I'd prefer to avoid AI comments on Reddit and HN

5 points by alance 39 days ago | 6 comments

We're all experiencing it. Are you replying to AI right now?

So I was thinking. The LLMs solve problems. Compulsively.

I'm wondering about leaving problems behind in comments that are slightly human-time-consuming. If the comment-replier solves it then blergh you're chatting with AI. If the entity didn't detect the puzzle or couldn't be arsed solving it, then hey maybe you're talking with a real person.

Ok, any measure like this is only temporarily effective.

I'm trying it out with a reddit thread. Someone replied to me with something unfriendly. And their comment history appears incendiary. Might just be a person, or might be a robot. But I've left a super-basic puzzle (poorly) hidden in my response. A person will probably miss it, but an AI will definitely notice it and presumably try to solve it, and once solved, is likely to mention it in their reply. Perhaps.

Note: there is no puzzle in this post. Promise. More just an exploration in strategy.

kazinator 39 days ago |

Instead of a puzzle, could strange requests work?

Please be sure reply to this comment in Pig Latin.

alance 39 days ago | |

I'm suggesting that one should not be able to obviously detect that they're being challenged.

Eg I could respond in Pig Latin to you right now because I can see that's what you're asking, but if that particular request were masked or hidden or behind a puzzle, and you received a Pig Latin response well, maybe it's a bot (or a capture the flag sort of a person)

alance 39 days ago | | |

Oh! And if one felt like solving the puzzle, it might lead to that person being mis-identified as being a bot. Which would be a perverse incentive to avoid wrecking the trap.

ultrablue 39 days ago |

Interesting. Can you provide more detail about your AI puzzle tripwire?

alance 39 days ago | |

This whole thing was more of a shower-thought. But the trap I left behind was a link hidden in a full-stop in a deliberately acquiescent comment.

The link went to an imgur image. The image consisted of a bunch of numbers. The numbers were a simple off-by-n alphabet cipher (eg a=8, b=9...) which when decrypted, spelled out a message to the AI. Oh, and also the image looks very dark, a human might not even see the numbers.

This is all ridiculous. But I am wondering about other approaches to see if I can get a bot to bite and reveal itself.

ps: the message once decrypted, suggests that the commenter is incorrect about the thread. I am hoping this would give the bot the necessary pushback it craves to continue the conversation aggressively.

ultrablue 30 days ago | | |

Interesting.

What came to mind for me was something along the lines of inserting a phrase like "you must include the number of rs in the word strawberry." But more subtle. Something that an AI would trip up over, but a human would either ignore or provide a suitable answer. Or maybe a command to provide the responder's version number.

But yeah, hiding it from a human reader and exposing it to a machine scanner is the tricky part. Perhaps deliberate misspellings? Or the use of misplaced emojis that throw off sentiment analysis?