Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 40 of the 40 references it cites were found in a registry or an index; 27 of the 43 sources its claims cite were checked for support.

Warning: Humans cannot reliably detect speech deepfakes

Open access
Mai, Kimberly T., Bray, Sergi, Davies, Toby, Griffin, Lewis D. · 2023 · PLOS ONE

Abstract

Speech deepfakes are artificial voices generated by machine learning models. Previous literature has highlighted deepfakes as one of the biggest security threats arising from progress in artificial intelligence due to their potential for misuse. However, studies investigating human detection capabilities are limited. We presented genuine and deepfake audio to n = 529 individuals and asked them to identify the deepfakes. We ran our experiments in English and Mandarin to understand if language affects detection performance and decision-making rationale. We found that detection capability is unreliable. Listeners only correctly spotted the deepfakes 73% of the time, and there was no difference in detectability between the two languages. Increasing listener awareness by providing examples of speech deepfakes only improves results slightly. As speech synthesis algorithms improve and become more realistic, we can expect the detection task to become harder. The difficulty of detecting speech deepfakes confirms their potential for misuse and signals that defenses against this threat are needed.

Overview

CoChat Summary

## TL;DR

This study tested 529 people on whether they could distinguish genuine from synthetic speech in English and Mandarin. Human detection was unreliable: participants identified deepfakes correctly 73% of the time in the unary setting, with no significant language difference and only a small boost from familiarization.

## Key findings

  • In the unary task, participants classified clips correctly 70.35% of the time overall; they spotted deepfakes correctly 73% of the time and bona fide clips correctly 67.78% of the time.
  • In the binary task, performance improved: participants correctly recognized deepfake audio in 85.59% of trials, and this improvement from contextual reference audio was statistically significant (p < 0.001).
  • Familiarization with five deepfake examples improved detection accuracy by 3.84% on average (p = 0.001); in the unary and binary regressions the gains were 3.76% (p = 0.017) and 3.85% (p = 0.032), respectively.
  • Detection did not significantly differ by language: Mandarin-speaking participants outperformed English-speaking participants by 1.79%, but this was not statistically significant (p = 0.202).
  • Human performance was stable over time and not helped by repeated listening: clip replays had no relationship with accuracy (ρ = −0.05, p < 0.001), and total task time also showed no meaningful improvement (ρ = 0.10, p = 0.018).

## Methods

  • Online experiment with 529 Prolific participants fluent in English or Mandarin; mean age was 28.9 years and 50.6% identified as male.
  • Stimuli were 100 clips total: bona fide speech from LJSpeech (English) and CSMSC (Mandarin), plus deepfakes generated from the same sentences using pre-trained VITS text-to-speech models; clip lengths ranged from 2 to 11 seconds.
  • Participants were randomly assigned to unary vs binary judging, and to familiarization vs control; the main analysis used linear regression on confidence-weighted accuracy scores, with additional comparisons to automated LFCC-LCNN detectors and AUROC/EER metrics.

## Limitations

  • The paper states that the setup “does not necessarily reflect more realistic scenarios” because the deepfake proportion was balanced 50/50, whereas AI-generated content is still rare, and participants were warned about deepfakes in advance.
  • The stimuli minimized contextual information, so the study does not test cases where speaker identity, emotional status, conversational context, or political affiliation might help detection.
  • The deepfakes were generated with an older approach and “may not faithfully reflect performance under the most current conditions” — stated in the paper.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1371/journal.pone.0285333

Cites

40 of 40 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig 1 Diagram of a typical generative speech synthesis model.
Fig 1 Diagram of a typical generative speech synthesis model.CC BY · PubMed Central
Fig 2 Screenshots of the task interface.
Fig 2 Screenshots of the task interface.CC BY · PubMed Central
Fig 3 Box plot summarizing human performance across the different groups.
Fig 3 Box plot summarizing human performance across the different groups.CC BY · PubMed Central
Fig 4 Scatter plot showing the relationship between clip length and confidence-scaled accuracy.
Fig 4 Scatter plot showing the relationship between clip length and confidence-scaled accuracy.CC BY · PubMed Central
Fig 5 Scatter plot showing the relationship between the number of times played and confidence-scaled accuracy.
Fig 5 Scatter plot showing the relationship between the number of times played and confidence-scaled accuracy.CC BY · PubMed Central
Fig 6 Scatter plot showing the relationship between minutes taken to complete and correctness scores.
Fig 6 Scatter plot showing the relationship between minutes taken to complete and correctness scores.CC BY · PubMed Central
Fig 7 Histogram of correct responses across question number.
Fig 7 Histogram of correct responses across question number.CC BY · PubMed Central
Fig 8 Receiver operator curves under the unary scenario.
Fig 8 Receiver operator curves under the unary scenario.CC BY · PubMed Central
Fig 9 Word clouds containing justifications for the English-language clips.
Fig 9 Word clouds containing justifications for the English-language clips.CC BY · PubMed Central
Fig 10 Word clouds containing justifications for the Mandarin-language clips. Note participants for the Mandarin tasks provided justifications in both Mandarin and English.
Fig 10 Word clouds containing justifications for the Mandarin-language clips. Note participants for the Mandarin tasks provided justifications in both Mandarin and English.CC BY · PubMed Central

