MR-0001 · Human Detection of Deepfakes: Accuracy, Bias, Confidence, and Training

Question: How accurately can people tell AI-generated faces, voices, and videos from real ones, and what changes that accuracy?
Answer in one line: On average, not reliably better than a coin flip: 55.54% pooled across 56 papers, with a 95% CI of 48.87–62.10 that includes chance. Synthetic still faces are often judged more real than real faces. Cloned voices of unfamiliar speakers are barely distinguishable from the real speaker. People are confident regardless of accuracy, and training mostly adds about 0–20 points.
Version: 0.4.0 · Updated: 2026-10-19 · Sources: 33 (13 checked against full text, 17 against the official abstract, 0 DOI-only, 3 unverified; see Verification status).

TL;DR

• Near chance overall: 55.54% pooled (95% CI 48.87–62.10), 56 papers, 86,155 participants diel-2024-deepfake-detection-meta.
• Still faces: StyleGAN2 faces 48.2%, rated 7.7% more trustworthy nightingale-farid-2022; White AI faces judged human 65.9% vs 51.1% miller-2023-ai-hyperrealism; 69% judged real by 2,203 people pehlivanoglu-2026-machines-humans.
• Confidence is uninformative or inverted (r = .385, errors vs confidence) miller-2023-ai-hyperrealism. Training: +0–20 points, short-term; AI-text baseline 57.4% (F7).
• More channels help: 57% (transcript) → 85–86% (video+audio) groh-2024-political-speech. Voice clones pass as the speaker (MR-0003).
• AI advice helps only when right kramer-2026-aggregation-human-ai; plausible fakes slip through plohl-2026-political-deepfakes-plausibility; experts and LLMs are similar on AI chest X-rays (75.8% vs 77.9%) lee-2026-ai-chest-radiographs.

Key numbers

Key findings (citable)

Confidence (editorial): High = ≥3 consistent peer-reviewed studies, or one verified large-N study with replications; Moderate = fewer or heterogeneous; Low = one study or unverified.

Evidence table: accuracy by study

Accuracy = % correct real-vs-fake judgments; chance 50%; "≈" = approximate or pending verification.

Audio-only studies: MR-0003. n/r = not recorded. Machine-readable versions: /data/studies.csv and /data/studies.json.

Chart: pooled accuracy by modality

Text equivalent: Diel et al. (2024) report these pooled accuracies:

• audio 62.08% [38.23, 83.18], k = 8
• video 57.31% [47.80, 66.57], k = 26
• images 53.16% [42.12, 64.64], k = 18
• text 52.00% [37.42, 65.88], k = 15
• overall 55.54% [48.87, 62.10]
• with improvement strategies 65.14% [55.21, 74.46]

Every modality CI except the "strategies" one includes 50%. The line marks chance.

xychart-beta
    title "Pooled human deepfake-detection accuracy (Diel et al. 2024)"
    x-axis ["Audio", "Video", "Images", "Text", "Overall", "With strategies"]
    y-axis "Accuracy (%)" 0 --> 100
    bar [62.08, 57.31, 53.16, 52.00, 55.54, 65.14]
    line [50, 50, 50, 50, 50, 50]

Theme 1 — Detection is at or near chance

Pooled accuracy across 56 papers is 55.54%; the CI (48.87–62.10) and pooled d′ include chance diel-2024-deepfake-detection-meta. Most of a 3,002-person sample in the USA, Germany and China was "effectively guessing" frank-2024-representative-study; a 1,276-person US study found mean detection near 50% cooke-2024-coin-toss. Most studies use pre-2023 generators and warn participants, so real-world accuracy is likely lower (see Gaps).

