AI's Legal Image Dilemma: Trust in Visual Evidence Takes a Hit

Visual evidence in courtrooms used to be rock-solid. Not anymore. AI advancements are throwing that reliability out the window. Humans and the latest multimodal large language models (MLLMs) are struggling big time to tell real photos from AI-generated doppelgangers, especially in civil cases.

The SLED-1400 Experiment

Meet SLED-1400, a dataset crafted with 200 genuine evidence images and 1,200 synthetic ones made by six top text-to-image AI generators across ten categories. In an experiment, 136 regular folks and four leading MLLMs, including GPT-5.1 and Gemini-3-Pro, were put to the test. The results? Humans nailed it only 64.8% of the time overall, which is basically a coin toss when facing the big guns like Gemini-3-Pro-Image and Flux-2-Max.

MLLMs, on the other hand, didn't mess up real images at all, scoring a perfect 100% specificity. Sounds great until you hear they only detected 5.9% of synthetic outputs from Gemini-3-Pro-Image. That's barely scraping by.

Why This Matters

So what does this mean for the justice system? If neither humans nor machines are reliable enough on their own, is any visual evidence truly safe from doubt? It seems visual evidence in courtrooms isn't the slam dunk it once was. It's time to re-evaluate how we treat these images.

We need a hybrid approach. Trained human reviewers should team up with MLLMs, backed by solid provenance infrastructure like C2PA Content Credentials. Relying on just one method is like building a sandcastle right before high tide.

The Big Question

Here's the kicker: How long until these AI tools get a seat at the table in legal processes? It's not about if, it's about when. With a system that's clearly flawed, can we afford to wait?

, Solana doesn't wait for permission. Neither should the legal system when faced with innovations in AI.