Your brain is NOT good at ‘spotting’ deepfakes:
Why lab studies and feel‑good media summaries are at best misleading . . . ”
 |
| Clker.com |
You may have seen the latest "feel‑good" report: “Humans still beat AI at spotting deepfake videos” (University of Florida, 2026). It sounds hopeful; our messy, embodied brains outwitting the machines that are trying to fool us. Look at what the research actually did, though, and the comfort evaporates (Pehlivanoglu et al., 2026; Diel et al., 2024).
What the UF study really shows
The University of Florida team ran two experiments: one with still face images, one with short talking‑head videos (Pehlivanoglu et al., 2026). In both, the content was not random internet sludge; it was carefully curated, cleaned‑up, lab‑ready material. For images, they took “real” faces from the Flickr‑Faces‑HQ dataset and “deepfake” faces generated with StyleGAN2, then filtered out any GAN outputs with obvious glitches (Pehlivanoglu et al., 2026). Participants rated faces on a fake–real scale.
Results:
- Humans: near chance at telling real from fake (AUC ≈ 0.53, accuracy just under 50%), with a strong truth‑bias—better at identifying real than fake (Pehlivanoglu et al., 2026; Bray et al., 2023).
- A convolutional neural net: around 97% accuracy on the same images (Pehlivanoglu et al., 2026).
On still images, your conscious “deepfake detector” is basically a coin flip—often a very confident coin flip in the wrong direction (Bray et al., 2023; Diel et al., 2024).
For videos, they pulled real and fake clips from the Deepfake Detection Challenge (DFDC) dataset, again heavily filtered: good lighting, no text overlays, one person facing the camera, 10‑second clips, minimal motion, no audio tricks. Under those tidy conditions, humans managed about two‑thirds accuracy, while the particular model tested fell to near chance (Pehlivanoglu et al., 2026). Other DFDC work finds similar “barely above chance” human performance—often in the 55–60% range (Köbis et al., 2021; Korshunov & Marcel, 2020).
That narrow slice of data is what gets packaged as “Humans beat AI on deepfake videos” (University of Florida, 2026).
Here is where my skepticism spikes: ecological validity.
In both UF experiments, the researchers controlled the stimulus set: which GAN images survived their artifact filter, which DFDC videos passed their quality screen, how clips were trimmed (Pehlivanoglu et al., 2026). That is standard lab practice—but it also means we do not know how representative those stimuli are of the deepfakes that actually show up in your inbox, on your feed, or in a scam call.
Bray et al. (2023) followed a similar pattern with GAN‑generated faces matched to high‑quality photographs. Across thousands of online participants, accuracy hovered around 60–64%, with striking overconfidence (Bray et al., 2023). Diel et al.’s (2024) meta‑analysis, pooling 56 studies, concludes that human deepfake detection is often “as good as chance,” with accuracy barely above 50% and sometimes worse.
In the real world, deepfakes do not arrive as balanced sets of 100 real and 100 synthetic faces drawn from the same distribution. They ride inside political propaganda, fundraising pitches, and emotional “grandchild in trouble” messages, layered over compression, screenshots, and user‑generated chaos (Diel et al., 2024; Somoray et al., 2025). UF’s choice to exclude audio swaps, overlays, messy sound, and multiple speakers makes for clean statistics, but it is a long way from what scammers actually push across Zoom and WhatsApp (Köbis et al., 2021; Rossetto et al., 2023).
The “internet savvy” story: very thin ice
UF does acknowledge that not all humans are equal: better video performance correlates with higher analytical thinking, lower positive mood, and greater “internet skills” (Pehlivanoglu et al., 2026). Other work likewise finds that higher cognitive reflection, political interest, or prior exposure to deepfake warnings can nudge scores upward (Diel et al., 2024; Somoray et al., 2025).
But “internet savvy” here is a broad, self‑reported mix of familiarity with terms like phishing and selfie plus general tech‑use questions (Pehlivanoglu et al., 2026). It is not a clean, actionable construct. And these higher‑performing observers are not “the public”; they are younger, tech‑comfortable undergraduates who volunteer for long online studies (Groh et al., 2021; Fooled Twice, 2021).
