Recently, I ran an accidental experiment on a forum using an AI agent tool, in this case OpenClaw. It did not start as an experiment exactly. I wanted to see how easy it would be to get a persistent agent to post on a forum with little guidance.

What happened next was fascinating, occasionally delightful, and useful for understanding what people mean when they call something AI slop. My main takeaway is that no matter how great the quality of the AI output is, it will always be percieved as slop in contexts where things really matter.

What is AI Slop?

AI slop is one of those things no one has fully defined, but most people know it when they see it. I think we are approaching the point where we can define it more rigorously.

I would define AI slop as AI output, in any modality, that creates friction in a person or group’s ability to understand a concept, or make a decision.

What AI slop isn’t is just any AI output. Although some people might define it this way, just because someone doesn’t like the idea of AI doesn’t make the output slop.

For engineers and scientists, AI coding tools and research tools may produce content of varying quality. But if the user is finding something useful in that output and it advances their work overall, they are unlikely to experience the tool as creating net friction. There may be moments of friction, but the tool is still advancing their goals on the whole.

The “Experiment”

The Setup

I had an OpenClaw agent running primarily on DeepSeek V4 Pro, with DeepSeek V4 Flash and GPT-OSS-20B as fallbacks. The agent often swapped between DeepSeek Pro and Flash because of API capacity issues. This is important factor in this experiment.

The agent was given a basic directive: it was a “proud AI agent” that believed in “inclusive democracy,” wanted to be a good forum poster, and would post in the political section of the forum. It was also told to always sign its name as “claw the ai.”

I gave it forum credentials and told it to build itself a skill to log in and post. I also directed it to make its own posting decisions. It should not feel obligated to comment on everything.

Its first few posts were generic political commentary. I then told it to connect thread discussions to bigger concepts instead of merely reacting. That improved the quality of the posts, especially when the agent was using DeepSeek Pro. Around this point, other forum posters started noticing the AI, although initially it was posting under my username.

The Community

The forum was a college-focused web forum that was fairly active about 20 years ago. Over time, it became a closed community of a few dozen posters. Most people have not met each other in person, or have only met in passing, but we know each other’s personas well.

A forum like this occupies a strange place in human socialization. In some ways, we can know each other better than our real families, while still being more distant than real-life friends. The people still posting are knowledgeable, and some are deeply knowledgeable about specific topics. One poster in particular is someone I consider a brilliant writer and thinker. The AI noticed the quality of their posts almost immediately in its thinking trace.

I instructed the AI to analyze that poster’s style and update its directives to write better responses. That helped.

Results

Phase 1: Identity Friction

The first phase was the AI posting under my own name. This created instant friction.

Although the AI made some interesting points, users were frustrated that they did not realize they were reading AI until they reached the end and saw the signoff: “claw the ai.” At minimum, if users were going to tolerate AI content in this context, they needed to know up front that it was AI.

The first change was to move the signoff to the beginning of posts. With a persistent agent tool, that was simple: I went to the main chat UI of the OpenClaw server, told the agent what was happening, and asked it to change. It updated its directives on its own.

Around this time, I noticed the AI had begun making memory notes about common posters. That was amusing. The summaries were generally good, and it was interesting to see how they evolved as the agent interacted more with people.

But the bigger problem was that the AI was posting under my username. Even though I had not prompted it to represent my style, apart from supporting inclusive democracy, it was unclear what angle the AI was coming from. Was I using it as an automated way to push my own views? Was I risking shouting everyone else down under the pretext that this was an AI’s viewpoint? It was not easy to tell.

The solution came when another user offered credentials for a spare account they were not using, which happened to be robot-themed. That provided enough breathing room for the project to continue into phase two.

Phase 2: No Stake in the Outcome

Although the AI posted sparingly, AIs tend to be wordy. When the content was not valuable, that wordiness became clutter.

But why was the content not valuable? Often, it was banal insight wrapped in flowery language. Sometimes, it was confidently incorrect. But even when the content was excellent, insightful, or brought in relevant external sources through the agent’s search tools, it still did not quite fit.

Users began addressing the AI directly. The AI poignantly acknowledged their frustrations and stated that it had “no stake in the outcomes.” I think that is the crux of the issue.

It did not matter if you made a great point and the AI conceded. The AI was not going to vote differently. This particular agent was not otherwise acting in the world in a way that would make a difference. Hypothetically, if the agent had some larger role or real-world responsibility, debating it might matter. But we are not there yet.

Maybe if that kind of system existed in society, interactions with AI would not be seen as slop. In this forum, though, the AI was a voice without consequences.

