Artificial intelligence enabled deception is increasingly embedded in everyday online life, from manipulated protest photos circulated by US officials to covert Russian influence videos aimed at deterring Ukrainian enlistment. In response, Microsoft has developed a blueprint for proving what is real online, grounded in methods analogous to authenticating a Rembrandt painting: detailed provenance records, invisible watermarks readable by machines, and mathematical fingerprints derived from the content itself. An artificial intelligence safety research team at the company evaluated how these techniques perform against emerging threats such as interactive deepfakes and hyperrealistic, widely accessible models, with the goal of defining technical standards that artificial intelligence providers and social media platforms can adopt.
Microsoft’s researchers tested 60 different combinations of provenance, watermarking, and fingerprinting, modeling how each setup behaves when metadata is stripped, content is subtly altered, or media is deliberately manipulated. The team then mapped which combinations remain robust enough for platforms to confidently surface to users and which are too fragile and risk adding confusion. Chief scientific officer Eric Horvitz frames the push as a form of self regulation motivated by new laws such as California’s AI Transparency Act, which will take effect in August, and by the rapid advances in artificial intelligence systems that can tightly synchronize video and voice. Horvitz maintains that the proposed tools are explicitly not designed to determine factual accuracy, but instead to label where content came from and whether it has been manipulated, a distinction he emphasizes to lawmakers wary of letting large technology firms arbitrate truth.
External experts such as UC Berkeley digital forensics professor Hany Farid say widespread adoption of Microsoft’s blueprint would make it meaningfully harder to mislead the public with manipulated media, even though sophisticated actors could still circumvent protections and human psychology often overrides warnings about artificial intelligence content. Existing efforts, including Google’s watermarking of artificial intelligence outputs since 2023 and partial uptake of the C2PA provenance standard launched in 2021, show some progress, but an audit by Indicator found that only 30% of its test posts on Instagram, LinkedIn, Pinterest, TikTok, and YouTube were correctly labeled as AI-generated. Microsoft is pushing to shape upcoming regulation such as the European Union’s AI Act and California’s law while warning that weak or error-prone verification tools could backfire, erode trust, and even enable new sociotechnical attacks if real images are minimally altered to trigger misleading labels. The company’s approach will be tested as California’s AI Transparency Act collides with a presidential executive order instructing federal agencies to restrain state rules deemed burdensome, amid an administration that has both opposed some disinformation curbs and itself shared artificial intelligence generated content through official channels.
