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

The DOI matches its Crossref record; 103 of the 103 references it cites were found in a registry or an index; 13 of the 41 sources its claims cite were checked for support.

Deepfake detection by human crowds, machines, and machine-informed crowds

Open access
Groh, Matthew, Epstein, Ziv, Firestone, Chaz, Picard, Rosalind · 2022 · Proceedings of the National Academy of Sciences

Abstract

Significance: The recent emergence of deepfake videos raises theoretical and practical questions. Are humans or the leading machine learning model more capable of detecting algorithmic visual manipulations of videos? How should content moderation systems be designed to detect and flag video-based misinformation? We present data showing that ordinary humans perform in the range of the leading machine learning model on a large set of minimal context videos. While we find that a system integrating human and model predictions is more accurate than either humans or the model alone, we show inaccurate model predictions often lead humans to incorrectly update their responses. Finally, we demonstrate that specialized face processing and the ability to consider context may specially equip humans for deepfake detection.

Overview

CoChat Summary

TL;DR

In two online experiments, ordinary people performed near or above the leading 2020 deepfake-detection model on minimal-context DFDC videos, and crowd averages matched or exceeded the model. Combining human and model predictions improved accuracy overall, but incorrect model predictions often led participants to update toward wrong answers.

Key findings

  • Experiment 1: 5,524 individuals completed 26,820 trials; among 882 participants who saw at least 10 pairs, 82% outperformed the leading model, which achieved 65% accuracy on the holdout dataset. Experiment 2: recruited participants accurately identified deepfakes in 66% of attempts, nonrecruited participants in 69% of trials (72% for those seeing at least 10 videos), versus the leading model’s 80% on sampled videos; 13% of recruited, 27% of nonrecruited with ≥10 videos, and 37% of nonrecruited with <10 videos outperformed the model.
  • Crowd wisdom: the crowd mean accurately identified 74% of videos for recruited participants, 80% for nonrecruited participants, and 86% for the 1,879 nonrecruited participants who saw at least 10 videos, compared with 80% for the leading model. After seeing the model’s prediction, recruited participants’ accurate identification increased from 66% to 73% (P < 0.001), but on 10 videos with incorrect or equivocal model predictions, participants became 2.7% less accurate on average; on political deepfakes the model was confidently wrong (2%, 8%, and 1% deepfake probability) and participants’ accurate identification dropped from 56% to 34% for Kim Jong-un and 70% to 55% for Vladimir Putin.
  • Specialized face processing: inversion, misalignment, and eye occlusion reduced human accuracy by 4.3 percentage points (P = 0.002), 6.3 percentage points (P < 0.001), and 4.4 percentage points (P = 0.004) on a base rate of 65.5%; the model was affected only by inversion, dropping 12.1 percentage points (P = 0.005). Anger elicitation had no overall significant effect (P = 0.280), but on real videos the anger group underperformed controls by 5.2 percentage points (P = 0.032).
  • Video features: a flickering face increased recruited participants’ accuracy by 24.2 percentage points (P < 0.001) and the model’s by 21.7 percentage points (P = 0.170); two people increased recruited participants’ accuracy by 7.6% (P < 0.001) but decreased the model’s accuracy by 21.9% (P = 0.023) on real videos; dark-skin presence decreased recruited participants’ accuracy by 8.8% (P < 0.001) and the model’s by 12.0% (P = 0.192).

Methods

  • Two randomized online experiments on the Detect Fakes website. Experiment 1 used a two-alternative forced-choice design with 56 pairs of videos from the DFDC training dataset (n = 5,524; 26,820 trials; analysis focused on 882 participants who saw at least 10 pairs). Experiment 2 used a single-video design (n = 9,492: 304 recruited from Prolific and 9,188 organic visitors) with 50 videos randomly sampled from the DFDC holdout set (half deepfake, half nonmanipulated), 4 political-leader videos, and a deepfake attention check; participants reported confidence from 50% to 100% in 1% increments and could update after seeing the model’s prediction.
  • Embedded randomized interventions tested anger elicitation and obstruction of specialized face processing via inversion, misalignment, and occlusion. Analyses used linear regression with robust SEs clustered on participants unless otherwise stated; accuracy was normalized between 0 and 1, with crowd means, ROC/AUC comparisons, and hand-labeled video-level features.

Limitations

  • Paper states: the study evaluated 167 videos (84 deepfake and 83 authentic); only the two political deepfake videos include lip-syncing manipulations; no expert fact-checkers or forensic analysts were recruited, so results generalize only to ordinary people; the comparison is limited to the best model in 2020; minimal-context videos may not resemble the most problematic deepfakes; the half-real/half-deepfake base rate does not reflect real misinformation base rates; and emotion results were mixed, so no firm conclusions are drawn.
  • Paper states: future work is needed on persuasive manipulated video, larger samples of political deepfakes, social media context, and additional or integral emotions.
  • Evident from text: the nonrecruited participants were not preregistered and arrived organically/self-selected via the public website, while recruited participants came from Prolific.

