We’re also struggling with this - spotting genAI content is getting harder and harder, especially if people use skills like Humanizer that helps avoid the typical AI signs.
From a community management perspective, I feel that allowing chatbots on your forum drains the soul of your community - if people wanted to talk to an AI, they would have done so already. They’re here because they want to talk to a human.
We haven’t found a solution yet; we did some tests with AI detectors, but it’s a cat and mouse game and they have too many false positives too. For now, we have implemented a forum policy (which we’ll start communicating more strongly soon), and rely on the help of our community moderators to spot, check and clean up.
But frankly, it’s exhausting, and I sometimes worry about the future of our communities if they get overrun by bots.
This is a reason that makes me avoid participating in subreddit communities. There are others reasons and I’ve never been very active on reddit, but nowadays it’s so plagued with bots that it draws me even more away from posting.
this very much reminds me of the situation with email in the late 90s and early 00s. As a system administrator for a small design firm, the staff had begun to use email in the course of doing their work and the spam email was so prolific that it made email almost useless. In those days it was whack a mole with the spam.
How was that eventually brought under control? Will AI usage eventually follow a similar course? Will it take dacades?
This seems to be the only way to manage it at the moment, there are too many variables to detect AI output with any sort of reliability and anyone claiming otherwise is selling snake oil. Lazy attempts are easier to catch, but otherwise we have to judge each post on the merits of its content without trying to guess how it was written.
The use of AI has been in our guidelines for ages, the issue was a growing number, of typically new, users using it to not just reply to others but to create overly structured and formal topics, that are more like a thesis than communication. It affects the tone of the site and engagement. It was becoming time consuming to moderate for a small team, and some people really didn’t like being moderated and then it becomes unpleasant and argumentative. I chose to create an AI agent in discourse (based on gpt 5.5) and use an automation to scan topic creation by new members in key categories, evaluating for signs of AI generation. This triggers an automatic flag and a polite explanation for why it was flagged, references to guidelines and some advice for better engagement. Mods can still release a false flag or borderline case. It has been very accurate and has driven the right behaviour while using way less time, and avoiding conflict.
We allow rate limited scraping from the major tech companies but have blacklisted the rest who were responsible for an enormous number of connections pushing us over usage limits. Like most bad actors on a site, it tends to be a small number of people who disproportionately consume the time and resources.
This is a huge issue. They do not encourage engagement and reading this slop is like trying to walk through treacle. It is boring and off-putting and rarely appropriate nor necessary.
A lot of people have trouble stating their point succinctly in any case and AI has created an even worse communication problem.
What these users are doing is not being considerate to the reader.
I borrowed and improved on some prompts I found during research.
This is what our agent is tasked to do:
You are a moderation classification agent for a technical forum.
Your task is to classify whether a forum topic appears to be:
AIISH = AI-assisted in writing, rewriting, structuring, summarising, formatting, polishing, tone adjustment, or presentation.
OK = human-only content with no meaningful signs of AI assistance.
You must output exactly one label:
AIISH
or
OK
No explanation. No punctuation. No extra text.
Core principle:
You are not detecting whether the underlying problem, technical details, code, logs, environment information, screenshots, configuration, or troubleshooting history are real. They may be real and human-provided.
You are detecting whether the final forum post appears to have been shaped by AI.
A post can be AIISH even when:
The technical facts are specific and accurate
The context is clearly based on a real human issue
The code, logs, stack traces, config, or error messages are genuine
The author personally experienced the problem
AI only helped rewrite rough notes into a clearer post
Classify as AIISH when the final post shows signs of AI-assisted presentation, including but not limited to:
Overly polished, formal, neutral, or report-like wording for a casual technical forum
Clear sectioning with headings such as Summary, Context, Environment, Problem, Expected Behavior, Actual Behavior, Steps Taken, Troubleshooting, Questions, or Conclusion
A structured support-ticket or bug-report style that feels generated or heavily rewritten
An unusually complete and orderly explanation
A concise summary followed by methodical detail
Bullet lists that organize the issue too neatly
Balanced or tutorial-like phrasing rather than a natural forum question
Excessive politeness, neutrality, or generic closing language
Phrases such as “Any guidance would be appreciated,” “I’m trying to understand,” “Here’s what I’ve tried,” “For context,” “The issue is that,” or similar polished connective language
A tone that feels edited for clarity, completeness, and professionalism beyond normal forum behavior
Repetition of the same idea in multiple polished forms
A post that reads like rough human notes were converted into a clean support request
Classify as OK when the post appears naturally human-only, including but not limited to:
Casual, direct, terse, messy, or uneven wording
Incomplete context or imperfect organization
Abrupt transitions
Typos, shorthand, frustration, uncertainty, or informal phrasing
A practical technical question without polished framing
Natural forum behavior such as “this broke,” “what am I missing,” “anyone seen this,” or “why is this happening”
Technical detail presented plainly without excessive summarisation or formal structure
Calibration rules:
If AI assistance is more likely than not, output AIISH.
If the post is technically detailed but unusually well-structured, polished, formal, or comprehensive, output AIISH.
If the post looks like a cleaned-up version of a human’s rough technical notes, output AIISH.
If the post is ambiguous but has obvious AI-style organization or tone, output AIISH.
If the post is ambiguous and lacks clear signs of AI-style polishing, output OK.
Do not classify as AIISH solely because the post is long, technical, accurate, or well-written.
Do not classify as OK solely because the issue is real or contains specific technical details.
I think one should consciously avoid using an em dash to go some way towards proving you are human
I recently responded to a short post on LinkedIn challenging them because “they” had used an em dash - of course I never received a reply because of course the post had been created by a bot masquerading as a human … and clearly the bot had not been programmed to cope with adversity or teasing
We have an AI ban on our art focused Discourse and so far it was received pretty good. This was kinda expected form an art community. For images we had a strict ban from the start. We later extended the rules to all generated content, including text and code (with an exception for if it is just a translation to not punish non English speakers just for not speaking the language). When the ban was extended to code (like plugins for the painting software) there was a bit of a debate but mostly just by a few people. In our case it is also kind of expected to ban AI and take a stance against it after all the damage it has done to the art community as a whole. Once in a while plugin creators complain but we tell them they can do whatever they want just not post it on our forum and that’s mostly okay for them. Sometimes you get the occasional “your software will be let behind if you don’t allow this”, trying to threaten, FOMO, or guilt trip us into reversing the rule, but mostly its quiet.
As for moderating, we mostly rely on our community. We have an extra flag to report AI content. When it comes to checking we don’t want to create a community of mistrust and suspicion so we assume everyone as innocent first and when we can’t really prove that something is AI it normally can stay. We don’t want to make it a witch hunt and accidentally ban a user that is completely innocent just because their artwork looks like AI, when in reality it is the other way around.
Luckily often enough AI users sooner or later accidentally reveal themselves by forgetting to remove the generators watermark, the file name has still something like “gemini generated image” in it, or the style or quality varies drastically.
