Attention
Real or AI Face Task
Also called: AI-generated face detection, synthetic face discrimination, AI hyperrealism
Measure how people distinguish photographs from AI-generated faces, with sensitivity, accuracy and judgments of each kind.
4-6 minUntimed
Free to start: 10 free participants per account, then $0.99 per participant for academic research, or any yearly plan. Try it here first, with no account.
- Papers, 2021 to 2025
- 255
- Evidence grade
- C
- Registered clinical studies
- 0
What it measures
Measures how people distinguish photographs from AI-generated faces. Participants judge each face with no time limit. Scores include sensitivity, accuracy and how often each kind of face is judged real.
Constructs
- Synthetic face discrimination
- Face perception
- Reality judgments
Research uses
- Research on perception of AI-generated faces
- Research on synthetic media and misinformation
How the paradigm works
AI-generated face detection measures how people judge whether a face belongs to a real person or was made by an image model. Miller and colleagues found that some AI faces were judged human more often than actual human faces. This task records that pattern as well as correct discrimination.
More about the paradigm
Participants view 48 pictures by default, half London Set photographs and half faces generated locally with Stable Diffusion 1.5. The short setting has 24. Each picture appears once in a seeded order, with counterbalanced answer positions. Four practice pictures give feedback; the main trials do not. At the end, participants are told which pictures were generated.
The primary score is d prime, with AI as the signal and log-linear corrected rates. Accuracy, response criterion, rates of real judgments for each source and their difference describe the result. Responses under 100 ms are counted as anticipations and excluded from response-time summaries.
This is a research version with new images, grayscale crops and no confidence ratings. Its scores describe these pictures and this model. Norms are off; use the raw measures for within-study and group comparisons.
Scores
- d primed'
- Sensitivity to AI-generated faces: corrected AI hit rate minus corrected false-alarm rate on the z scale.
- Informational
- Accuracy%
- Percent of answered pictures judged correctly.
- Informational
- Photograph accuracy%
- Percent of photographs judged real.
- Informational
- AI-face accuracy%
- Percent of generated pictures judged AI.
- Informational
- Photographs judged real%
- Percent of photographs called real.
- Informational
- AI faces judged real%
- Percent of generated pictures called real.
- Informational
- Difference in real judgments%
- Difference in percentage points: AI faces judged real minus photographs judged real. Positive means more AI faces were judged real.
- Informational
- Response criterionz
- Negative values indicate a tendency to answer Made by AI.
- Informational
- Median response timems
- Median response time, excluding responses under 100 ms.
- Informational
- Faces answeredcount
- Judgments with a recorded answer and a known picture.
- Quality check
- Planned facescount
- 24 or 48 test pictures.
- Informational
- Anticipationscount
- Presses under 100 ms, including those the task ignores.
- Quality check
Use and evidence
255 papers from 2021 to 2025 name this task in their title or abstract, up 750% compared with 2016 to 2020. That ranks it 113 of 582 test families in our census. See the counts for every test.
Show the numbers
| Year | Papers |
|---|---|
| 2025 | 145 |
| 2024 | 59 |
| 2023 | 29 |
| 2022 | 13 |
| 2021 | 9 |
| 2020 | 7 |
| 2019 | 10 |
| 2018 | 5 |
| 2017 | 4 |
| 2016 | 4 |
| 2015 | 6 |
| 2014 | 2 |
| 2013 | 9 |
| 2012 | 2 |
| 2011 | 3 |
| 2010 | 16 |
| 2009 | 5 |
| 2008 | 9 |
| 2007 | 3 |
| 2006 | 5 |
| 2005 | 7 |
| 2004 | 6 |
| 2003 | 2 |
| 2002 | 4 |
| 2001 | 3 |
| 2000 | 1 |
Use grade C
100 to 299 papers from 2021 to 2025.
Evidence grade C
At least 2 test-retest papers, 30 reliability or validity papers or 3 registered clinical studies.
- Papers on test-retest reliability
- 1
- Papers on reliability or validity
- 62
- Registered clinical studies using it as an outcome
- 0
Measured on ConductCognition
Pooled across the studies that feed our platform norms, by device group. A figure needs 50 or more people.
No pooled figure yet.
Note on the counts: includes deepfake studies that use videos.
How the grades work
Use: A is 1,000 or more papers from 2021 to 2025, B 300 to 999, C 100 to 299, D 30 to 99 and E fewer than 30. Evidence: A means at least 25 papers report its test-retest reliability and at least 25 registered clinical studies use it as an outcome; B, at least 8 test-retest papers, and either 5 registered clinical studies or 150 reliability or validity papers; C, at least 2 test-retest papers, 30 reliability or validity papers or 3 registered clinical studies; D, little published measurement evidence found. The counts measure use. They include every paper that names the task in its title or abstract. Grades describe the published research on this task, not ConductCognition's version of it.
Source: ConductCognition test census 0.39.0, counted 8 November 2026 from OpenAlex and ClinicalTrials.gov.
Task design
The trial sequence as the test runs in ConductCognition.
- Faces
- 24 photographs and 24 AI-generated faces, or 12 of each. Each picture appears once; the order comes from the session seed, with no more than four of one kind in a row.
- Pictures
- London Set photographs and fictional adult faces generated locally with Stable Diffusion 1.5. Both use grayscale oval crops, 280 by 320 pixels, on black. No source label appears during test trials.
- Answer
- Real photograph or Made by AI, by click, tap, or F and J, with no time limit. Left and right answers are counterbalanced across session seeds. A 400 ms blank follows each judgment.
