---
type: source
id: pehlivanoglu-2026-machines-humans
title: Is this real? Susceptibility to deepfakes in machines and humans
authors:
- Didem Pehlivanoglu
- Mengdi Zhu
- Jialong Zhen
- Aude A. Gagnon-Roberge
- Rebecca K. Kern
- Damon Woodard
- Brian S. Cahill
- Natalie C. Ebner
year: 2026
venue: 'Cognitive Research: Principles and Implications'
volume: '11'
pages: '3'
doi: 10.1186/s41235-025-00700-y
peer_review: peer-reviewed-journal
evidence_tier: B
design: 'Study 1: humans rated 200 faces 1 (fake) to 10 (real) plus individual-difference measures; compared with ResNet-50 CNN and frequency-domain SVM. Study 2: same comparison for deepfake videos'
modality:
- image-face
- video
stimuli: 'Study 1: 300 FFHQ real faces and 300 artefact-screened StyleGAN2 faces (sets of 200 per participant). Study 2: deepfake videos (details not read)'
n_participants: 2203
n_note: Study 1 N=2,203 undergraduates (age 18-58, M 19.64); Study 2 N not read
topics:
- human-detection
- human-vs-machine
- truth-bias
- confidence
- individual-differences
summary: 'On StyleGAN2 faces humans were at chance (49%, AUC 0.53) with a strong truth bias while a CNN scored 97%; on videos the pattern reversed: machines near chance, humans better.'
results:
- metric: Study 1 human accuracy (faces)
  value: 49%; mean AUC 0.53
  note: N=2,203
  source_location: null
- metric: Study 1 deepfakes judged real
  value: 69%
  note: 'real judged real 67%: truth bias'
  source_location: null
- metric: Study 1 confidence
  value: real 6.77 vs deepfake 4.11 (1-10 realness scale)
  note: ''
  source_location: null
- metric: Study 1 machine accuracy
  value: CNN 97%; frequency-domain SVM 79%
  note: ''
  source_location: null
- metric: Study 2 (videos)
  value: machines near chance; humans outperformed machines
  note: abstract only; values not read
  source_location: null
moderators:
- 'Study 1: no individual-difference measure predicted face discrimination'
- 'Study 2 (abstract): higher analytical thinking, lower positive affect and greater internet skills predicted better video detection'
limitations:
- Undergraduate sample, 75% female
- Single GAN for faces
- Online, uncontrolled viewing
- Study 2 values not checked for this record
verification:
  status: verified-fulltext
  confidence: medium
  checked: 2026-10-12
  via: link.springer.com full text (Study 1 methods and results); Study 2 from abstract only
url: null
pmid: null
pmcid: null
arxiv_id: null
openalex_id: null
oa_url: null
cited_in:
- MR-0001
updated: '2026-10-12'
---

# Is this real? Susceptibility to deepfakes in machines and humans

> **Source record** `pehlivanoglu-2026-machines-humans` · 2026 · Cognitive Research: Principles and Implications  
> **Evidence tier:** B (peer-reviewed-journal) · **Verification:** verified-fulltext  
> **Bottom line:** On StyleGAN2 faces humans were at chance (49%, AUC 0.53) with a strong truth bias while a CNN scored 97%; on videos the pattern reversed: machines near chance, humans better.

## Citation {#citation}

Didem Pehlivanoglu, Mengdi Zhu, Jialong Zhen, Aude A. Gagnon-Roberge, Rebecca K. Kern, Damon Woodard, Brian S. Cahill, Natalie C. Ebner (2026). Is this real? Susceptibility to deepfakes in machines and humans. *Cognitive Research: Principles and Implications*, 11, 3. https://doi.org/10.1186/s41235-025-00700-y

Verification entry (identifiers, resolvers, claims with source locations): [/data/citations.json#pehlivanoglu-2026-machines-humans](/data/citations.json#pehlivanoglu-2026-machines-humans)

## Study design {#design}

| Field | Value |
|---|---|
| Design | Study 1: humans rated 200 faces 1 (fake) to 10 (real) plus individual-difference measures; compared with ResNet-50 CNN and frequency-domain SVM. Study 2: same comparison for deepfake videos |
| Modality | image-face, video |
| Stimuli | Study 1: 300 FFHQ real faces and 300 artefact-screened StyleGAN2 faces (sets of 200 per participant). Study 2: deepfake videos (details not read) |
| Participants (N) | 2203 |
| N detail | Study 1 N=2,203 undergraduates (age 18-58, M 19.64); Study 2 N not read |
| Topics | human-detection, human-vs-machine, truth-bias, confidence, individual-differences |

## Results {#results}

| Metric | Value | Note |
|---|---|---|
| Study 1 human accuracy (faces) | 49%; mean AUC 0.53 | N=2,203 |
| Study 1 deepfakes judged real | 69% | real judged real 67%: truth bias |
| Study 1 confidence | real 6.77 vs deepfake 4.11 (1-10 realness scale) |  |
| Study 1 machine accuracy | CNN 97%; frequency-domain SVM 79% |  |
| Study 2 (videos) | machines near chance; humans outperformed machines | abstract only; values not read |

## Moderators {#moderators}

- Study 1: no individual-difference measure predicted face discrimination
- Study 2 (abstract): higher analytical thinking, lower positive affect and greater internet skills predicted better video detection

## Limitations {#limitations}

- Undergraduate sample, 75% female
- Single GAN for faces
- Online, uncontrolled viewing
- Study 2 values not checked for this record

## 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-fulltext` |
| Compiler confidence | medium |
| Checked on | 2026-10-12 |
| Via | link.springer.com full text (Study 1 methods and results); Study 2 from abstract only |