AI Photos Still Look Fake Because They Can't Get Light Right
Bad lighting is the giveaway that exposes a fake image — and AI still flunks it.
A team of researchers has published a new way to grade AI image generators — not on how pretty their pictures look, but on whether the light in them makes physical sense. Their test, called "Shedding Light," takes real photographs containing simple objects that act like light probes (think of a white ball that shows you exactly where the sun is). They then ask an AI to add that same object somewhere else in the photo, filling in the missing part of the picture. Finally, they measure the shadow and glow coming off the AI's version and compare it with the real thing.
The findings are humbling for an industry that loves to boast about realism. Even the best image generators often put shadows in the wrong direction, choose the wrong colour of light, or make objects glow when they should be dim. In other words, these systems are very good at copying the look of photographs, but they do not appear to reason about how light actually behaves. A sunlit kitchen from the left should light an apple from the left — and frequently it doesn't.
So what does that mean for you? Lighting is the single biggest tell that something is fake. It's how you catch a doctored photo, a suspicious online listing, or a deepfake video. If AI masters it, that tell disappears — great news for advertisers, movie effects artists and anyone selling products online, less great for anyone trying to spot a fake. And when small businesses can generate realistic product photos for pennies, the people who used to shoot and retouch them feel it.
The honest catch: this paper is a measuring stick, not a fix. It doesn't make any AI better at lighting, it just gives everyone a fair way to compare models going forward. It also tests simple objects in real photographs rather than busy, complicated scenes, so the failures it finds may understate how badly things break down in messier situations. Still, the work was accepted to SIGGRAPH Asia 2026, one of the field's biggest stages, and the code and data are free for anyone to use.
- Researchers created a public test that checks whether AI image generators get shadows, light direction and light colour physically right — not just whether the picture looks nice.
- The method is simple: ask AI to drop the same object into a real photo in a new spot, then compare its shadow to the real one.
- Results suggest today's image models copy the look of photos without truly understanding how light works — which is why AI pictures often feel slightly 'off'.
Why It Matters
Lighting is how you spot a fake photo or deepfake. If AI nails it, spotting fakes gets much harder.