You’re Trying to Spot AI-Generated Faces Wrong

For a while, AI-generated faces were easy to dismiss on sight. Distorted teeth, asymmetrical features, and nonsensical backgrounds gave them away almost immediately. That window has closed. The generative models behind synthetic portraits have improved to the point where the old checklist of visual artifacts is no longer a reliable guide, and leaning on it may actually make people worse at detecting fakes.

The core problem is that most people were trained - informally, through social media posts and word of mouth - to look for specific flaws that early diffusion models and GANs reliably produced. Those flaws were not fundamental to how AI generates images; they were limitations of particular model versions at a particular moment in time. As models iterated, the artifacts faded, but the detection habits didn't update alongside them.

What researchers and detection specialists now suggest is a shift away from hunting for obvious deformities and toward subtler, more systemic cues. These include things like lighting consistency across the entire image, the way fine textures behave at high zoom, and statistical patterns in how pixels are distributed - none of which are intuitive to the untrained eye. Human perception is poorly suited to this kind of analysis, which is part of why automated detection tools, despite their own limitations, are increasingly being used alongside human judgment rather than being replaced by it.

The broader implication is that public confidence in the ability to spot AI-generated faces by eye may be outpacing actual ability. As synthetic media becomes more common in advertising, social profiles, and news contexts, the gap between perceived and real detection skill carries real consequences. The more useful approach, according to current thinking, is a combination of skepticism about provenance, use of metadata and content credentials where available, and an awareness that visual inspection alone is no longer sufficient.