Training human super-recognizers' detection and discrimination of AI-generated faces

Area of Science:

• Computer Vision
• Psychology
• Artificial Intelligence

Background:

• Generative Adversarial Networks (GANs) produce highly realistic synthetic faces, posing challenges for detection.
• Existing training programs for synthetic face detection show mixed results.

Purpose of the Study:

• To investigate synthetic face detection and discrimination abilities in super-recognizers (SRs) compared to typical controls.
• To evaluate the effectiveness of a training procedure designed to highlight rendering artifacts.

Main Methods:

• Two experimental designs were used with a total of 283 SRs and 381 control participants.
• Participants performed synthetic face detection and discrimination tasks, with and without a specific training intervention.

Main Results:

• SRs demonstrated superior performance in detecting and discriminating synthetic faces compared to control participants.
• Control participants performed below chance without training, but improved significantly after training.
• Training enhanced performance in both SRs and controls, with a similar magnitude of improvement across groups.

Conclusions:

• SRs appear to utilize cues beyond rendering artifacts for synthetic face detection.
• A straightforward training procedure can significantly boost synthetic face detection accuracy in both SRs and typical individuals.
• Trained SRs offer potential for real-world applications in synthetic media forensics.