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

Integrity signals

Registry

Crossref · 10.1073/pnas.2110013119

Cites

103 of 103 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1. One of these two images is the first frame of a deepfake from experiment 1; the other is the first frame of the original, authentic video from which the deepfake was created. Experiment 1 asked whether participants can tell which is which, using a two-alternative forced-choice paradigm (i.e., selecting which of two video clips is a deepfake). Experiment 2 presented a single video and asked participants for their confidence the video is a deepfake or not. (Left) The deepfake; the man was not mustachioed at the time of filming. (Right) Authentic.
Fig. 1. One of these two images is the first frame of a deepfake from experiment 1; the other is the first frame of the original, authentic video from which the deepfake was created. Experiment 1 asked whether participants can tell which is which, using a two-alternative forced-choice paradigm (i.e., selecting which of two video clips is a deepfake). Experiment 2 presented a single video and asked participants for their confidence the video is a deepfake or not. (Left) The deepfake; the man was not mustachioed at the time of filming. (Right) Authentic.CC BY-NC-ND · PubMed Central
Fig. 2. (A) The distribution of participant performance across experiments compared to the model’s performance via violin plots where the white dots indicate the mean and the black bars indicate the interquartile range. R, recruited participants; NR, nonrecruited participants; E1, experiment 1; E2, experiment 2. In experiment 1 (two-alternative forced choice), accuracy is defined as identifying a deepfake from a pair of videos correctly. In experiment 2 (single-video design), accurate identification is defined as responding with the correct answer with more than 50% confidence. The model’s performance represents a single observation in each instance, and, as such, we present the model’s performance as a horizontal black line with a white dot in the middle. The crowd mean distributions are obtained by bootstrapping CIs based on 1,000 randomly drawn samples that are each half of the total observations. (B) A scatter plot of the model’s accuracy and the distribution of participants’ accuracy scores for each video. The x axis is an index of the videos, and it is ordered by experiment, true class of each video, and participant’s average accuracy. The teal lines represent the interquartile range of recruited participants’ responses. (C) The distribution of changes in recruited participants’ accuracy after updating their response based on whether the model’s prediction is correct, incorrect, or indecisive. (D) The receiver operator characteristic curves of computer performance, recruited participants’ collective performance, and recruited participants’ collective performance with the model’s decision support across the 50 DFDC videos in experiment 2.
Fig. 2. (A) The distribution of participant performance across experiments compared to the model’s performance via violin plots where the white dots indicate the mean and the black bars indicate the interquartile range. R, recruited participants; NR, nonrecruited participants; E1, experiment 1; E2, experiment 2. In experiment 1 (two-alternative forced choice), accuracy is defined as identifying a deepfake from a pair of videos correctly. In experiment 2 (single-video design), accurate identification is defined as responding with the correct answer with more than 50% confidence. The model’s performance represents a single observation in each instance, and, as such, we present the model’s performance as a horizontal black line with a white dot in the middle. The crowd mean distributions are obtained by bootstrapping CIs based on 1,000 randomly drawn samples that are each half of the total observations. (B) A scatter plot of the model’s accuracy and the distribution of participants’ accuracy scores for each video. The x axis is an index of the videos, and it is ordered by experiment, true class of each video, and participant’s average accuracy. The teal lines represent the interquartile range of recruited participants’ responses. (C) The distribution of changes in recruited participants’ accuracy after updating their response based on whether the model’s prediction is correct, incorrect, or indecisive. (D) The receiver operator characteristic curves of computer performance, recruited participants’ collective performance, and recruited participants’ collective performance with the model’s decision support across the 50 DFDC videos in experiment 2.CC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

5 Valid · 1 Not valid · 18 Not checked · 170 Own results and interpretations

126 more claims

Create a free account to see all 194 claims and to verify the 28 references still open.

See all claims — free account
  • It may state two separate claims as one. 34
  • Probably not a substantial claim. 31
  • The quote only partly states this claim. 2

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?

12 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 100 references

Cited by

See all 248

Metadata

Authors
Groh, M., Epstein, Z., Firestone, C., & Picard, R.
Affiliations
  • MIT Media Lab (US)
  • Johns Hopkins University (US)
Corresponding author
Groh, Matthew
Venue
Proceedings of the National Academy of Sciences
Published
2021-12-28
Volume
119
Issue
1
ISSN
  • 0027-8424 (ISSN-L)
  • 1091-6490
OpenAlex
W3209972423
Cited by
248
References
100
Citations by year
2026: 64 · 2025: 55 · 2024: 65 · 2023: 43 · 2022: 20 · 2021: 1
FWCI
11.59
Type
journal article
Language
en
Open access
Locations
  • Proceedings of the National Academy of Sciences (publishedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • PhilPapers (PhilPapers Foundation) (submittedVersion)
  • arXiv (Cornell University)
Funders
MIT Media Lab member companies (NA)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Face Recognition and Perception
Keywords
deepfake detection, human-machine comparison, face processing
MeSH terms
Facial Recognition, Artificial Intelligence, Communication, Humans, Deception, Video Recording, Forensic Sciences, Social Media

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 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (qwen3.8-omni-flash and 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: Deepfake detection by human crowds, machines, and machine-informed crowds. https://app.cochat.ai/papers/10.1073/pnas.2110013119
The paper
Groh, M., Epstein, Z., Firestone, C., & Picard, R. (2022). Deepfake detection by human crowds, machines, and machine-informed crowds. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2110013119

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.