Re: 🇳🇿 ‘Zero to 100 ’: ChatGPT disputes clogging up the rental system, agents warn of price hikes
Richard Kettlewell <[email protected]>
| Newsgroups | comp.misc |
|---|---|
| Organization | terraraq NNTP server |
| Message-ID | <[email protected]> |
Lawrence D’Oliveiro <[email protected]> writes: > 🇳🇿 tenants are increasingly using ChatGPT to help them write > complaints for submission to the Tenancy Tribunal > <https://www.oneroof.co.nz/news/zero-to-100-chatgpt-disputes-clogging-up-the-rental-system-agents-warn-of-price-hikes-49682>, > and this is not always a good thing. The closing lines sum it up best: > > However, [Tenancy Advisory director Sarina Gibbon] describes AI as > an amplifier, not an equaliser. > > “If you have a Tribunal case that has merit and understand how the > law applies to your situation, AI can amplify that. But if you > have a weak case and little understanding of tenancy law, AI will > amplify that too.” > > How does a user tell whether they’re on the right track or completely > the wrong one? AI will happily drive them further in the same > direction, regardless ... They don’t seem to be able to say “no”. Two experiences: ChatGPT has at some point learned to count the number of Rs in the word strawberry. But while exploring this, I asked it increasingly nonsensical questions (like “which eight days of the week have a Q in?”). Although it consistently got the numeric part of the answer right, it continued to generate responses stating that I was engaging in wordplay - rather than just asking it stupid questions. A human with any common sense whatsoever would have told me to stop being so silly. That’s just mucking around, but the other one was an attempt to get something somewhat useful done: https://blog.trailofbits.com/2026/06/12/factoring-short-sleeve-rsa-keys-with-polynomials/ publishes a couple of weak RSA public keys, in the form of a screenshot. I wanted to have a play to see if I could reproduce pattern 1 from a plausible key generation strategy (bugs in key generation are professionally relevant to me), and didn’t fancy copy-typing 512-digit numbers, so I asked Copilot to do it for me. It identified immediately that there were two columns, but struggled with two things: the exact boundaries between them, and the zero digits, which I never got it to reproduce at all. I spent several iterations pointing out its mistake, and each time it generated a response apologizing and stating it would be more careful, and then getting output wrong in the essentially the same way again. Obviously after a few tries I gave up and copy-typed pattern 1 myself: it would all have been much quicker if it had admitted from the start that it couldn’t do this task that a child could have easily achieved. -- https://www.greenend.org.uk/rjk/