---
type: source
id: dawel-2026-training-ai-faces
title: Training humans to detect AI-generated faces
authors:
- Amy Dawel
- Tanya George
- Eric Y. Mah
- James D. Dunn
- Clare A. M. Sutherland
- Nick Argument
- James W. Tanaka
year: 2026
venue: Proceedings of the National Academy of Sciences (PNAS)
volume: 123(27)
pages: e2602122123
doi: 10.1073/pnas.2602122123
peer_review: peer-reviewed-journal
evidence_tier: B
design: Pre-post training study on untrained test faces; test-retest control study; online replication
modality:
- image-face
stimuli: AI-generated and real faces (details not in abstract)
n_participants: 45
n_note: Main pre-post study N=45; control and online replication Ns not in abstract
topics:
- human-detection
- training
- confidence
- hyperrealism
summary: Training people to notice the global impressions AI faces systematically evoke (rather than artefacts) improved every participant, roughly doubled mean accuracy, and made confidence well calibrated; a test-retest control ruled out practice effects.
results:
- metric: Pre-post accuracy
  value: all participants improved; mean accuracy nearly doubled
  note: exact values not in abstract
  source_location: null
- metric: High performers
  value: near-perfect detection after training
  note: ''
  source_location: abstract
- metric: Confidence
  value: well calibrated only after training
  note: ''
  source_location: abstract
- metric: Controls
  value: test-retest control ruled out practice effects; online replication supported scalability
  note: ''
  source_location: abstract
moderators: []
limitations:
- Small main sample
- Retention not reported in abstract
- Values not checked against full text
verification:
  status: verified-abstract
  confidence: medium
  checked: 2026-10-12
  via: OpenAlex abstract record for doi:10.1073/pnas.2602122123 (publisher page not reachable)
url: null
pmid: null
pmcid: null
arxiv_id: null
openalex_id: null
oa_url: null
cited_in:
- MR-0001
updated: '2026-10-12'
---

# Training humans to detect AI-generated faces

> **Source record** `dawel-2026-training-ai-faces` · 2026 · Proceedings of the National Academy of Sciences (PNAS)  
> **Evidence tier:** B (peer-reviewed-journal) · **Verification:** verified-abstract (compiler confidence: medium)  
> **Bottom line:** Training people to notice the global impressions AI faces systematically evoke (rather than artefacts) improved every participant, roughly doubled mean accuracy, and made confidence well calibrated; a test-retest control ruled out practice effects.

## Citation {#citation}

Amy Dawel, Tanya George, Eric Y. Mah, James D. Dunn, Clare A. M. Sutherland, Nick Argument, James W. Tanaka (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

Verification entry (identifiers, resolvers, claims with source locations): [/data/citations.json#dawel-2026-training-ai-faces](/data/citations.json#dawel-2026-training-ai-faces)

## Study design {#design}

| Field | Value |
|---|---|
| Design | Pre-post training study on untrained test faces; test-retest control study; online replication |
| Modality | image-face |
| Stimuli | AI-generated and real faces (details not in abstract) |
| Participants (N) | 45 |
| N detail | Main pre-post study N=45; control and online replication Ns not in abstract |
| Topics | human-detection, training, confidence, hyperrealism |

## Results {#results}

| Metric | Value | Note |
|---|---|---|
| Pre-post accuracy | all participants improved; mean accuracy nearly doubled | exact values not in abstract |
| High performers | near-perfect detection after training |  |
| Confidence | well calibrated only after training |  |
| Controls | test-retest control ruled out practice effects; online replication supported scalability |  |

## Limitations {#limitations}

- Small main sample
- Retention not reported in abstract
- Values not checked against full text

## Cited in {#cited-in}

- [MR-0001](/metareports/human-detection-accuracy.md): Human Detection of Deepfakes: Accuracy, Bias, Confidence, and Training

## Verification {#verification}

| Field | Value |
|---|---|
| Status | `verified-abstract` |
| Compiler confidence | medium |
| Checked on | 2026-10-12 |
| Via | OpenAlex abstract record for doi:10.1073/pnas.2602122123 (publisher page not reachable) |

Values on this page have not all been checked against the full text. Confirm them against the primary source before high-stakes use.