from the mod log:

Mojng5xTnVM7YHd.png

https://s.faf-pb.xyz/lXxek was the old url but they took it down - I have a copy at https://join.piefed.social/wp-content/uploads/2026/05/evidence-llm-used-for-banning.txt

The output from the LLM was provided as proof that someone needed to be banned.

I didn't want to do this but my hand was forced by their dissembling, minimization and bullshit. Note how the output includes a nice big logo with a slogan and ascii graphics - this was not a one-off experiment, this is an app they were proud to call their own.

you are viewing a single comment's thread
view the rest of the comments
[–] 38 points 6 months ago* (10 children)

And? It's in the public modlog, Rimu. Please, can you stop harassing me? How many hater posts are you gonna make this week exactly?

this was not a one-off experiment, this is an app they were proud to call their own.

It's a python script that calls the lemmy/piefed api to grab comment histories. Calling it an "App" is quite a stretch. Luminous just likes pretty things, bless his cotton socks.

You are right about one thing, we will no doubt keep using the script to pull comment histories on the few occasions when we want a deep dive into someone's profile.

For the record, we were never planning to use it as an automod or hook it to an LLM. You hallucinated that, ironically. That's why we never announced it. It was nobody else's business.

How it was actually intended to be used was as a tool to supplement the existing lemmy mod tooling, so that mods can make better, more informed decisions, in a more efficient way. How is it ever a bad thing to have all the relevant information to hand? And it's being open sourced soon, so everyone can use it. Unlike some instances [nudge, nudge], we try to take a considered approach to user bans. And no, it's no linked to an LLM in any way. If you chose to use one, you'd have to paste in the user history manually. But it also very useful just for skimming through, much faster than with Lemmy UI.

Also for the record, we do not use any form of automod with automated banning and never have. All ban decisions are made by a human mod or admin. Neither do we auto-scan user profiles for bad political takes, like Zionism. What we do do, is act on user reports, conduct an investigation and then make a mod decision. Nothing has changed in that respect, except now we have an extra tool to make sure we get the decision correct.

Finally, just because I included an LLM summary of the comment history in the public modlog, it does not logically follow that the ban occurred because of the LLM summary. It simply happened to correspond very well to my own manual review, which I have already explained to you. And I thought the summary did a much more comprehensive justification of the ban than I could have fitted in the modlog, so why not include it?

In conclusion, much ado about absolutely nothing. Again.

  • source
  • hideshow 10 child comments
  • [–] 14 points 6 months ago* (7 children)

    Aside from all the drama, is there a reason you are using the gpt-5-mini api from OpenAI (and aparently pay them for that) instead of a self-hosted model? This seems to be also against your instance policy and helps OpenAI to harvest training data.

  • source
  • parent
  • hideshow 7 child comments
  • [–] 9 points 6 months ago (6 children)

    We added the chatgpt reference because we thought it would be funny, luminous talked about it somewhere already. We are pirates, and self-hosters. This is surely obvious. And we don't have cash to splash around on corporate AI tokens.

  • source
  • parent
  • hideshow 6 child comments
  • [–] 19 points 6 months ago (1 child)

    Sadly this is a learning moment that inside-jokes don't belong in modlogs. Any nuance it would have would be immediately lost and twisted beyond any recognition.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 6 points 6 months ago (3 children)

    So what model are you using? Models are biased by their training data and guardrails put in by the trainers, so for transparency sake it is important to mention the model used.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 5 points 6 months ago

    Flatworm, looking for clarification. @db0@lemmy.dbzer0.com alleges you didn't use AI in your ban process. But then I read your comments and they are far more difficult to parse.

    Finally, just because I included an LLM summary of the comment history in the public modlog, it does not logically follow that the ban occurred because of the LLM summary. It simply happened to correspond very well to my own manual review

    • you claim you had a manual review process, but never show your work. (If you did, where is it?)
    • you imply you didn't use the LLM to ban the user, but you don't say it outright.
  • source
  • parent