- Practice
- Four separate pictures, two of each kind, with 1,500 ms feedback. A rest after 24 pictures in the 48-picture setting.
- Debrief
- The end screen gives the picture numbers that were generated and states that the others were photographs.
This version follows the real-or-AI judgment format with our own generated faces and London Set photographs, fewer trials and no confidence ratings. The scores describe judgments of this image set, including below-chance discrimination; norms are off. The photographs are credited to DeBruine and Jones under CC BY 4.0. Including practice, the photographs share people with these tests: Face Memory Test (12), Face-name Memory Task (10), Face Matching Task (26), Chimeric Faces Task (7), Part-Whole Face Task (13), Thatcher Illusion Task (8), Two-Tone Pictures Task (11), Verbal Overshadowing Task (17). Run face-memory tasks before this task or in a separate session, because seeing a face can affect later judgments. The Uncanny Valley Task shares 6 test people, one practice person, 6 generated test identities and one generated practice identity. Including practice pictures and foils, this task shares people with Face Impressions Task (26).
Quality checks
Each session is marked valid, needs review or invalid, with the reason, before you see its scores.
Scored rows: real_or_ai_face_test from the scored run. Below-chance accuracy and repeated answers are retained because misidentifying AI faces can be the research finding. Answers under 200 ms on at least half the faces invalidate. An incomplete run is flagged for review.
Why these rules
Miller et al. (2023) found below-chance discrimination with AI faces, so chance and response bias are outcomes, not exclusion rules. An answer under 200 ms is faster than a simple visual reaction, so it cannot reflect the item. Following response-time effort screening for untimed tests (Wise and Kong 2005), the session is invalid only when at least half the answers are that fast.
No percentile. Scores and quality flags only.
Running this test
Settings you can change
| Setting | Standard | Allowed |
|---|---|---|
| Faces of each kind | 24 (48 pictures, standard) | 24 (48 pictures, standard), 12 (24 pictures) |
Yearly plans can change settings within these limits. Free and pay-as-you-go studies use the standard settings.
Need a setting this test does not have?
Devices and languages
- Untimed test: runs on desktops, laptops, tablets and phones. Participant devices
- Languages: English, Spanish, French, German, Italian, Portuguese (Brazil), Chinese (Simplified) and Japanese.
- Available on: Every plan, including the free plan.
- No percentile. Scores and quality flags only. How norms work
- Recruiting online? See the setup guides for Prolific, Sona Systems, CloudResearch Connect and Qualtrics.
Cite and describe
Participants completed the Real or AI Face Task in a web browser (ConductCognition task version 1.0.0, scoring version 1.0.0; Miller et al., 2023), with the standard settings (faces of each kind: 24 (48 pictures, standard)). The main score was d prime (sensitivity to AI-generated faces: corrected AI hit rate minus corrected false-alarm rate on the z scale).
ConductCognition versions: task version 1.0.0, scoring version 1.0.0. Every export records the versions each session used, for a methods section.
- Miller, E. J., Steward, B. A., Witkower, Z., Sutherland, C. A. M., Krumhuber, E. G., & Dawel, A. (2023). AI Hyperrealism: Why AI Faces Are Perceived as More Real Than Human Ones. Psychological Science, 34(12), 1390-1403. https://doi.org/10.1177/09567976231207095Real-or-AI judgments, response bias and the possibility of below-chance discrimination.
- DeBruine, L., & Jones, B. (2017). Face Research Lab London Set [Data set]. figshare. https://doi.org/10.6084/m9.figshare.5047666.v5Real photographs, CC BY 4.0.
Source papers
Questions
How long does the Real or AI Face Task take?
About 4 to 6 minutes, as listed in the test library.
What device does a participant need?
Untimed test: runs on desktops, laptops, tablets and phones. The browser must be Chrome, Edge, Firefox or Safari.
What scores does it produce?
d prime, Accuracy, Photograph accuracy, AI-face accuracy, Photographs judged real, AI faces judged real, Difference in real judgments, Response criterion, Median response time, Faces answered, Planned faces and Anticipations. Each session is also marked valid, needs review or invalid, with the reason.
How widely is the Real or AI Face Task used in research?
255 papers from 2021 to 2025 name it in their title or abstract, up 750% compared with 2016 to 2020. That ranks it 113 of 582 test families in the ConductCognition test census (use grade C). Source: ConductCognition test census 0.39.0, counted 8 November 2026 from OpenAlex and ClinicalTrials.gov.
How much reliability evidence does the Real or AI Face Task have?
In the published research, 1 paper reports its test-retest reliability and 0 registered clinical studies use it as an outcome (evidence grade C in the ConductCognition test census).
Does it report a percentile?
No. Scores and quality flags only.
Which settings can a researcher change?
Faces of each kind (24 (48 pictures, standard), 12 (24 pictures)). Settings can be changed on yearly plans; free and pay-as-you-go studies use the standard settings.
Which languages can participants use?
Participants can take this test in English, Spanish, French, German, Italian, Portuguese (Brazil), Chinese (Simplified) and Japanese. English is on every plan, and the other languages come with yearly plans. Have a native speaker read the machine-translated text before launch.
Is this a clinical test?
No. ConductCognition is for research use only. It runs, scores and exports the test; clinical interpretation is outside the platform.
Every test: the full test list.
Run this test in your next study
Start free with 10 free participants on any test, then use participant credits or a yearly plan.
Related tests