Theme 2 — AI hyperrealism: synthetic faces look more real than real faces

• Effect: White StyleGAN2 faces judged human 65.9% vs 51.1% (reanalysis 69.5% vs 52.2%); non-White at chance miller-2023-ai-hyperrealism. AI audio, images and text rated more human-like in three countries frank-2024-representative-study; unverified reports agree (lago-2022-more-real-than-real, tucciarelli-2022-realness).
• Mechanism: proportions, familiarity and memorability misread as human; a classifier on 14 rated attributes reached 94% miller-2023-ai-hyperrealism. AI faces are face-space-central, a cue super-recognisers use dunn-2026-super-recognizers-average, and better remembered sulimani-2026-recognition-ai-faces.
• Race dependence: StyleGAN2 data ≈69% White miller-2023-ai-hyperrealism; male White faces classified worst nightingale-farid-2022. See MR-0002.

Theme 3 — Synthetic faces are trusted more

Synthetic faces averaged 4.82 vs 4.48 for real faces on a 1–7 scale, 7.7% higher (t(222) = 14.6, p < 0.001, d = 0.49); not explained by smiling (65.5% of real vs 58.8% of synthetic faces smiled); women's faces rated more trustworthy, 4.94 vs 4.36 (d = 0.82) nightingale-farid-2022. A synthetic persona starts with a small trust advantage.

Theme 4 — Overconfidence and response bias

• Inverted confidence: r = .385 between errors and confidence; meta-d′ negative for 59% miller-2023-ai-hyperrealism; mean confidence 6.83/10, unrelated to accuracy bray-2023-deepfake-faces.
• "Seeing is believing": people overestimate their ability and lean "authentic" kobis-2021-fooled-twice; 69% of StyleGAN2 faces judged real pehlivanoglu-2026-machines-humans.
• Modality-specific insight: metacognitive AUROC .59 faces, .48 voice clones ivory-2026-faces-voices-domain; none for video kramer-2026-aggregation-human-ai.
• Industry (not peer reviewed): ≈0.1% of ≈2,000 consumers got every item right iproov-2025-consumer-study.

Theme 5 — Training and interventions

Pattern: artifact training and accurate AI help most; generic advice, awareness and incentives do little; some gains are criterion shifts; no retention data.

Theme 6 — Modality: more channels, more cues

• Political speeches: 57% / 64% / 81% / 85–86% (transcript / silent video / audio / video+audio; Exp 3: 58% / 62% / 65% / 74%); TTS fakes caught 72% vs 83% for voice actors groh-2024-political-speech.
• Pooled: audio 62.08% > video 57.31% > images 53.16% > text 52.00%, CIs overlapping diel-2024-deepfake-detection-meta. Foreign-language, single-modality and mixed items lower accuracy cooke-2024-coin-toss.
• Speech and voice clones: 70.35% single TTS clip, 85.59% paired mai-2023-speech-deepfakes; 59.11% in Spanish/Japanese listeners san-segundo-2026-audio-language-style; clones d′ ≈ 0.1 lavan-2025-voice-clones-hyperrealism, 53.95% vs 65.48% for faces ivory-2026-faces-voices-domain. Full synthesis: MR-0003.

Theme 7 — Humans vs machines, and teaming

• Video: human crowds ≈ leading DFDC model, different errors groh-2022-crowds-machines; guidance helped only when accurate kramer-2026-aggregation-human-ai; humans and AI attend to different cues chen-2026-humans-ai-deepfake-teaming.
• Still images: humans 38.7% error vs best model 13% lu-2023-fake2m; a CNN scored 97% on StyleGAN2 faces (humans 49%) but was near chance on video pehlivanoglu-2026-machines-humans.
• Experts and LLMs: on 160 chest X-rays (75% synthetic), all four human readers caught 82–100% of diffusion images, but radiologists called 38–45% of real ones fake and nonradiologists were near chance; LLMs caught 32–72% of LLM-generated images (κ = 0.213) lee-2026-ai-chest-radiographs.
• Detector brittleness: speech AUROC 100% in-domain, 25.31% out-of-domain; listener crowds 95.51% mai-2023-speech-deepfakes. An X-ray detector: 98.3% in-distribution, 12% sensitivity on a held-out generator lee-2026-ai-chest-radiographs. Benchmark accuracy does not reliably transfer to in-the-wild media naseem-2026-in-the-wild-detection-review.
• Folk strategies: 12 strategies in 13,098 Reddit posts and 222,060 comments shifted over 32 months yeung-2026-isthisai-strategies. Implicit signals: EEG separates real from AI faces moshel-2022-decoding-ai-faces; see MR-0002.

