Here's a claim that'll get creatives talking: Google says it's proved, mathematically, that AI image generators actually invent new images, rather than just copying the photos and artwork they were trained on. It's a big claim, and one that goes right to the heart of the argument many artists have been having with the tech industry for the past few years.
The paper in question, snappily titled On the interpolation effect of score smoothing in diffusion models, was written by Google research scientist Zhengdao Chen and presented at ICLR 2026, a major AI research conference. It's dense, technical stuff, but the headline idea is simple enough to unpack.
What Google's saying
AI image generators work by starting with random noise, a screen of meaningless static, and gradually reshaping it into a picture. Something inside the model has to decide which direction to nudge that noise in at every step, and Google calls that thing the score function. Think of it a bit like an invisible hand, pulling scattered dots towards a recognisable shape.
If that invisible hand worked with total precision, Google argues, it would always pull the noise back into an exact copy of something in the training data. That would be memorisation, not creativity: a fancy photocopier, basically.
But neural networks don't work with total precision. The training process naturally blurs things slightly, a side effect known as score smoothing. That blur softens the model's grip on any single training image, so instead of snapping back to a photo it's seen before, the output can settle somewhere in between two or more of them. Google argues that's where new, plausible images come from: not copying, but interpolating between real examples.
Why artists aren't convinced
To be fair to Google, this isn't marketing BS. It's a clear, well-expressed mathematical account of why these models behave the way they do, and it pushes back hard on the idea that AI image tools are simply collage machines.
Ask people who make a living from creative work, though, and the use of the word "creativity" in this paper starts to look like it's doing a lot of unearned work.
Interpolating between data points is not the same thing as having something to say. A painter who spends years developing a style is drawing on lived experience, taste, doubt and choice. A model finding the mathematical midpoint between two training images is doing geometry, not self-expression.
What makes real art important
Van Gogh's swirling skies weren't the mathematical midpoint of two paintings he'd seen. They came from a specific, hard-won mix of Japanese woodblock prints, the emotional weight of his own mental health struggles, and years of technical experimentation with colour and brushwork.
Monet didn't paint Rouen Cathedral 30 times because he was interpolating between data points. He was chasing how a single subject changed under shifting light, at different hours, in different weather, because that specific question mattered to him. In both cases, the "novelty" came bundled with a reason; a lived problem the artist was trying to solve.
A diffusion model settling into the gap between two training images has no equivalent problem it's trying to work through. It has a slightly blurred force field, and nothing resembling a stake in the outcome.
There's also the small matter of where the training data comes from in the first place. Explaining that a model doesn't copy any single image exactly does little to settle the argument over whether it should have been trained on millions of artists' work without permission or payment in the first place. Interpolation might be mathematically real, but so is the original source material it's interpolating between.
Yes, Google's maths looks solid, and it's a useful corrective to the idea that these tools are just stitching together stolen fragments. But solving how a model generates something new is a very different problem to proving that what it's doing counts as creativity… at least, in any sense a working artist would recognise. Or to put it another way: your numbers may be right, Google, but I'm not so sure about the words that go with them.
