Faces instead of distorted letters: people were correct about 94% of the time
In 42,000 responses in a 2014 study, people were correct about 94%, but blending faces heavily with the background reduced human accuracy.
The key idea
A test designed to stop bots has two sides: how much it stops them and how difficult it is for people.
The study, in plain language
The challenge showed a collage of distorted faces and asked people to choose two images of the same person. Researchers varied rotation, noise, and how much faces blended with the background, aiming to make machines fail without making the task impossible for people.
Across two sessions, they collected 42,000 responses. Accuracy differed substantially between sessions: about 98% in one and about 85% in the other, with a weighted average around 94%. During calibration, heavy blending of faces with the background reduced human accuracy. In a comparison with two other designs, most people chose this version. A tension remained: make the task difficult for programs without making it difficult for people.
What they found
- The test asked people to match two images of the same person instead of transcribing distorted letters.
- Heavily blending faces into the background reduced human accuracy and revealed the cost of making the task harder.
- In a comparison, this version was preferred over other designs.
How they studied it
Researchers designed a face-matching challenge with distortions and adjusted its parameters. They then compared people’s accuracy, design preferences, and performance of facial-recognition algorithms of that era.
What the study cannot show
- The algorithms evaluated came from a 2014 study. The result does not establish resistance to current tools or accessibility for everyone.
Something to think about
When verification is difficult, there are two different questions: how much it stops bots and how much it burdens a person. Both can be measured.
The original study
FR-CAPTCHA: CAPTCHA Based on Recognizing Human Faces
Source under CC BY 4.0. English adaptation and explanation by PaperFlow, with AI assistance. The text has been reorganized for a general audience; it is not the complete paper and does not imply endorsement by its authors.
One study alone does not represent a field's consensus. Educational content does not replace professional advice. Report a possible error.
