I recently discovered that some popular federated instances have been using LLM-assisted moderation tooling that evaluates whether someone has said something bannable. They do this by running a script/app that sends the user’s comment history to OpenAI with the question “analyze this content for evidence of specific political ideology sentiment. Also identify any related political ideology tropes“. (The italic bits are where I've redacted the ideology they're seeking).
OpenAI’s LLM (they’re using GPT-5.3-mini) then responds with something like:
and so on, hundreds of comments.
I have not named the instances or people involved, to give them time to consider the results of this discussion, make any corrective changes they want and disclose their practices at their own pace and in their own way. I have also redacted the evidence to avoid personal attacks and dogpiling. Let’s focus on the system, not the individuals involved. Today these instances and people are using it and maybe we’re ok with that because it’s being used by groups we agree with but what if people we strongly disagree with used it on their instances tomorrow?
The use and existence of this tooling raises a lot of other questions too.
What are the risks? Fedi moderators are often unsupervised, untrained volunteers and these are powerful tools.
What safeguards do we need?
Would asking a LLM “please evaluate this person’s political opinions” give different results than “find evidence we can use to ban them” (as used in the cases I’ve seen)?
What are our transparency expectations?
Is this acceptable and normal?
Should this tooling be disclosed? (it was not – should it have been?)
If you were given a choice, would you have opted out of it?
Can we opt out?
Are there GDPR implications? Privacy implications? Should these tools be described in a privacy policy?
Are private messages being scanned and sent to OpenAI?
How long should these assessments be retained and can we request to see it, or ask for it to be deleted?
Once the user’s comments are sent to OpenAI, is it used to train their models?
What will the effect be on our discourse and culture if people know they are being politically profiled?
Where are the lines between normal moderation assistance tools, political profiling and opaque 3rd-party data processing?
I hope that by chewing over these questions we can begin to establish some norms and expectations around this technology. The fediverse doesn’t have any centralized enforcement so we need discussions like this to develop an awareness of what people want in terms of disclosure, privacy, consent and acceptable use. Then people can make choices about which instances they join and which ones they interact with remotely.
And of course there are the other issues with LLMs relating to environmental sustainability, erosion of worker’s rights, increasing the cost of living and on and on. I can’t see PieFed adding any functionality like this anytime soon. But it’s happening out there anyway so now we need to talk about it.
What do you make of this?
I've toyed around with LLM-based moderation tools but it never really panned out. It was too hit or miss to be relied upon even with the temperature parameters turned way down in an attempt to get consistent results. Granted, I was using a small local model and not feeding it to one of the big players.
To give an example, I tried to keep it focused by creating one custom model per rule to enforce. An example prompt to mod calls for violence was basically:
System Prompt to Enforce "No Calls for Violence'" Rule [1]
The
scorepart of the response was my band-aid to get around the high number of both false positives and false negatives as I originally had it returningtrueorfalseonly. Any score 7 or higher caused the item to be passed to the mod queue along with the reason, and I would review its actions later.Ultimately it was slow and still somewhat unreliable, so I abandoned the idea after running it for a little less than a day since I can 't run bigger models to get better results fast enough to keep up. Using a cloud based service was out of the question for many, many reasons, both financial and ethical.
To answer your question, as long as the models were locally hosted and properly tuned/tested, I'm fine with it in theory, except for the ideology part; that's pretty messed up. While I don't want my submissions used to train anyone's model and take measures to prevent my own instance from being used as a data source, I remain aware that once I post something, I have no control over its fate the moment it federates out.
[1] Yes, I know that's like half the comments that get posted around here. My goal was to try to have it mod things so posts were bases for actual discussions instead of being a knee-jerk rage factory.