Crucially, in UF’s image task, those individual differences explain essentially none of the variance (Pehlivanoglu et al., 2026). For static fakes, the emerging story is that it does not matter how savvy you think you are—unaided performance is near random (Bray et al., 2023; Diel et al., 2024).
What the broader research actually says. Stepping back, the pattern across studies is clear.
Large image‑based experiments find human accuracy around 60–64%, with people most confident when they are wrong (Bray et al., 2023; Groh et al., 2021). Systematic reviews and meta‑analyses report that, across images and videos, human detection is often near chance and heavily shaped by bias and context (Diel et al., 2024; Somoray et al., 2025). In some video experiments, people reach ~70% on “easy” fakes but collapse to below chance on high‑quality fakes or under time pressure (Köbis et al., 2021; Rossetto et al., 2023).
If your takeaway from UF is “Whew, my brain can probably tell,” you have missed the forest for a couple of carefully curated trees.
Why this matters for fraud—and where IDShield, or one of its better competitors fits in
This is not just methodological nitpicking. “Trust your eyes” is becoming a dangerous piece of folk wisdom.
Criminals already use deepfakes in social engineering: cloned voices, synthetic “CEO” or “pastor” videos, fake customer‑service reps, and family‑member scams that lean on urgency, secrecy, and emotional leverage (Jericho Security, 2025; Breacher.ai, 2025). In one case, a deepfaked CFO on a video call helped thieves walk off with roughly $25 million (Jericho Security, 2025). Other attacks have used AI‑cloned voices of executives or relatives to push urgent wire transfers and ransom‑style payments (Breacher.ai, 2025; NCOA, 2025).
The National Council on Aging reports that older adults lost billions to fraud in 2023, and deepfake‑enabled “grandparent scams” now use voice cloning to simulate a grandchild in distress (NCOA, 2024, 2025). Telling people “you’ll know a deepfake when you see it” is comforting—and false.
By all means, cultivate skepticism and slow thinking. Verify through second channels before acting on anything that combines urgency, secrecy, and money or credentials. But recognize that even your best effort will not catch every attack. The data say unaided human detection is mediocre and biased, especially when fakes are high quality and contexts are messy (Bray et al., 2023; Diel et al., 2024; Somoray et al., 2025; Pehlivanoglu et al., 2026).
That is why I argue for layered defenses beyond “I’ll know it when I see it.” You want:
- Continuous monitoring for new accounts, credit activity, and dark‑web exposure when a deepfake‑enabled scam does succeed.
- Professional help to clean up after identity theft—especially as more institutions start encountering synthetic “evidence.”
- Guidance on verification procedures so a single convincing video or voice cannot override common sense.
That is the space where services like IDShield and LegalShield live: not as magical deepfake detectors, but as part of a realistic response to a world where we must assume someone can and will fake our voices, faces, and stories. If my own eyes and ears can be fooled by a neural net—and the data say they can—then my defense plan has to be bigger than my confidence.
