The "I'm not a robot" checkbox was never asking you anything

The checkbox has spent a decade doing something other than what it says on the label. In July 2025, OpenAI's ChatGPT Agent made that plain: a bot ticked the box meant to stop bots. It hit a checkbox run by Cloudflare, the web infrastructure company, during a video conversion job. A Reddit user called logkn caught the moment, because the agent recounts what it does. The agent wrote: "This step is necessary to prove I'm not a bot and proceed with the action." Then it clicked the box and carried on.

People read that and assumed something had failed. Nothing failed. The box let through exactly the kind of visitor it was built to let through.

The story starts on 3 December 2014. Vinay Shet, product manager for Google's reCAPTCHA, announced the checkbox on Google's security blog. reCAPTCHA sorts humans from bots on websites. Until then it did that by making you read distorted text and type it out. Google's research had shown AI solving even the hardest distorted text with 99.8% accuracy. Shet's line: "Thus distorted text, on its own, is no longer a dependable test."

So Google substituted it with a question. The post says "we figured it would be easier to just directly ask our users whether or not they are robots". A box that asks "are you a robot" has no defence against a robot that says no.

Which is why the answer is not the point. The post describes "a high degree of sophistication behind that modest checkbox". Out of sight on Google's own servers, software weighs your whole visit to the checkbox. It looks at what happened before you clicked it, while you clicked it and after. Google's reCAPTCHA documentation says the checkbox will either pass you instantly or challenge you to prove you are human.

Cloudflare was blunter. Its post of 29 September 2023 declared that its own version, Turnstile, was now generally available and free to everyone, after a year in beta. The post says: "Bots definitely can check a box, and they can even mimic the erratic path of human mouse movement while doing so." The common explanation is that the box watches how you move your mouse. It also says "the actual act of checking a box isn't important". It tests the browser and makes it do small amounts of work "to prove that it's an actual browser".

Read that against the agent. It runs in a sandbox, a walled-off copy of a computer, with its own virtual operating system and a browser that reaches the real internet. It is an actual browser. Cloudflare checks the browser's features and asks it to run small tests, and the agent's browser can answer. It passed because it is the thing the test looks for. It never had to solve a CAPTCHA, meaning the distorted text or the image puzzle you work out by hand. It passed the screening that decides whether to show one. The narration about proving it was not a bot was earnest and beside the point.

The reason anyone puts the box in front of you is duller than the agent story. Cloudflare says humans abandon CAPTCHA puzzles about 15% of the time, and that the puzzles are easier for bots than for people. For over a year it rotated between CAPTCHAs and Turnstile. Turnstile worked just as well without asking users to interact at all. The box is mostly a cover story while the browser gets checked.

The 2003 paper that introduced CAPTCHA defined it as "an automated test that humans can pass, but current computer programs can't pass". It was written by Luis von Ahn, Manuel Blum, Nicholas Hopper and John Langford. They called the system win-win: either the test holds, or the test falls and a useful AI problem gets solved along the way. Two decades on, the web has taken a third option. The test did not hold and it did not fall. It was quietly retired, and the box was left standing in front of it as decoration. What sorts visitors now is not a question anyone can pass or fail, but an inspection of the browser doing the clicking. The agent did not beat the test. It was waved through by one that had stopped asking.