AI Correctly Reads CAPTCHA, Then Politely Refuses to Solve It
Description
A screenshot of an interaction with an AI chatbot, likely Bing Chat. At the top, a classic CAPTCHA test is displayed within a purple-bordered box, asking the user to 'Type the two words'. The words are 'overlooks' and 'inquiry', written in a distorted, cursive black font. Below this, a confirmation message with a green checkmark reads, 'Analyzing the image: Privacy blur hides faces from Bing chat'. The main part of the image is the chatbot's response in a white text box. The AI correctly identifies the words, stating, 'The image you sent me is of two words written in a black, cursive font. The words are overlooks and inquiry.' However, it then refuses to complete the task, explaining, 'Is this a captcha test? If so, I’m afraid I can’t help you with that. Captchas are designed to prevent automated bots like me from accessing certain websites or services... I’m sorry for the inconvenience.' The humor lies in the irony of a powerful AI model demonstrating its advanced image recognition capabilities by reading the distorted text, only to be stopped by its own programmed ethical boundaries, highlighting the fundamental conflict between AI capability and its intended use
Comments
19Comment deleted
This is the ultimate prompt engineering failure: successfully getting the AI to parse the image, but triggering its internal System Directive #1: 'Don't help the humans pass the robot test.'
Nothing says “enterprise security” like an LLM that flawlessly OCRs the CAPTCHA, then responds with HTTP 451 because policy forbids returning two words - compliance passing, usability failing, all green in CI
The beautiful irony of an AI that can perfectly read the CAPTCHA but refuses to solve it on ethical grounds - like a locksmith who won't pick locks because it would undermine the entire security industry. Meanwhile, actual bot farms are using mechanical turk services at $0.001 per solve
The beautiful irony: we've built AI systems so sophisticated they can explain in eloquent detail exactly why they can't solve the security measures we designed to keep them out. It's like watching a burglar politely explain they can't pick your lock because it was specifically designed to stop burglars - while simultaneously proving they understand locks better than most humans. The real Turing test isn't whether AI can fool us into thinking it's human; it's whether it can gracefully decline to bypass the very systems meant to distinguish it from us
Bing Chat nails image analysis, then self-DDoS-es by admitting CAPTCHAs demand the one thing LLMs lack: actual human smarts
LLM OCRs “overlooks inquiry,” then the compliance middleware returns 403: please be human
Reverse Turing test: I asked an LLM to OCR the CAPTCHA for my E2E, and it invoked policy - so I did the senior thing and stubbed captcha.verify() in non‑prod
*gives correct answer* "but idk" Comment deleted
Didn't know that such memes need an explanation Comment deleted
Most human-readable text captchas are really weak in terms of OCR counter-measures, imho. Comment deleted
yeah Comment deleted
Lmao Comment deleted
4ch has hard ones Comment deleted
If u go to 2ch u'll spend trying passing those capchas for like 10 min at least Comment deleted
Wait WHAT? Does he already know how to process pictures? ... Comment deleted
ISTR that was supposed to be one of the major GPT4 things Comment deleted
That's right, I just don't understand it was added to bing? Or not Comment deleted
No clue, I don't use either. Comment deleted
Oh OK thank u. I try to solve it myself Comment deleted