Free lesson

Test Your Eye: How Well Can You Really Spot AI?

Test Your Eye: How Well Can You Really Spot AI? — Spot Deepfakes and AI Fakes: Check Photos, Videos and Voices. Start with a test, not a list of clues; What your score can—and cannot—mean; Plausibility and trust are not evidence; A fake often wins before checking begins; Replace the sharper-eye test with a checkable process; Record your starting point.…

Demo edition. The materials are undergoing editorial review.

Topics covered

  • Start with a test, not a list of clues
  • What your score can—and cannot—mean
  • Plausibility and trust are not evidence
  • A fake often wins before checking begins
  • Replace the sharper-eye test with a checkable process
  • Record your starting point

Start with a test, not a list of clues

First measure the ability people often rely on: visual judgment. If whichfaceisreal.com is available, complete 20 rounds, choosing the real face from each pair. If it is unavailable, use a university or newsroom real-versus-AI test with known answers. Before starting, rate your confidence from 1 to 5, where 1 means “I am guessing” and 5 means “I am very unlikely to be wrong.” Do not study clues first, and record every answer. After 20 rounds, write down your score out of 20 and rate your confidence again. This is not a skill certification: the sample is small, and difficulty depends on the images used. Its purpose is to compare your feeling of certainty with a measured result.

What your score can—and cannot—mean

A single quiz result cannot be turned into a universal personal accuracy rate. Research does, however, show why unaided sight is insufficient. A 2024 meta-analysis combined 56 papers and 86,155 participants. Its pooled deepfake-detection accuracy was 55.54%, with a 95% confidence interval of 48.87% to 62.10%, which crossed chance. The studies used different media and tasks, so 55.54% is not a promised score for every person or test. In a separate 2022 study, 315 participants classified StyleGAN2 faces with 48.2% average accuracy. Another 219 participants received a short tutorial and answer-by-answer feedback and reached 59.0%. Training helped, but it did not turn visual judgment into reliable proof.

Plausibility and trust are not evidence

In the same 2022 study, a separate group of 223 participants rated faces for trustworthiness on a scale from 1 to 7. Real faces averaged 4.48; synthetic faces averaged 4.82. That finding belongs to a particular, carefully selected portrait set and is not a rule about every image. A 2023 study also found “AI hyperrealism”: White faces made by StyleGAN2 were judged human more often than photographs of White people. The authors did not find the same effect for non-White faces in reanalysed data. Reddit discussion raised a useful limitation: the selection and matching of portraits may affect what looks real. The practical conclusion is narrow but important. A face that seems pleasant, familiar, or candid does not prove an image’s origin, and strong confidence does not repair weak evidence.

A fake often wins before checking begins

Image quality is only part of the problem. A message may push immediate action: “share this now,” “do not call anyone,” or “pay urgently.” The deceiver then needs no perfect fake; the material only needs to hold up until the recipient acts. Fictional example: Lars is a 47-year-old purchasing manager at a company with about 120 employees. He may approve supplier payments up to $20,000 without a second signature. In a 14-person family chat, he sees a video labelled “bridge collapsing in a storm,” reacts with a shocked emoji, and forwards it to two colleagues. He never establishes whether it is real. His error did not occur during careful frame analysis. It occurred because no check began at all.

Replace the sharper-eye test with a checkable process

Visual oddities can justify further checking, but they are not final verdicts. Compression, motion blur, unusual lighting, and editing can produce defects in authentic recordings. A well-made synthetic item may contain no obvious defect. The working question is therefore: “Where did this come from, and what would it cost to be wrong?” The rest of the course moves from a quick decision about whether a claim deserves checking to source tracing and the examination of photos, place, time, and video. It then covers cloned voices, video calls, provenance labels, detectors, attempts to dismiss real evidence, fast-moving events, personal protection, and a reusable verification protocol. The aim is trust proportionate to the evidence—not automatic belief or blanket suspicion.

Record your starting point

Create a short record that you can revisit at the end of the course. Include: confidence before the test from 1 to 5; correct answers out of 20; percentage calculated as correct ÷ 20 × 100; confidence after the test; the last photo, video, or audio item you believed without checking; where it reached you; what you did; and whether you checked it later. For example, a fictional score of 11 out of 20 is 55%, because 11 ÷ 20 × 100 = 55. Do not treat a match with a research average as a diagnosis of your ability. Look for a gap between confidence, performance, and behaviour: you might score well while still forwarding consequential material without checking its source.

Practical assignment

My starting point

Continue to learning

Course outline

  1. Test Your Eye: How Well Can You Really Spot AI?
  2. Seconds: Is This Worth Checking?
  3. Where Did This Picture Come From? A Two-Minute Reverse Search
  4. Anatomy of an AI Image: Clues, Not Verdicts
  5. Real Photo, False Story: Cheapfakes
  6. Where and When Was This Taken?
  7. Video: Keyframes, Seams and AI Clips
  8. A Voice on the Phone: Clones and the Callback Rule
  9. The Video Call From “Your Boss”
  10. Labels and Watermarks: What Provenance Can and Can't Tell You
  11. AI Detectors: One Piece of Evidence Among Many
  12. “That’s a Deepfake!” When People Lie About Fakes
  13. Breaking News, Elections and Disasters: The First Hours
  14. If Someone Makes a Deepfake of You or Someone You Love
  15. Protect Your Voice and Face
  16. How to Correct a Relative in the Family Chat
  17. Your Verification Protocol: 30 Seconds, 5 Minutes, 30 Minutes