Researchers gave AI models some colored pencils to copy the Mona Lisa. They kept making it worse

This article was featured in the July 24, 2026, edition of the IBM Think newsletter. Get it in your inbox.

Have you ever revised something to death? Sent off the fourth draft (probably named “Project_final_FINAL_ACTUALLY.doc”) before realizing the second one was probably better? Congratulations, you now have something in common with a frontier model that just took 27 cracks at drawing the Mona Lisa.

In an experiment released this week by startup TryAI, researchers gave GPT-5.6 Sol, Claude Fable 5, Grok 4.5 and Gemini 3.6 Flash a drawing toolset with all the shortcuts removed—all they could do was pick a color, set pressure, lay down strokes, smudge and erase—and had them reproduce da Vinci’s Mona Lisa and van Gogh’s Starry Night. Five additional tests involved written prompts with no reference image, things like “A tabby cat curled asleep on a windowsill in the afternoon sun.”

The researchers judged the results two ways: with an objective structural-similarity score against the target, and their own human eyeball test. (The full harness is open source on GitHub for anyone who wants to run their own models through it.) At any point, a model could stop, inspect its canvas, compare it to the target and decide what to fix.

And boy, do these models like to check their own work.

First though: Some drawings were surprisingly good! Others had that cute quality you’d find in a sixth-grade art project stuck on a refrigerator door. And a few—well, remember that botched restoration of Spain’s Ecce Homo fresco that launched a thousand memes? Think that.

In all eight scored runs the final drawings rated below their mid-run peaks. Gemini 3.6 Flash reviewed its Mona Lisa 22 times, hit its best score around attempt five, then spent the next 17 reviews making it worse. Claude Fable 5 checked its work 27 times on the same painting; its score was essentially flat after the fifth look.

“More reviewing did not translate into a better finish,” the researchers wrote. “The models kept editing past their own best performance.”

The models also differed on quality and cost. GPT-5.6 Sol produced the best output, at the lowest price. Claude Fable 5 cost a whopping USD 160.58 across seven drawings, about 20 times more than any competitor, for results that came in second. Meanwhile, Grok 4.5 made 1,349 tool calls across its runs, 65% of them setup adjustments; the researchers described its final pieces as “basically garbage.”

TryAI’s experiment is fun, and a little weird, and in the end not really about art. Self-critique loops are a core feature of today’s AI agents, based on the idea that stopping to check your work should improve the final result. After watching four frontier models repeatedly paint over their own best sketches, maybe we should also be asking how often they do the same thing on tasks where we can’t see every revision.