Moderators summary

Training, plausibility, media literacy and AI-platform engagement: F7, F11, F14. Audio moderators: MR-0003.

Contradictions and caveats

• Headline accuracies (≈48% to ≈86%) differ by task (modality, generator, curation, pairing, warnings), not by disagreement.
• The pooled CI includes chance but the "strategies" CI does not: "at chance" holds for unaided judgments.
• Frank et al. find significant predictors in a population that was "effectively guessing".

Evidence gaps

1. No peer-reviewed human-detection study of 2025–2026 video generators (Sora 2, Veo 3, Kling era) or real-time face swaps in live calls was found.
2. No retention studies of detection training.
3. Few tests under realistic conditions: no warning, real context, something at stake.
4. Hyperrealism untested for non-White faces with current diffusion generators.
5. Voice-clone gaps: see MR-0003.
6. Combined-channel attacks (video call, voice, chat) untested.

Implications (editorial inference, not study findings)

• Do not make perception the control; verify by process: callbacks, passphrases, multiple approvers (F1, F2, F5, F12).
• Measure calibration, and report hit and false-alarm rates in training (F5, F7).
• "Looks real" and "looks trustworthy" are not evidence (F2, F4).
• Pair humans with in-domain detectors; people follow wrong advice (F9, F10).
• Prefer comparison with a known-genuine reference (MR-0003.F6).

Answers to common questions

Q: What percentage of people can spot a deepfake? A: 55.54% on average; the CI includes chance diel-2024-deepfake-detection-meta. Curated synthetic faces: 48.2% (MR-0001.F1, F2).

Q: Can people tell a cloned voice from the real person? A: Mostly not, for unfamiliar speakers (MR-0001.F12; detail in MR-0003).

Q: Does deepfake awareness training work? A: Partially and briefly: +0–20 points, some of it criterion shift, no retention data (MR-0001.F7).

Q: Are humans or AI detectors better? A: Models win on still images; out-of-domain detectors can fail badly while human crowds stay robust (MR-0001.F10).

Method of this metareport

• Type: narrative synthesis; pooled values from Diel et al. (2024). Inclusion: empirical human real-vs-synthetic judgments, peer-reviewed preferred; one labelled industry report (tier E).
• Verification: each Key-numbers row names its level; unverified values are marked "≈".
• Search limits: v0.1.0 had no live search; v0.2.0–v0.4.0 added 2024–2026 studies via PubMed, OpenAlex and reference lists, without general web search. Coverage is incomplete.

Verification status

Values from unverified records are marked "≈" or "unverified" in this report. Check them against the primary source before high-stakes use. Each source page states exactly what was checked and how.

References

Each entry links to its citations.json entry (identifiers, claims, verification) and its source record.

