regulations deemed "burdensome" to industry.

His administration has also opposed disinformation-curbing efforts, canceling related research grants last year. Government channels, including the Department of Homeland Security, have used AI video generators for public content.

When asked if fake content from official sources worried him, Horvitz said, "Governments have not been outside the sectors that have been behind various kinds of manipulative disinformation, and this is worldwide."