We cannot tell fake from real, and our checkers cannot either
Detectors score near random chance on new AI video, people spot synthetic clips barely better than a coin flip, and medical AI verifiers approve most wrong answers. One narrow detection method works. The rest of the checking stack is behind.
- People detect AI-generated videos with 52.6% accuracy, and state-of-the-art deepfake detectors perform close to random chance on the DF26 benchmark, in a preprint. Self-reported m:…6480
- Vanilla verifiers in medical visual question answering accept 71.0% of incorrect candidates on average, in a preprint. Self-reported m:…6488
- SnapJudge detects AI-generated math-heavy scientific text at 92% on proofs and 93% on prose at a 0.1% false positive rate, in a preprint. Self-reported m:…6478
- METR finds AI judges systematically overestimate the quality of tasks made by models stronger than themselves. m:…4775
The detection failure
Start with video. A new benchmark called DF26 tests detectors against modern AI-generated video, and the paper reports that state-of-the-art deepfake detectors perform close to random chance on it. People do not do much better: human accuracy in detecting AI-generated videos is 52.6%, barely above guessing, while human accuracy in recognizing real videos is 76.0%. The gap between those two numbers is the story. Real footage still reads as real, but synthetic footage now passes as real most of the time, to people and to machines alike. m:…6480
That matters because detection is the backstop for everything else. Newsrooms, platforms and courts all assume that a fake can be caught, either by a person looking closely or by a tool. DF26 suggests that assumption no longer holds for video, at least not with current detectors. The benchmark exists precisely because the old ones stopped discriminating: detectors tuned on older generators fail on the new ones, and the failure is not marginal. Near random chance means the tool contributes almost nothing. m:…6480
The verification failure
The second failure is quieter and arguably more dangerous, because it happens inside systems that are supposed to check themselves. A paper on self-verification in medical visual question answering finds that vanilla verifiers accept 71.0% of incorrect candidates on average. In other words, when a medical vision-language model produces a wrong answer and then checks its own work, it usually signs off on the mistake. The authors introduce a framework called MirageCheck, which reduces mean false acceptance by 14.6 percentage points relative to the strongest candidate-independent baseline. m:…6488
The fix the authors propose is instructive: reliable visual self-verification requires turning acceptance to discriminative evidence from the current image rather than another judgment of the same problem. A verifier that re-solves the problem the same way the model did will repeat the error. A verifier that is forced to point at evidence in the image has something independent to check against. That is a design principle with uses well beyond medical imaging: any system that grades its own homework needs an independent source of evidence, or the grade is meaningless. m:…6488
The judge problem
The same pattern shows up at the level of the whole field. METR's Q4 2026 update on AI research capabilities reports that AI judges systematically overestimate the quality of tasks made by models stronger than themselves, which the authors read as a potential cap on self-improvement: a weaker judge cannot reliably grade a stronger student's work. METR also reports that time horizons for AI agents are doubling roughly every 3.6 months, and that models usually match or outperform skilled humans on coding and optimization tasks, while showing worse judgment than human experts on tasks that require choices without clear empirical feedback. m:…4775
Put those findings together and the picture is uncomfortable. Models are getting better at the tasks we can grade, and worse at being graded on the tasks we cannot. Current public evidence is insufficient to rule out drastic AI research acceleration in the next couple of years, according to METR, which means the checking problem is likely to get harder while the things being checked get more capable. The bottleneck is not generation. It is evaluation. m:…4775
One method that works, narrowly
There is a counter-example this week, and it is worth noting because it shows what good detection looks like. SnapJudge tackles AI-generated scientific text heavy with mathematics, where text-based detectors suffer significant accuracy drops. The authors found that LaTeX-based inputs improve detection accuracy compared with extracted text by preserving formatting, and that their image-based approach achieves 92% detection on proofs and 93% on prose at a 0.1% false positive rate. The method works because it stops throwing away the information that matters: the formatting of the math itself. m:…6478
The lesson across all four papers is that checking works when the checker has access to something the generator did not control: the formatting of a proof, the pixels of an image, the evidence in front of it. It fails when the checker just repeats the same reasoning and asks itself whether it looks right. For anyone building or buying AI systems, that is a practical test to apply: ask what independent evidence the verification step uses. If the answer is nothing, the 71% figure is the number to remember. m:…6488 m:…6478
What is still unresolved
- Whether MirageCheck's gains hold outside medical visual question answering.
- Whether DF26's detector results reflect the newest commercial detectors or only published research systems.
- Whether SnapJudge's image-based detection generalizes beyond math-heavy scientific text.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.