There was a moment, sometime around 2022, when spotting AI art felt almost too easy. Six fingers. Melting wrists. Eyes like cracked marbles. Midjourney V3 got standard hand poses right about 40% of the time. The discourse settled into a comfortable rhythm: AI is impressive but broken, and broken in ways any human can catch.
That comfort is gone. Midjourney V7, released in April 2025, hits roughly 85 to 90% accuracy on standard hand poses. Text in images is readable. Faces no longer drift into soft horror. The tools that a vocal slice of the internet used to dismiss AI art wholesale have been, if not eliminated, dramatically blunted. Community testing and the V7 documentation both confirm this, and anyone who's used the tool in the past six months has felt it firsthand.
The New Tells Are Subtle, and That's the Point
Here's what researchers at Northwestern's Kellogg School have started cataloguing as the actual failure modes now: lighting that doesn't match between a subject and their background, backpack straps that merge implausibly into a shirt, shadows that fall from the wrong angle or from nowhere at all. They call these 'functional implausibilities' and 'violations of physics.' The name is clinical, but the experience is that uncanny valley feeling you can't quite name. Something is wrong and you can't say what.
This matters because it changes what kind of attention you need to catch it. Counting fingers is a fast, conscious check. Noticing that the light on a person's face couldn't possibly come from the window behind them requires a different kind of literacy, one most people aren't trained to apply casually. And the detection numbers bear this out brutally. A Conjointly survey from September 2025 found consumers correctly identify AI images only 52% of the time, essentially a coin flip, down from 56% in June 2023. FLUX.1-dev, one of the newer models, was detected at only 29% in compiled study data, meaning viewers were actively, confidently wrong, labeling AI images as real photographs.
What We Actually Owe This Moment
The old critique of AI art wasn't wrong, it was just temporary. The people who said 'this technology has real limits' were right, for a window of time that has now largely closed. What would be wrong is pretending the critique still lands the same way, or that the same quick glance test still works.
The case for skepticism hasn't weakened, it's just gotten harder to act on. You can still find the seams, but you have to look for physics now instead of anatomy. You have to ask whether the shadows agree with each other, whether the objects in a scene interact like objects actually do. That's genuinely more demanding, and it demands more of everyone: consumers, editors, platforms, and anyone sharing images as documentation of reality.
The fingers were a gift, honestly. An obvious, democratizing tell that required no expertise. What replaced them requires expertise. That asymmetry should make us more cautious, not less.
