Evidence, not verdicts.
In 2026, AI video crossed a line. Google's Veo 3.1 and Kling 3.0 now render cinematic lighting, natural motion, fabric and water and hair that behave, even synced audio. Sora 2 produces clips most people cannot tell apart from a phone recording. The industry's own word for this moment is "production-ready." For anyone trying to tell what's real online, that's the whole problem in one sentence: the fakes now look exactly like the truth.
The instinct is to reach for a detector — an app that scans a clip and announces "94% fake" or "98% authentic." It feels like the answer. It isn't. Here's why, and what to do instead.
Detection vendors love a big number. The trouble is where that number comes from. Leading detectors score 90–99% on the data they were trained and tested on — and 50–65% on deepfakes they haven't seen before. In real-world conditions, measured accuracy drops 45–50% from the lab benchmark. That's a coin flip wearing a lab coat.
Three forces break detectors in the wild:
That last point is why a confident verdict is worse than no verdict. When a tool stamps "authentic, 98%" on a fake, it doesn't just miss — it launders the fake with false authority. Trust spent that way doesn't come back.
There is no magic pixel that says "AI." But there are independent signals, and weighed together the picture usually resolves. No single one is proof on its own — that's the point.
1. Provenance (Content Credentials / C2PA). The most durable answer is a cryptographic record of where a file came from and how it was edited. The open standard — backed by Adobe, the BBC, Microsoft and Intel — is no longer theoretical: Google's Pixel 10 signs every photo by default with hardware-backed keys; Leica, Nikon, Canon, Fujifilm and Samsung ship C2PA-capable cameras; and tools like DALL·E and Sora embed credentials that declare AI origin. The catch: social platforms strip this data on upload — so its absence proves nothing, while its presence is gold.
2. Where it appears. Run the image across the web. Does it trace back to a news wire or a documented event — or surface first on AI-image sites, or only on anonymous social accounts? A genuine news photo leaves a credible trail; a fabricated one usually doesn't.
3. The fact-check record. Has this exact clip already been examined? And read carefully what was rated false: a "false" verdict most often means miscaptioning — real footage paired with a false story — not that nothing happened. Real event, wrong video is the most common form of visual misinformation, and the one a blunt "FAKE" label gets wrong.
4. The forensic read. An AI model's close look — malformed hands, melted text, impossible physics, light that doesn't agree. Useful as one signal; fallible as a verdict. And never forget the frame problem: a clean still cannot clear the whole video.
Weigh those four together and you get the thing a detector cannot give you: an honest read, with its reasons exposed, that you can check for yourself.
This is the entire reason Relity exists, and why our rule is "evidence, not a verdict." We don't hand you a number and tell you to trust it. We show you the provenance, where the image appears, the fact-check record, and an AI vision read — each labeled with its caveats — converging into one honest best guess. When the signals conflict, we say so. When we genuinely can't tell, we say that, too. And when fact-checkers have flagged a real clip as miscaptioned, we tell you the event may well be real and the footage is what's being misused — not that "the news is fake."
The result is the opposite of a black-box detector: you see what we found, you see what we couldn't, and you make the call. In a world where the fakes are perfect, that humility isn't a weakness. It's the only honest position left.
The future of telling real from fake isn't a smarter detector. It's provenance you can verify, a plurality of signals you can weigh, and a human — you — making the final call. Anyone selling you certainty is selling the one thing that's hardest to come by.