• barrington-2025-voice-clones · record · Barrington, S., et al. (2025). People are poorly equipped to detect AI-powered voice clones. Scientific Reports, 15, 11004. · https://doi.org/10.1038/s41598-025-94170-3
• bray-2023-deepfake-faces · record · Bray, S. D., et al. (2023). Testing human ability to detect 'deepfake' images of human faces. Journal of Cybersecurity, 9(1), tyad011. · https://doi.org/10.1093/cybsec/tyad011
• chen-2026-humans-ai-deepfake-teaming · record · Chen, X., et al. (2026). How Humans and AI Detect Deepfakes Differently: Complementary Strategies for Human-AI Teaming. Proceedings of the Human Factors and Ergonomics Society Annual Meeting. · https://doi.org/10.1177/10711813261487968
• cooke-2024-coin-toss · record · Cooke, D., et al. (2025). As good as a coin toss: Human detection of AI-generated images, videos, audio, and audiovisual stimuli. Communications of the ACM, 68, 100-109. · https://doi.org/10.1145/3729417
• dawel-2026-training-ai-faces · record · Dawel, A., et al. (2026). Training humans to detect AI-generated faces. Proceedings of the National Academy of Sciences (PNAS), 123(27), e2602122123. · https://doi.org/10.1073/pnas.2602122123
• diel-2024-deepfake-detection-meta · record · Diel, A., et al. (2024). Human performance in detecting deepfakes: A systematic review and meta-analysis of 56 papers. Computers in Human Behavior Reports, 16, 100538. · https://doi.org/10.1016/j.chbr.2024.100538
• dunn-2026-super-recognizers-average · record · Dunn, J. D., et al. (2026). Too good to be true: Synthetic AI faces are more average than real faces and super-recognizers know it. British Journal of Psychology. · https://doi.org/10.1111/bjop.70063
• frank-2024-representative-study · record · Frank, J., et al. (2024). A Representative Study on Human Detection of Artificially Generated Media Across Countries. IEEE Symposium on Security and Privacy (S&P 2024). · https://arxiv.org/abs/2312.05976
• gray-2025-super-recognizers · record · Gray, K. L. H., et al. (2025). Training human super-recognizers' detection and discrimination of AI-generated faces. Royal Society Open Science, 12(11), 250921. · https://doi.org/10.1098/rsos.250921
• groh-2022-crowds-machines · record · Groh, M., et al. (2022). Deepfake detection by human crowds, machines, and machine-informed crowds. Proceedings of the National Academy of Sciences (PNAS), 119(1), e2110013119. · https://doi.org/10.1073/pnas.2110013119
• groh-2024-political-speech · record · Groh, M., et al. (2024). Human detection of political speech deepfakes across transcripts, audio, and video. Nature Communications, 15, 7629. · https://doi.org/10.1038/s41467-024-51998-z
• ivory-2026-faces-voices-domain · record · Ivory, M. R., et al. (2026). Human detection of AI-generated faces and voices is not domain-general. Scientific Reports, 16, 28639. · https://doi.org/10.1038/s41598-026-59364-3
• kramer-2024-crowds-synthetic-faces · record · Kramer, R. S. S., & Cartledge, C. (2024). Crowds Improve Human Detection of AI-Synthesised Faces. Applied Cognitive Psychology, 38(5). · https://doi.org/10.1002/acp.4245
• kramer-2026-aggregation-human-ai · record · Kramer, R. S. S., & Lister, V. P. M. (2026). Improving human performance in deepfake detection: limits of aggregation and human–AI collaboration. Cognitive Research: Principles and Implications. · https://doi.org/10.1186/s41235-026-00758-2
• kobis-2021-fooled-twice · record · Köbis, N. C., et al. (2021). Fooled twice: People cannot detect deepfakes but think they can. iScience, 24(11), 103364. · https://doi.org/10.1016/j.isci.2021.103364
• lago-2022-more-real-than-real · record · Lago, F., et al. (2022). More real than real: A study on human visual perception of synthetic faces. IEEE Signal Processing Magazine, 39(1), 109-116. · https://doi.org/10.1109/MSP.2021.3120982
• lavan-2025-voice-clones-hyperrealism · record · Lavan, N., et al. (2025). Voice clones sound realistic but not (yet) hyperrealistic. PLOS ONE, 20(9), e0332692. · https://doi.org/10.1371/journal.pone.0332692
