The Arms Race We’re In

We owe you honesty. As a company that builds AI image detection tools, we have a financial incentive to overstate our capabilities and understate the difficulty of the problem. Every detection company does. The pitch is seductive: upload an image, get a verdict, know the truth. The reality is significantly more complicated, and we think you deserve to know how complicated.

Here is where detection stands in early 2026, as clearly as we can state it.

The best current detection systems—including fusion strategies that combine GPT-4o-level reasoning with specialized classifiers—achieve approximately 93.4% accuracy on benchmark datasets. That’s better than CNNSpot’s 91.8%, and meaningfully better than the best human annotators at 86.3%. These numbers sound good. They obscure important caveats.

First: accuracy varies dramatically by generation method. GAN-generated images—the older technology—are detected at 88 to 95 percent by advanced hybrid systems. Diffusion model outputs—which now dominate the landscape—are detected at 88 to 94 percent by the best systems and as low as 62 to 80 percent by standard CNN detectors. The gap between best-case and worst-case is enormous, and most users don’t know which case applies to the image they’re analyzing.

Second: benchmark accuracy and real-world accuracy are different things. Benchmark images are clean, well-formatted, and drawn from known distributions. Real-world images are screenshotted, compressed, filtered, cropped, partially edited, and shared across platforms that each apply their own processing. Every one of those transformations degrades the artifacts our models look for. An image that our system identifies correctly at 97% confidence in its original format might drop to 70% after being posted to Instagram, screenshotted on an Android phone, and shared via WhatsApp.

Third: the landscape is a moving target. When we train a model to detect outputs from DALL-E, Midjourney V6, and Stable Diffusion XL, that model works well for those generators. When Midjourney V7 ships with a “totally different architecture,” our model has never seen its outputs. When FLUX introduces inference speeds eight times faster with a twelve-billion-parameter model, the artifacts change. When Google’s Gemini 3 Pro Image launches, it’s a new fingerprint we have to learn.

This is the arms race, and it’s not a metaphor. Generation and detection are locked in a co-evolutionary dynamic where each advance on one side drives adaptation on the other. The generators are not trying to evade detection—but the optimization pressures that make images more realistic also, as a side effect, make them harder to detect. Realism and undetectability are not the same thing, but they’re correlated enough to keep us running.

The most promising advances in detection are coming from two directions.

The first is multimodal analysis. Systems like GADNet use Gram matrix-based attention to detect spatially uneven artifact distribution—essentially, they look for patterns in how artifacts are distributed across an image rather than looking for individual artifacts. This is more robust because it targets properties of the generation process rather than specific outputs of specific models.

The second is using large language models for detection. This sounds circular—using AI to detect AI—but the approach works differently than you’d expect. Multimodal LLMs analyze both visual and textual signals to provide not just a detection verdict but interpretive reasoning: here’s why this image appears to be AI-generated, here’s what specific patterns suggest synthetic origin. Explainability is as important as accuracy, because users need to understand and evaluate the detection result, not just accept it.

Now the numbers that keep us up at night. Over three billion images per month are being generated using diffusion-model platforms. Up to thirty-two percent of all images shared on major social platforms show evidence of partial or full AI augmentation. AI-generated content overtook human-made content in raw volume by November 2024 and reached fifty-two percent by May 2025. The ratio is still climbing.

Detection at our current accuracy rates means, practically speaking, that for every hundred AI-generated images we analyze, we correctly identify roughly ninety-three and miss roughly seven. Seven percent of three billion is two hundred and ten million images per month that could evade current detection. That’s not a rounding error. That’s a flood.

What do we do with this honesty? We don’t think it’s a reason for nihilism. Ninety-three percent accuracy is dramatically better than the fifty percent that humans achieve unaided. Detection tools are essential for journalists verifying sources, platforms moderating content, researchers studying misinformation, and individuals protecting themselves from fraud. The tool doesn’t have to be perfect to be valuable. A smoke detector that catches ninety-three percent of fires is a lot better than no smoke detector.

But we think it’s a reason for humility. When we give you a confidence score, we want you to treat it as what it is: a probability, not a verdict. An 85% confidence that an image is AI-generated means there’s roughly a one-in-seven chance we’re wrong. A 60% confidence means we’re barely better than a coin flip. The numbers mean something, and they mean something specific, and treating them as binary—real or fake—is a misuse of the tool.

The path forward is layered defense. Watermarking for images from cooperating generators. Provenance standards like C2PA for images within cooperating ecosystems. Forensic detection for everything else. Platform policies that require disclosure. Legislation that creates consequences for deceptive use. Media literacy education that gives people the cognitive tools to evaluate what they see.

No single layer is sufficient. All of them together might be. The arms race we’re in is not one we expect to win decisively. It’s one we expect to fight indefinitely, iteration by iteration, model by model, artifact by artifact. That’s not the inspiring pitch. But it’s the true one.