Verification

Claims the paper makes

22 Valid · 0 Not valid · 13 Not checked · 63 Own results and interpretations

39 more claims · 2 with their cited sources checked

Create a free account to see all 98 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 41
  • It may state two separate claims as one. 18
  • The quote only partly states this claim. 3

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 37 references

Research works printed but not linked: 10

  1. I Goodfellow, 2016, Deep Learning
  2. N Schick, 2020, Deep Fakes and the Infocalypse: What You Urgently Need To Know
  3. Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, et al. Generative Adversarial Nets. In: Advances in Neural Information Processing Systems; 2014. Available from: https://proceedings.neurips.cc/paper_files/paper/2014/file/5ca3e9b122f61f8f06494c97b1afccf3-Paper.pdf
  4. Geirhos T, Zimmerman RS, Bilodeau BL, Brendal W, Kim, B. Don’t trust your eyes: on the (un)reliability of feature visualizations. arXiv preprint arXiv: 230604719. 2023 Jun 13 [Cited 2023 Jun 13].
  5. Watson G, Khanjani Z, Janeja VP. Audio Deepfake Perceptions in College Going Populations. arXiv preprint arXiv:211203351. 2021 Dec 6 [Cited 2023 Jan 19].
  6. Dolhansky B, Bitton J, Pflaum B, Lu J, Howes R, Wang M, et al. The Deepfake Detection Challenge Dataset. arXiv preprint arXiv:200607397. [Cited 2023 Mar 27]. Dataset.
  7. K Kumar, 2019, Proceedings of Advances in Neural Information Processing Systems
  8. Kim J, Kong J, Son J. Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech. In: International Conference on Machine Learning. PMLR; 2021. p. 5530–5540. Available from: https://proceedings.mlr.press/v139/kim21f.html.
  9. Ma H, Yi J, Wang C, Yan X, Tao J, Wang T, et al. FAD: A Chinese Dataset for Fake Audio Detection. arXiv preprint arXiv:220712308. 2022 Jul 12. [Cited 2023 Jan 19].
  10. Chesney R, Citron D. Deepfakes and the new disinformation war: The coming age of post-truth geopolitics. Foreign Affairs. 2019 Jan/Feb;98:147.

Cited by

See all 98

Metadata

Authors
Mai, K. T., Bray, S., Davies, T., & Griffin, L. D.
Affiliations
University College London (GB)
Venue
PLOS ONE
Published
2023-08-02
Volume
18
Issue
8
Pages
e0285333
ISSN
1932-6203 (ISSN-L)
OpenAlex
W4385483249
Cited by
98
References
37
Citations by year
2026: 42 · 2025: 34 · 2024: 18 · 2023: 4
FWCI
9.21
Type
journal article
Language
en
Open access
Locations
  • PLoS ONE (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
Funders
  • Engineering and Physical Sciences Research Council (EP/R513143/1, EP/S022503/1)
  • Dawes Centre for Future Crime
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Speech Recognition and Synthesis
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Physical Sciences › Computer Science › Signal Processing › Speech and Audio Processing
Keywords
deepfake detection, synthetic speech, speech synthesis, AI misuse, decision-making rationale
MeSH terms
Artificial Intelligence, Humans, Language, Phonetics, Speech, Speech Perception

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-30, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-09-30 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Warning: Humans cannot reliably detect speech deepfakes. https://app.cochat.ai/papers/10.1371/journal.pone.0285333
The paper
Mai, K. T., Bray, S., Davies, T., & Griffin, L. D. (2023). Warning: Humans cannot reliably detect speech deepfakes. PLOS ONE. https://doi.org/10.1371/journal.pone.0285333

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.