• lee-2026-ai-chest-radiographs · record · Lee, S., et al. (2026). Complementary Error Profiles of Human Readers and Large Language Models in Detecting AI-Generated Chest Radiographs: A Comparative Reader Study. Journal of Imaging Informatics in Medicine. · https://doi.org/10.1007/s10278-026-02339-7
• lu-2023-fake2m · record · Lu, Z., et al. (2023). Seeing is not always believing: Benchmarking human and model perception of AI-generated images. NeurIPS 2023 Datasets and Benchmarks Track. · https://arxiv.org/abs/2304.13023
• mai-2023-speech-deepfakes · record · Mai, K. T., et al. (2023). Warning: Humans cannot reliably detect speech deepfakes. PLOS ONE, 18(8), e0285333. · https://doi.org/10.1371/journal.pone.0285333
• manhede-2026-online-markets-training · record · Manhede, T., et al. (2026). Human discernment of artificial intelligence in online markets can be shaped by experience and training. AI & Society. · https://doi.org/10.1007/s00146-026-03373-3
• miller-2023-ai-hyperrealism · record · Miller, E. J., et al. (2023). AI hyperrealism: Why AI faces are perceived as more real than human ones. Psychological Science, 34(12), 1390-1403. · https://doi.org/10.1177/09567976231207095
• moshel-2022-decoding-ai-faces · record · Moshel, M. L., et al. (2022). Are you for real? Decoding realistic AI-generated faces from neural activity. Vision Research, 199, 108079. · https://doi.org/10.1016/j.visres.2022.108079
• naseem-2026-in-the-wild-detection-review · record · Naseem, M. T., et al. (2026). From curated datasets to social-media reality: In-the-wild, multimodal, and generalizable deepfake detection. Computer Science Review, 63, 101081. · https://doi.org/10.1016/j.cosrev.2026.101081
• nightingale-farid-2022 · record · Nightingale, S. J., & Farid, H. (2022). AI-synthesized faces are indistinguishable from real faces and more trustworthy. Proceedings of the National Academy of Sciences (PNAS), 119(8), e2120481119. · https://doi.org/10.1073/pnas.2120481119
• pehlivanoglu-2026-machines-humans · record · Pehlivanoglu, D., et al. (2026). Is this real? Susceptibility to deepfakes in machines and humans. Cognitive Research: Principles and Implications, 11, 3. · https://doi.org/10.1186/s41235-025-00700-y
• plohl-2026-political-deepfakes-plausibility · record · Plohl, N., et al. (2026). How deepfake plausibility, media literacy, and personal attitudes shape detection, liking, and intentions to share political deepfakes on social media. Humanities and Social Sciences Communications. · https://doi.org/10.1057/s41599-026-09122-7
• rehman-2025-training-synthetic-faces · record · Rehman, R. S., et al. (2025). Training humans for synthetic face image detection. Frontiers in Artificial Intelligence, 8, 1568267. · https://doi.org/10.3389/frai.2025.1568267
• san-segundo-2026-audio-language-style · record · San Segundo, E., et al. (2026). Human perception of audio deepfakes: the role of language and speaking style. Speech Communication, 185, 103487. · https://doi.org/10.1016/j.specom.2026.103487
• sulimani-2026-recognition-ai-faces · record · Sulimani, N., et al. (2026). Recognition of real and artificial intelligence-generated faces. Perception, 55(10), 924-937. · https://doi.org/10.1177/03010066261442089
• tucciarelli-2022-realness · record · Tucciarelli, R., et al. (2022). On the realness of people who do not exist: The social processing of artificial faces. iScience, 25.
• yeung-2026-isthisai-strategies · record · Yeung, C., et al. (2026). Is This AI? Longitudinal Analysis of Strategies Used for AI Detection on Two Subreddits. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES), 9(4), 2800-2812. · https://doi.org/10.1609/aies.v9i4.43351
• iproov-2025-consumer-study · record · iProov (2025). iProov deepfake blindspot consumer study (industry report). Industry report / press release.