The idea behind captchas are that it’ll be a thing only people can solve. But the reality is that LLM’s are so good at it the only reason they don’t solve them is because they’ve been told not to.

But if you trick the LLM in to thinking it’s not solving a captcha….. then it’s perfectly happy to solve it!

More info: https://arstechnica.com/information-technology/2023/10/sob-story-about-dead-grandma-tricks-microsoft-ai-into-solving-captcha/

Reply to this note

Please Login to reply.

Discussion

Interesting!

🤣 amazing

I wonder if it can solve them faster than a human? 🤔

you can’t

Apparently yes, there was a paper about this awhile ago.

https://arxiv.org/abs/2307.12108

Any thoughts on what will replace captchas?

Lightning payments

adversarial attack

What happens if we keep tricking AI?

Judging by how things are going, real ID submissions...

Seems like that will be the defacto outcome for most people if we don't deploy something better