Most people have adjusted a crooked image until it looks straight. Some websites now use a circular puzzle instead of the usual “I am not a robot” checkbox. A small slice of a photo must be rotated to fit its outer ring. The task is simple but frustrating.
Two researchers at Bern University of Applied Sciences created a script that solves these puzzles in 0.006 seconds. Their tool relies on circle-detection algorithms from the 1970s and signal-matching methods older than today’s AI security tools. It passed all ten test puzzles without error, completing each faster than a human blink.
The AI models fell short
The same puzzles stumped expensive AI models. Gemini 3.1 Pro took 67 seconds and solved seven out of ten correctly. GPT-4o and Grok each managed only one right answer. With only eight possible positions, random guessing would have performed better.
The real surprise emerged when the models received a working solution. The researchers allowed them to use the fast script as a helper, feeding them the correct answer. GPT-4o and Grok accepted it and scored perfectly. Gemini rejected the help.
One puzzle aligned best at 129 degrees, but the site only accepted answers in 45-degree steps, so the script suggested 135. Gemini ignored the suggestion, deciding the alignment still looked wrong. That choice cost it 20% of its score.
The issue wasn’t the AI’s ability to see. It was the tendency to second-guess a tool that was already right. If a language model is supposed to oversee a working system, this creates a risk: the model might override the correct answer because it assumes it knows better.
Related: Vendor’s PQC Comments Reveal Their Readiness
The models recognized the answer but couldn’t act on it
The same pattern appeared in other tests. A puzzle displayed colored rings and asked users to click the one with a gap. Most AI models failed. When forced to explain, they described the correct ring—”a cyan ring in the lower central area”—without clicking it accurately.
Another test showed a grid of nine squares, asking users to select the ones with insects. Grok identified the insects but mislabeled the squares, calling the middle-left square “5” when it was “4.” At one point, it even answered “square 10” in a grid that only had nine. When the researchers turned on its reasoning mode—without changing the image—accuracy improved sharply.
These puzzles exist because the websites using them can’t rely on JavaScript. Many users disable it to avoid tracking, removing the behavioral signals modern CAPTCHAs depend on. Without those, the challenge becomes a basic shape-matching problem—one computer vision solved decades ago on hardware slower than a modern phone.
The security measure protecting these sites is a geometry exercise a laptop can solve 160 times per second. The real weakness isn’t the puzzle. It’s the assumption that AI will always recognize when it’s wrong.
In practice, the models often see the right answer but refuse to trust it. That hesitation isn’t a vision problem. It’s a supervision problem. If the goal is to catch errors, that’s the part that needs attention.
