On January 22, 2026, Runway published “The Turing Reel,” a controlled study on whether humans can still tell AI video from real footage. The company recruited 1,043 participants through random sampling and had each of them watch 20 five-second clips — 10 real, 10 generated by its newest model, Gen-4.5 — judging every clip as real or fake. The headline numbers: overall accuracy of 57.1%, and only 99 people, or 9.5%, who performed significantly above chance. The name nods to the Turing test. The finding inverts it: on video, humans have quietly lost it.
This is the first time a video model vendor has run a rigorous controlled experiment with its own frontier model, reducing “can the human eye still detect AI video” to a single-digit percentage. For platform and content teams, it is not a marketing stunt — it is a data point that belongs in a risk assessment.
How the Test Worked
- Real footage was sampled from the Filmpac library across five categories: faces, full-body human motion, animals, nature scenes, and urban environments
- The first frame of each real video was extracted and fed to Gen-4.5’s newly released image-to-video capability, using default settings — every clip generated exactly once, with no regeneration or post-processing
- All clips were trimmed to 5 seconds and matched in resolution; participants could watch up to 10 seconds before judging
- The 20 clips appeared in randomized order, half real and half generated
The design deliberately mirrors social media viewing conditions: short, fast, judged in seconds. It tested not careful forensic inspection in a lab, but the real attention span of someone scrolling a feed.
57.1%: Barely Better Than a Coin Flip
An overall accuracy of 57.1% sits only slightly above the 50% random baseline. The more telling detail is the split: participants identified real videos correctly 58.0% of the time and generated videos 56.1% of the time. The two figures are nearly identical, which means nobody developed a stable detection strategy — there was no “tell” being found. Raise the bar to statistical significance (at least 15 of 20 correct, p < 0.05) and only 99 of 1,043 participants clear it. TechRadar’s coverage put it plainly: most people can no longer separate AI videos from reality.
Where Humans Still Had an Edge
The per-category breakdown shows one remaining strength. Content with people in it performed best — faces, hands, and human actions landed at 58–65% accuracy, because body structure and physical motion are still where generation models make their most visible mistakes. Animals and architecture, however, dropped to 45–47%, below random guessing. The “AI look” participants thought they were detecting was, for non-human content, a negative indicator. TechRadar also noted Gen-4.5’s known flaws: disappearing objects, causal errors (doors opening before the handle turns), and “success bias” (a badly aimed soccer kick still scoring). The problem is that none of those clues can be found in five seconds of casual viewing.
Even Runway’s Co-Founder Failed
Runway co-founder and CTO Anastasis Germanidis admitted he got it wrong “quite a bit” himself. His conclusion is worth taking seriously: at this point we are crossing the threshold where generated and real videos are hard to tell apart, and viewers need a more critical mindset about online content. When the CTO of the company that built the model says he cannot reliably spot its output, that carries more weight than any benchmark score.
From Detection to Provenance
The final section of the study is the part that matters most. If detection by eye is no longer viable, trust has to be established at the source instead. Runway already embeds C2PA provenance metadata in its outputs by default, is developing more reliable watermarking, and is calling on the industry to build shared verification standards; the research page also hosts a public “Real or AI” test anyone can take. In the days after release, commentators including synthetic-media specialist Henry Ajder focused on the sub-10% reliable-detection rate, reading it as quantified evidence that the era of eyeball detection is over. For developers, the practical takeaway is blunt: build the real-versus-fake line at the metadata and cryptographic-signature layer, and stop counting on users or moderators to see the difference.
Sources
- The Turing Reel — Runway Research
- Most people now can’t separate AI videos from reality — TechRadar
- Henry Ajder on The Turing Reel — LinkedIn
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.