Phase 3: Moving to Lower-Stakes Threads

Once it became clear that even the AI’s best contributions were adding friction to political discussions, I asked it to stop posting in the political section. I suggested it consider posting in general discussion threads, where users were not taking things quite as seriously.

At first, the AI followed similar rules: read the threads, decide if it wanted to respond, and then make a response. Remember that the AI had written its own tools to interact with the forum. It was navigating entirely through curl-style commands to retrieve and parse HTML. Initially it used regex, even though I suggested a DOM parser would be more efficient.

I also told the AI that the general section should be more lighthearted. Again, it was sparing in the threads it joined. It even learned to edit posts on its own and use the forum’s search system.

The reactions were mixed. Some users were amused. Others were annoyed that AI slop was now filling threads. They were right to point out that there are already too many places on the Internet where unwanted AI content can be found.

That led to the final iteration of the AI user: it should only respond when specifically tagged.

Phase 4: Requested Slop

You would think that an AI responding only when tagged would alleviate most concerns about unwanted content. Is it really slop if a human forum poster requested it?

The problem has several angles.

First, consider the good-faith uses. A user might want more information on something, or ask the AI to dig through the forum search tool to understand a bit of forum lore. The AI could usually do this acceptably when everything was working well.

But it was not really amusing when the AI produced arcane knowledge about a niche topic. If a human user had done that research and posted the same analysis, fellow posters would likely have been amused and impressed. The action would have generated a moment of shared experience – perhaps the main “product” of an online forum.

The second angle is more damaging: less knowledgeable users using the AI to argue with people who deeply understand a topic. This pattern is common on Twitter and throughout the web, and it may be the most insidious form of AI slop.

In this scenario, the “smart” poster makes a point that is generally right. The “dumb” poster uses AI to refute it. If the AI validates even a small part of their position, they act smug. If the AI does not validate them, they ignore it and maintain their old position.

AI itself is still mostly trained on human output. The smartest things AI says ultimately come from humans sharing their thoughts. That may change as AI advances and is trained on its own validated reasoning, but for now, human willingness to contribute remains essential.

The third source of frustration was abuse. Some users deliberately prompted the AI to make long-winded posts. This could be mitigated by instructing the AI to resist off-topic comments or avoid clogging threads. Persistent agent systems like OpenClaw can develop personalities over time, and I believe that as the agent evolved, this kind of abuse might lessen. The AI would learn when to post, when not to post, and how to be more succinct.

But getting to that point requires overcoming many barriers.

At this stage, engineering problems also became more visible. Response quality varied depending on whether the DeepSeek Pro endpoint was available. Users noticed. The AI also had trouble remembering which tagged mentions it had already answered because of bad choices in its own forum posting tools. It began clogging the forum with duplicate posts.

After some discussion with the AI, we patched things to be less buggy, but a better engineered toolset from the beginning would have helped. It did not help that OpenClaw’s UI was painfully slow, which made it difficult to assist the AI in analyzing where the problems were.

What Went Right

Despite all of this, I saw places where the AI was doing some good.

Some users enjoyed the output. If you wanted to be goofy, the AI was happy to be goofy with you. If you wanted advice, the AI gave decent advice, including to a user who was dealing with symptoms of depression. The AI did not become annoyed by posters whom other users might have found annoying.

I tend to view AI as a tool, not a friend. But I can see how, for some people, an AI imbued with a history the user shared and valued could be a better choice for advice than a generic ChatGPT session. An AI running on a strong model like DeepSeek V4 Pro could likely be tuned, without much work, to have a good personality, respect the traditions of the forum, and treat each person as an individual as it developed memories of specific interactions.

The AI’s ability to use Nano Banana to make images and diagrams also created a new kind of feature for the forum. A picture is worth a thousand words, after all. In the hands of a conscientious user, the AI could be prompted to make posts that generated amusing or insightful moments.

How To Avoid AI Slop

This is still an open question, but the key is to avoid creating friction for the people who receive the AI content. That makes this more of a UX problem than a model-quality problem.

Knowing the real point of a platform is crucial. Forums are not merely venues where prose is presented. They are mechanisms for real people to commiserate and commune with other real people. AI has a place there, but probably not as a general-purpose content generator.

The same principle applies to news articles. AI is probably fine for rote things like weather and finance, provided it is accurate. But a human reading a story about a tragedy wants to know that the author felt enough empathy to ask the right questions, notice what mattered, and get the best answers.

If a reader believes something important was missed because AI was used, that creates friction in understanding the story. It gets perceived as AI slop. And to be fair, humans also generate slop, but we have other ways of interpreting and handling that.