What the research found
Researchers at Indiana University pooled data from 56 studies and over 86,000 participants, measuring how well people spot deepfakes across audio, images, text, and video. The average detection rate landed at 55.54%. A coin flip gets you 50%.
Accuracy varied by format. People caught 62% of fake audio clips, 57% of fake video, 53% of fake images, and 52% of fake text. Every number sits close to random guessing.
University of Florida researchers found something related. People who mistake a fake image for a real one share one thing in common, a default trust in what they see. Kevin Warren, a PhD student on the research team, described the pattern as wanting to believe what you hear or see is real.
Why training beats instinct
You can improve at this. The meta-analysis found three approaches that raised accuracy above chance, getting feedback after each guess, pairing your own judgment with an AI detection tool, and slowing down instead of reacting to a first impression. None of these require technical skill. All three require you to stop trusting your gut on this one specific task.
The practical takeaway
Treat any video, image, or audio clip you weren’t expecting with real suspicion before reacting, and give extra scrutiny to anything tied to money, a family emergency, or a request to act fast. Reverse image search a screenshot. Call the person back on a number you already have saved, not one sent in the message itself. Pause before sharing anything that triggers a strong emotional reaction. That reaction is often the point of the fake.
This meta-analysis measured detection under controlled study conditions, not a split-second scroll through a feed. Under real distraction and time pressure, accuracy in daily life may sit lower than these numbers.
Reference
Diel, A., et al. “Human performance in detecting deepfakes: A systematic review and meta-analysis of 56 papers.” Computers in Human Behavior Reports, 2024. https://www.sciencedirect.com/science/article/pii/S2451958824001714
Top comments (0)