---
type: source
id: ivory-2026-faces-voices-domain
title: Human detection of AI-generated faces and voices is not domain-general
authors:
- Matthew R. Ivory
- Carolyn McGettigan
- Helen E. Nuttall
- Sophie J. Nightingale
year: 2026
venue: Scientific Reports
volume: '16'
pages: '28639'
doi: 10.1038/s41598-026-59364-3
peer_review: peer-reviewed-journal
evidence_tier: B
design: 'Preregistered within-participant study: real-vs-synthetic classification of faces and of voices, each followed by confidence ratings; signal detection (d'', c) and type-II AUROC for metacognition'
modality:
- image-face
- audio
stimuli: 20 FFHQ real faces + 20 Adobe Firefly (diffusion) faces; 40 VCTK real speakers and ElevenLabs Instant Voice Cloning clones (each listener heard 20 real + 20 clones of other speakers)
n_participants: 172
n_note: 200 recruited on Prolific, 172 analysed; ages 18-30
topics:
- human-detection
- voice
- voice-clone
- confidence
- individual-differences
summary: Faces were detected above chance (65.48%) but cloned voices barely (53.95%); face and voice skill were not significantly related, and confidence tracked accuracy for faces but not voices.
results:
- metric: Face accuracy (diffusion faces)
  value: 65.48% [95% CI 63.75, 67.21]
  note: d' = .85
  source_location: null
- metric: Voice-clone accuracy
  value: 53.95% [95% CI 52.64, 55.27]
  note: d' = .20; 35.47% of participants at or below chance (faces 12.80%)
  source_location: null
- metric: Face d' predicting voice d'
  value: beta = .21, p = .053, R² = .02
  note: not significant; no evidence of a domain-general ability
  source_location: null
- metric: Response criterion c
  value: faces -.24; voices -.07
  note: both biased toward 'real'
  source_location: null
- metric: Metacognitive AUROC
  value: faces .59 (above chance); voices .48 (below chance)
  note: confidence informative for faces only
  source_location: null
moderators:
- Skin smoothness was used as an AI cue for faces and predicted accuracy; breathing predicted voice accuracy
limitations:
- Task difficulty not matched across modalities
- Static faces and short clips only, single play
- Participants warned that stimuli could be synthetic
- Young sample (18-30)
verification:
  status: verified-fulltext
  confidence: high
  checked: 2026-10-12
  via: nature.com full text, doi:10.1038/s41598-026-59364-3
url: null
pmid: null
pmcid: null
arxiv_id: null
openalex_id: null
oa_url: null
cited_in:
- MR-0001
- MR-0003
updated: '2026-10-12'
---

# Human detection of AI-generated faces and voices is not domain-general

> **Source record** `ivory-2026-faces-voices-domain` · 2026 · Scientific Reports  
> **Evidence tier:** B (peer-reviewed-journal) · **Verification:** verified-fulltext  
> **Bottom line:** Faces were detected above chance (65.48%) but cloned voices barely (53.95%); face and voice skill were not significantly related, and confidence tracked accuracy for faces but not voices.

## Citation {#citation}

Matthew R. Ivory, Carolyn McGettigan, Helen E. Nuttall, Sophie J. Nightingale (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

Verification entry (identifiers, resolvers, claims with source locations): [/data/citations.json#ivory-2026-faces-voices-domain](/data/citations.json#ivory-2026-faces-voices-domain)

## Study design {#design}

| Field | Value |
|---|---|
| Design | Preregistered within-participant study: real-vs-synthetic classification of faces and of voices, each followed by confidence ratings; signal detection (d', c) and type-II AUROC for metacognition |
| Modality | image-face, audio |
| Stimuli | 20 FFHQ real faces + 20 Adobe Firefly (diffusion) faces; 40 VCTK real speakers and ElevenLabs Instant Voice Cloning clones (each listener heard 20 real + 20 clones of other speakers) |
| Participants (N) | 172 |
| N detail | 200 recruited on Prolific, 172 analysed; ages 18-30 |
| Topics | human-detection, voice, voice-clone, confidence, individual-differences |

## Results {#results}

| Metric | Value | Note |
|---|---|---|
| Face accuracy (diffusion faces) | 65.48% [95% CI 63.75, 67.21] | d' = .85 |
| Voice-clone accuracy | 53.95% [95% CI 52.64, 55.27] | d' = .20; 35.47% of participants at or below chance (faces 12.80%) |
| Face d' predicting voice d' | beta = .21, p = .053, R² = .02 | not significant; no evidence of a domain-general ability |
| Response criterion c | faces -.24; voices -.07 | both biased toward 'real' |
| Metacognitive AUROC | faces .59 (above chance); voices .48 (below chance) | confidence informative for faces only |

## Moderators {#moderators}

- Skin smoothness was used as an AI cue for faces and predicted accuracy; breathing predicted voice accuracy

## Limitations {#limitations}

- Task difficulty not matched across modalities
- Static faces and short clips only, single play
- Participants warned that stimuli could be synthetic
- Young sample (18-30)

## Cited in {#cited-in}

- [MR-0001](/metareports/human-detection-accuracy.md): Human Detection of Deepfakes: Accuracy, Bias, Confidence, and Training
- [MR-0003](/metareports/voice-cloning-human-detection.md): Voice Clones and Synthetic Speech: How Well Do People Detect Them?

## Verification {#verification}

| Field | Value |
|---|---|
| Status | `verified-fulltext` |
| Compiler confidence | high |
| Checked on | 2026-10-12 |
| Via | nature.com full text, doi:10.1038/s41598-026-59364-3 |