Authenticity Becomes Infrastructure
Infrastructure is the set of things you only notice when they fail. The road that carries you to work, the current behind the wall, the clean water at the tap: each is invisible precisely because it is dependable. We build our lives on top of these systems without thinking about them, and that quiet reliability is what makes everything else possible. For most of modern history, one such system went entirely unnamed, because no one imagined it could break. That system was the basic correspondence between a recorded image and reality.
For a century, a photograph or a video carried an implicit guarantee. It was not perfect, and it could be staged, but at its core it recorded light that had genuinely fallen on a real scene. That guarantee was load-bearing. Journalism leaned on it. Courts leaned on it. So did insurance claims, family memories, and the simple act of believing a stranger's account of an event you did not witness. We never called it infrastructure because we never had to maintain it. Generative AI has changed that overnight, and the maintenance bill has come due.
The scale is not subtle. Detected deepfake incidents surged tenfold between 2022 and 2023, and the financial stakes are climbing with them. Deloitte projects that generative-AI-enabled fraud losses in the United States could rise from $12.3 billion in 2023 to roughly $40 billion by 2027. When the cost of misplaced trust runs into the tens of billions, proof of realness stops being a philosophical nicety. It becomes a thing that has to be engineered, funded, and kept running, like any other utility a society depends on.
When anything can be faked, proof of realness stops being a feature and becomes load-bearing.
The encouraging news is that serious people are building this infrastructure in the open. The Content Authenticity Initiative has grown to more than 3,700 members, with OpenAI joining the C2PA steering committee in 2024. The idea behind these provenance standards is elegant: attach a tamper-evident record to a piece of media at the moment of capture or creation, so that its origin and edit history travel with it. On a parallel track, Google DeepMind extended its SynthID watermarking to video in 2024, embedding signals into AI-generated output that machines can later detect. Provenance you can attach, and watermarks you can find: two complementary attempts to give realness a paper trail.
It would be dishonest to present these systems as a solved problem. Their greatest weakness is the same thing that makes the internet work: re-encoding. Every time a video is uploaded, compressed, clipped, or reposted, the platforms processing it can quietly strip the provenance metadata or degrade an embedded watermark. A Content Credential that survives a controlled lab does not always survive a screen recording shared to a group chat. Adoption is also voluntary and uneven, which means the absence of a credential tells you almost nothing. The honest fake will simply carry no provenance at all, indistinguishable, on that axis, from an honest video that was never labeled.
This is why the future of authenticity cannot rest on a single oracle that stamps each file true or false. The infrastructure has to be layered, the way real infrastructure always is. Roads have lanes, signs, lights, and shoulders, and none of them alone keeps you safe. Media authenticity will work the same way: provenance metadata where it survives, watermark detection where it is present, and, crucially, the wider context that no watermark can carry. Who published this? What does the description disclose? What are independent viewers saying about it? These public signals do not require the cooperation of whoever made the fake, which is exactly why they remain useful when provenance has been stripped away.
None of these layers delivers certainty, and any system that claims it does should be distrusted on principle. What a layered approach offers instead is something more durable: a reasoned estimate built from converging signals, presented honestly as a likelihood rather than a verdict. That is a less satisfying promise than a green checkmark, but it is the only one that survives contact with reality. We are trading the fantasy of perfect detection for the discipline of reading evidence well, and over time that discipline is what keeps the infrastructure trustworthy.
The work ahead is to make those signals legible to ordinary people, not just to forensic labs and platform engineers. An audience that can read provenance, weigh source reputation, and notice what the crowd is flagging is an audience that can keep functioning in an environment built to deceive it. That capacity, distributed widely, is the real utility we are building, the one that lets everything else, journalism, commerce, public trust, keep running.
In an AI world, authenticity becomes infrastructure. You can use a piece of it right now: analyze a video and read the signals for yourself.
- Deepfake incidents surged tenfold from 2022 to 2023, Sumsub. prnewswire.com
- GenAI fraud losses projected $12.3B (2023) to $40B (2027), Deloitte. deloitte.com
- Content Authenticity Initiative 3,700+ members; OpenAI joins C2PA, Adobe. blog.adobe.com
- SynthID expanded to video, Google DeepMind. deepmind.google