Do surgical masks make it harder to tell if two faces match?
Putting a surgical mask on one or both faces made people much worse at judging whether two photos showed the same person, for famous and unfamiliar faces alike.
Source
Surgical face masks impair human face matching performance for familiar and unfamiliar faces
Study at a glance
- Design
- Human experiment — Online between-subjects experiment: participants randomly assigned to no-mask, one-face-masked or both-faces-masked conditions completed two simultaneous face-matching tasks; a deep neural network did the same tasks.
- N
- N=138 · 201 UK adults recruited via Prolific; 138 remained after pre-registered exclusions (53 control, 43 mixed, 42 masked).
- Population
- UK adult online participants
- Outcome
- Face-matching sensitivity (d') and response bias (c) on the Glasgow Face Matching Test and a new famous/unfamiliar face matching task
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
138 UK adults judged whether pairs of photos showed the same person or two different people. They were randomly assigned to see no masks, a mask on one face in each pair, or masks on both faces, with masks digitally added to the same photos. Everyone did a standard unfamiliar-face test and a new task mixing celebrities and unknown models, followed by a recognition test to check which celebrities they actually knew. A face-recognition deep neural network was tested on the same pairs.
What they found
Masks lowered matching sensitivity with large effects (d between 1.02 and 1.21 versus no masks), and it did not matter whether one or both faces were masked. The drop was about as large for familiar as for unfamiliar faces, though masked familiar faces were still matched better than unmasked unfamiliar ones. Masks made people biased toward saying 'different' for unfamiliar faces and 'same' for familiar faces. The neural network stayed near ceiling except for familiar-task match pairs where only one face was masked.
The limits
What it doesn't show
The masks were photo-edited onto images, so real worn masks might preserve some face-shape information. The design was between-subjects, so it cannot show how the same person's performance changes with masks or identify individual differences. The study documents the impairment but does not test why it happens, for example missing features versus disrupted holistic processing. Results for the network may not generalise, as two of three commercial systems in the supplement were impaired.
Key terms
- Face matching
- Deciding whether two simultaneously shown face images depict the same person, a perceptual rather than memory task.
- d' (d-prime)
- A signal-detection measure of sensitivity: how well someone separates true matches from mismatches, independent of bias.
- Response bias (criterion c)
- A general tendency to answer 'same' (liberal) or 'different' (conservative) regardless of the evidence.
- Holistic processing
- Perceiving a face as an integrated whole, including the spatial relations among its features.
- Pre-registration
- Publicly recording hypotheses, design and analyses before data collection to reduce flexible analysis.
Flashcards
0 of 11 answers reviewed
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
What task did participants perform?
Common questions
Is it worse if only one person in the pair is masked?
Not for human sensitivity: the mixed and fully masked conditions were equally impaired compared with no masks.
Aren't familiar faces immune to this kind of disruption?
Surprisingly no. Masks reduced matching of familiar faces by a similar amount, although familiar faces remained easier overall.
Did the computer do better than people?
The research network matched or beat average human accuracy in most conditions, but it struggled when one face was masked in the celebrity task, and some commercial systems failed badly.
More on Face processing