If you told somebody in 2020 what AI image generation would look like in 2026, they would never believe you. We went from black-and-white, low-resolution blobs that barely resembled cows to images that stop you mid-scroll. Twelve years of progress compressed into something you can hold up side by side in a single tweet.
With Mythos dropping and GPT-5.5 on the horizon, it feels worth stepping back and looking at the visual side of where we are, because that is where something interesting is happening that most people are reading completely wrong.
What People Think They Know About AI Images
Most people think they can spot an AI image. They probably can. But they are not spotting AI. They are spotting ChatGPT. Those are not the same thing, and conflating them leads to a false sense of confidence that is going to become a real problem as these models get better.
GPT-generated images have a distinct look. Oversaturated colors, a kind of glossy perfection, hues that pop in a way that real photography almost never does. If you have seen enough of them, your brain flags it immediately. That recognition is real, but it is model-specific, not a universal AI fingerprint. The general assumption is that photorealism is the goal and that the closer something gets to photorealism, the harder it is to detect. That assumption is wrong in at least one direction, and OpenAI is the reason why.
OpenAI: Style Wins Leaderboards
Being indistinguishable from reality is not what people actually prefer when you give them a choice. On image preference leaderboards, where people compare two outputs and pick a winner, and in OpenAI’s own A/B testing, the thing that consistently wins is not photorealism. It is good style defaults. Images that are more saturated, more eye-catching, more visually striking.
OpenAI optimized for preference, not realism. That is a deliberate product decision. The result is that ChatGPT images are immediately identifiable as AI, often on first glance, but people like them more. They are designed to win the comparison, not to disappear into the world. There is real logic to this from a product standpoint. Most people using image generation want something that looks good for a post, a presentation, or a thumbnail. They are not trying to fool anyone. They want something that stands out. OpenAI read that correctly and built toward it.
Google: Raw Power, Actual Realism
Google went the other direction. If you use Nano Banana 2 or Nano Banana Pro with a solid prompt, you can produce results that are genuinely indistinguishable from reality to an untrained eye. No telltale saturation, no gloss, no obvious AI sheen. That is impressive. It is also less immediately attention-grabbing, which is not a flaw in the model. It is a flaw in the use case if what you want is a thumbnail that stops a scroll.
These images do not catch people the same way an oversaturated AI image does, but that is because they were not optimized for that. They were optimized for raw fidelity. Google did not lose the preference war by accident. They were playing a different game entirely. If you need an image to pass as real, Nano Banana Pro is the more capable tool. If you need an image to win a side-by-side comparison on a social feed, OpenAI’s defaults are probably going to outperform it.
How You Would Actually Detect These
For Google’s models, SynthID is the primary detection mechanism. It is a watermarking system embedded at generation time, and without it, you are left looking for artifacts. That requires a trained eye and close inspection: hands, background text, light source consistency, edge detail. The mistakes are there if you know where to look, but the average person is not looking, and even a trained eye can miss them on a well-generated output.
For OpenAI’s outputs, no technical inspection is necessary. The style is the tell. That is a product outcome, not a safety feature, and it is worth understanding the difference. OpenAI did not accidentally make their images look like AI. They made them look like AI because that is what tested well.
The important takeaway here is that public perception of what AI images look
… (truncated for brevity) …