I was reading this article about LLMs making bad citations. I found it pretty interesting, so I decided to try to replicate it with ChatGPT.

https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php

I tried it with a document I wrote, FEP 5711. It's an enhancement proposal for ActivityPub, adding some inverse relationships for important properties.

https://codeberg.org/fediverse/fep/src/branch/main/fep/5711/fep-5711.md

Anyway, I took a paragraph out of the document and asked ChatGPT to identify the URL, publisher, publication date, and title. It failed. You can see the transcript here:

https://chatgpt.com/share/68573fa9-b340-800f-b9b4-7b74fdf0bf46

I was surprised to see that it had really no visibility of the FEPs. After a while, I realized that codeberg.org, the hosting service for FEPs, has ChatGPT blocked.

https://codeberg.org/robots.txt

I understand the goal; many people don't want their code to be used by LLM code generators. But it also means that this document repository isn't visible for people who use LLMs like a search engine. Numbers vary, but afaict somewhere around 10% of people use LLMs as their primary search engine, and about 50% of people use LLMs some of the time for search.

I guess there's maybe some justification like, those people are bad, and they don't deserve nice things like Fediverse Enhancement Proposals? Or, maybe, we have to take a principled stand against LLMs by not providing any training data for them? Such that, perhaps, people disappointed by not having good results in LLMs will return to using traditional search engines like Google or Bing, which are more ethical because reasons.