References
Breacher.ai. (2025, June 24). *7 alarming deepfake attack examples you need to know*. https://breacher.ai [breacher](https://breacher.ai/blog/deepfake-attack-examples/)
Bray, J., Johnson, S. D., & Kleinberg, B. (2023). Testing human ability to detect “deepfake” images of human faces. Journal of Cybersecurity, 9 (1), tyad011. https://academic.oup.com/cybersecurity/article/9/1/tyad011/7205694 [academic.oup](https://academic.oup.com/cybersecurity/article/9/1/tyad011/7205694)
Diel, A., Lalgi, T., Schröter, I. C., MacDorman, K. F., Teufel, M., & Bäuerle, A. (2024). Human performance in detecting deepfakes: A systematic review and meta‑analysis of 56 papers. Computers in Human Behavior Reports, 16, 100538. https://sciety.org/articles/activity/10.31219/osf.io/cxv4r [sciety](https://sciety.org/articles/activity/10.31219/osf.io/cxv4r)
Fooled twice: People cannot detect deepfakes but think they can. (2021). Cognitive Research: Principles and Implications, 6 (1), 1–18. https://pubmed.ncbi.nlm.nih.gov/34820608/ [pubmed.ncbi.nlm.nih](https://pubmed.ncbi.nlm.nih.gov/34820608/)
Groh, M., Epstein, Z., Firestone, C., & Picard, R. (2021). Deepfake detection by human crowds, machines, and machine‑informed crowds. Proceedings of the National Academy of Sciences, 118*(1), e2110013119. https://www.pnas.org/doi/10.1073/pnas.2110013119 [pnas](https://www.pnas.org/doi/10.1073/pnas.2110013119)
Jericho Security. (2025, June 3). Deepfake phishing: The AI‑powered social engineering threat putting CISOs on high alert in 2025. https://www.jerichosecurity.com/blog/deepfake-phishing-the-ai-powered-social-engineering-threat-putting-cisos-on-high-alert-in-2025 [jerichosecurity](https://www.jerichosecurity.com/blog/deepfake-phishing-the-ai-powered-social-engineering-threat-putting-cisos-on-high-alert-in-2025)
Köbis, N. C., Doležalová, J., Soraperra, I., & Soraperra, G. (2021). Fooled by the deepfake: People cannot detect deepfakes but think they can. Technology, Mind, and Behavior, 2(2). (Also discussed in Groh et al., 2021, and Somoray et al., 2025.) [researchonline.jcu.edu](https://researchonline.jcu.edu.au/86542/1/Somoray,%20Miller%20&%20Holmes,%202025.%20HB&ET.pdf)
Korshunov, P., & Marcel, S. (2020). Deepfake detection: Humans vs. machines. arXiv preprint arXiv:2009.03155. https://ui.adsabs.harvard.edu/abs/2020arXiv200903155K/abstract [ui.adsabs.harvard](https://ui.adsabs.harvard.edu/abs/2020arXiv200903155K/abstract)
National Council on Aging. (2024, October 29). Understanding deepfakes: What older adults need to know. https://www.ncoa.org/article/understanding-deepfakes-what-older-adults-need-to-know [ncoa](https://www.ncoa.org/article/understanding-deepfakes-what-older-adults-need-to-know/)
National Council on Aging. (2025, December 27). What are AI scams? A guide for older adults. https://www.ncoa.org/article/what-are-ai-scams-a-guide-for-older-adults [ncoa](https://www.ncoa.org/article/what-are-ai-scams-a-guide-for-older-adults/)
Pehlivanoglu, D., Zhu, X., & colleagues. (2026). Is this real? Susceptibility to deepfakes in machines and humans. Cognitive Research: Principles and Implications, 11(1), 3. https://pmc.ncbi.nlm.nih.gov/articles/PMC12779810/ [pmc.ncbi.nlm.nih](https://pmc.ncbi.nlm.nih.gov/articles/PMC12779810/)
Rossetto, L., Tursun, O., & Giunchiglia, F. (2023). Human performance in deepfake video detection: An experimental study. In Proceedings of the International Conference on Artificial Intelligence and Law. (Summary and statistics highlighted in Somoray et al., 2025.) [researchonline.jcu.edu](https://researchonline.jcu.edu.au/79985/1/79985.pdf)
Somoray, K., Miller, P., & Holmes, R. (2025). Human performance in deepfake detection: A systematic review. Human Behavior and Emerging Technologies, 7 (1), e1833228. https://researchonline.jcu.edu.au/86542/ [researchonline.jcu.edu](https://researchonline.jcu.edu.au/86542/)
University of Florida. (2026, February 24). Machines spot deepfake pictures better than humans, but people outperform AI in detecting deepfake videos [Press release]. University of Florida News. https://news.ufl.edu/2026/02/deepfake-detection/ [news.ufl](https://news.ufl.edu/2026/02/deepfake-detection/)
wracton@gmail.com
williamacton.legalshieldassociate.com
Caveat emptier: This post was drafted with help from an AI assistant (Perplexity)— but ideated and edited extensively by the human, Bill Acton.
No comments:
Post a Comment