A new paper on arXiv proposes a taxonomy-guided study of human performance in detecting deepfake face images. The authors note that generative models for deepfakes have evolved rapidly, moving from GAN-based architectures to diffusion models and multimodal large language models (MLLMs). Despite this rapid progress, the paper highlights a gap: empirical research on how well humans can actually detect these synthetic faces has not kept pace.
The study's approach is to use a taxonomy to structure the existing literature, likely categorizing both generation methods and human detection tasks. The abstract indicates a focus on human performance, but the full findings are not available in the provided text. The paper appears to be a review or meta-analysis rather than a new experiment, aimed at identifying what is known and what remains unclear.
As a single source, this paper offers a framework for understanding the current state of human deepfake detection. It does not provide specific accuracy numbers or conclusions in the abstract, so readers should look to the full text for detailed results. The key takeaway is that the field is shifting toward more advanced generators, and human detection research needs to follow suit.