---
type: source
id: dunn-2026-super-recognizers-average
title: 'Too good to be true: Synthetic AI faces are more average than real faces and super-recognizers know it'
authors:
- James D. Dunn
- David White
- Clare A. M. Sutherland
- Elizabeth J. Miller
- Ben A. Steward
- Amy Dawel
year: 2026
venue: British Journal of Psychology
volume: null
pages: null
doi: 10.1111/bjop.70063
peer_review: peer-reviewed-journal
evidence_tier: B
design: Super-recognisers vs motivated controls classifying real vs AI faces; correlation with face-recognition tests (CFMT+, UNSW Face Test); face-space centrality from two identity DNNs (VGG2, ArcFace)
modality:
- image-face
stimuli: 100 StyleGAN2 and 100 FFHQ White faces; each participant saw one gender (50 real, 50 AI)
n_participants: 125
n_note: Super-recognisers n=36, controls n=89 after exclusions
topics:
- human-detection
- individual-differences
- hyperrealism
- mechanism
summary: Super-recognisers detected AI faces modestly above chance (57.3%, d' 0.41) while controls were at chance (50.7%); AI faces sit closer to the centre of face space, and super-recognisers use this 'hyper-averageness' as a cue.
results:
- metric: 'Accuracy: super-recognisers vs controls'
  value: 57.3% vs 50.7%
  note: d' 0.41 vs 0.04; t(66.0) = 2.58, p = .012
  source_location: null
- metric: Face recognition ability vs AI-face sensitivity (controls)
  value: r(87) = .35 [95% CI .15, .52]
  note: ''
  source_location: null
- metric: 'DNN centrality: AI vs real faces'
  value: AI faces closer to the face-space centroid
  note: b = -0.210, p < .001
  source_location: null
- metric: Use of centrality as AI cue
  value: super-recognisers yes; controls n.s.
  note: ''
  source_location: null
moderators:
- Face-recognition ability
limitations:
- White faces and StyleGAN2 only
- Effect size reported as d = 0.55 in abstract and 0.50 in results
verification:
  status: verified-fulltext
  confidence: high
  checked: 2026-10-12
  via: UNSWorks repository submitted version (hdl 1959.4/107078)
url: null
pmid: null
pmcid: null
arxiv_id: null
openalex_id: null
oa_url: null
cited_in:
- MR-0001
- MR-0002
updated: '2026-10-12'
---

# Too good to be true: Synthetic AI faces are more average than real faces and super-recognizers know it

> **Source record** `dunn-2026-super-recognizers-average` · 2026 · British Journal of Psychology  
> **Evidence tier:** B (peer-reviewed-journal) · **Verification:** verified-fulltext  
> **Bottom line:** Super-recognisers detected AI faces modestly above chance (57.3%, d' 0.41) while controls were at chance (50.7%); AI faces sit closer to the centre of face space, and super-recognisers use this 'hyper-averageness' as a cue.

## Citation {#citation}

James D. Dunn, David White, Clare A. M. Sutherland, Elizabeth J. Miller, Ben A. Steward, Amy Dawel (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

Verification entry (identifiers, resolvers, claims with source locations): [/data/citations.json#dunn-2026-super-recognizers-average](/data/citations.json#dunn-2026-super-recognizers-average)

## Study design {#design}

| Field | Value |
|---|---|
| Design | Super-recognisers vs motivated controls classifying real vs AI faces; correlation with face-recognition tests (CFMT+, UNSW Face Test); face-space centrality from two identity DNNs (VGG2, ArcFace) |
| Modality | image-face |
| Stimuli | 100 StyleGAN2 and 100 FFHQ White faces; each participant saw one gender (50 real, 50 AI) |
| Participants (N) | 125 |
| N detail | Super-recognisers n=36, controls n=89 after exclusions |
| Topics | human-detection, individual-differences, hyperrealism, mechanism |

## Results {#results}

| Metric | Value | Note |
|---|---|---|
| Accuracy: super-recognisers vs controls | 57.3% vs 50.7% | d' 0.41 vs 0.04; t(66.0) = 2.58, p = .012 |
| Face recognition ability vs AI-face sensitivity (controls) | r(87) = .35 [95% CI .15, .52] |  |
| DNN centrality: AI vs real faces | AI faces closer to the face-space centroid | b = -0.210, p < .001 |
| Use of centrality as AI cue | super-recognisers yes; controls n.s. |  |

## Moderators {#moderators}

- Face-recognition ability

## Limitations {#limitations}

- White faces and StyleGAN2 only
- Effect size reported as d = 0.55 in abstract and 0.50 in results

## Cited in {#cited-in}

- [MR-0001](/metareports/human-detection-accuracy.md): Human Detection of Deepfakes: Accuracy, Bias, Confidence, and Training
- [MR-0002](/metareports/uncanny-valley-and-synthetic-faces.md): The Uncanny Valley and Synthetic Media: Does It Exist, What Causes It, and Have Deepfakes Crossed It?

## Verification {#verification}

| Field | Value |
|---|---|
| Status | `verified-fulltext` |
| Compiler confidence | high |
| Checked on | 2026-10-12 |
| Via | UNSWorks repository submitted version (hdl 1959.4/107078) |