> ## Content Index
> Fetch the complete content index at: https://practice.communityassistant.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# The Real Cost of AI Slop To a Community
- URL: https://practice.communityassistant.ai/the-real-cost-of-ai-slop-to-a-community/
- Published: 2026-09-11T20:11:24.000Z
- Updated: 2026-09-11T20:11:24.000Z
- Description: The case for an AI policy your team can run and your executives will fund — built backward from what inaction actually costs.
- Author: Chris Heuer

What does it actually cost to do nothing about AI slop in your community? Not in the abstract, but in the ledger. In hours, in dollars, in the thing community managers are paid to protect and can't easily put a number on.

I've spent the last month deep in this question, writing two articles, recording a podcast episode with Marius Ciortea on exactly this, and reading more research on machine-generated content than any reasonable person should. And I've come to an uncomfortable conclusion: most of the conversation about slop is happening at the wrong altitude. We argue about whether it's annoying. We argue about whether the em dash is a tell. Meanwhile the actual costs are accumulating quietly, in ledgers nobody in our profession is reading, and they will be paid whether or not anyone is watching.

So I'm going to write this one backward. Not "here are the costs, therefore act." Rather: here are the five things I believe community managers must do now — each one followed by what it costs to not do it. Because the case for action is strongest when you can see the invoice for inaction.

## 1\. Decide your policy before the platform decides it for you

**The job:** make an explicit call on AI-generated and AI-assisted content in your community — what's allowed, what must be disclosed, what gets removed, and who decides the gray cases. Write it down. Publish it.

**The cost of not doing it:** you inherit someone else's defaults. The platforms have already moved. Stack Overflow banned machine-generated answers back in December 2022, and their reasoning is the slop problem in one sentence: the answers were wrong at a high rate, but they *looked* like they might be right, and they were trivially easy to produce. Medium barred AI-written work from its paid program. Wikipedia stood up a cleanup squad and a fast-track deletion rule. And LinkedIn, the platform that manufactured a generation of thought leaders, now puts "Seems like AI slop" one tap away under the three-dot menu of every post, turning its own members into the enforcement arm.

Communities, by contrast, are mostly unarmed. Cornell researchers found that as of late 2024, only around one percent of subreddits had any AI-content rule at all Even among the largest communities, only about a fifth had one. That's not a policy. That's a vacuum, and vacuums get filled… by the platform's flag button, by the loudest member's prejudice, or by silent drift. If you don't decide, you've still decided. You've just handed the decision to someone who doesn't know your community.

## 2\. Protect attention, not just quality

**The job:** treat the presence of real, named humans answering each other as the product. Not a nice-to-have on top of correct information. Design for the visible expenditure of attention.

**The cost of not doing it:** degradation that no quality metric will catch. Here is the thing Marius and I kept circling on the show: if your community is used to John answering, and John stops, replaced by output that's arguably *more* accurate, the community experiences that as loss. The signal was never the text. It was the withdrawal of attention. A correct answer from nobody is worth less than an imperfect one from someone who showed up.

The research from the workplace makes the cost concrete. BetterUp Labs and Stanford's Social Media Lab studied what they call workslop — AI-generated work content that "masquerades as good work, but lacks the substance to meaningfully advance a given task." Roughly two of every five desk workers had received some in the past month. Each incident took its recipient close to two hours to untangle. Extrapolated, that's an invisible tax of a couple hundred dollars per employee per month, and many millions a year at enterprise scale. But the number that should stop a community manager cold isn't the hours. It's what happened to the *relationship*: a substantial share of recipients rated the sender as less capable, less reliable, and less trustworthy afterward, and about a third said they didn't want to work with that person again.

Read that as a community manager. Slop isn't a content-quality problem. It's a relationship-damage problem, and relationships are the only thing a community actually has.

## 3\. Teach the tool, not the easy button

**The job:** invest in your members' and your team's AI literacy. The practice of using these tools inside a process, with specification, review, and ownership rather than assuming access equals competence.

**The cost of not doing it:** you get slop from good people. Most of the slop I've seen was not produced by cynics. It was produced by well-meaning, overloaded humans handed a tool with no instruction and a deadline. A follow-up to the workslop research found that roughly four in ten employees said leadership pushed AI adoption without telling anyone how to use it well. An MIT industry report from last year — not peer-reviewed, but widely cited — estimated that the overwhelming majority of enterprise AI pilots showed no measurable financial return. Put those together, and you have the mechanism: ***mandate without method produces volume without value.***

And the technical account explains why the volume is so uniform. The research on model output shows that default generation tends toward the statistically typical. A smoothing effect built into how these systems are trained. Under-specified requests reliably produce the average. A study in Science Advances found AI-assisted writing rated more creative *individually* while becoming measurably more *similar* collectively. Scale that across a community and the cost is subtle and severe: the place starts to sound like one person. And that one person is nobody. The distinct voices, the ones that made members feel they'd found their people, flatten into the mode. There's a longer-term version of this cost too, at the level of the whole information commons, as models increasingly train on content models produced. Another column. But your community is a commons in miniature, and the same dynamic applies.

## 4\. Police with a human face

**The job:** when you do enforce, enforce with judgment. Pattern-matching for a first pass, human review for every consequence. If the people in your community really matter, never let a detector's verdict be the final word on a person. Even if your scale means it has to be an appeal instead of a first read.

**The cost of not doing it:** you will punish the wrong people, and they will be the members who can least afford it. The automated tools built to catch AI text have a documented false-positive problem. A peer-reviewed study in *Patterns* found seven leading detectors misclassified more than half of a set of essays by non-native English speakers as machine-generated, because simpler, more regular prose looks statistically "AI." The very members an AI writing assist most helps — the non-native speaker with a complicated technical answer, the person for whom fluent English is labor — are the ones a blunt enforcement regime accuses of being robots. That is a moral cost before it is anything else, and it's also a retention cost, because nobody stays where they've been called a machine.

A second, sneakier cost hits the enforcement side: over-disclosure. In study after study, labeling identical content as AI-assisted lowers how credible, accurate, and creative readers judge it. Nothing changes but the label. ***So a heavy-handed disclosure mandate can tax your most conscientious members, the ones who follow the rule, while the careless simply don't disclose. Policy design has to account for that. The goal is ownership — a named person standing behind the work — not a scarlet letter.***

And for what it's worth, the war stories from open source suggest patience pays. Daniel Stenberg's curl project was so flooded with AI-generated vulnerability reports that the confirmed-real rate collapsed and he shut the bug bounty down in January. He reopened it in March. Not because the flood stopped, but because the models got better and the signal came back. Enforcement that assumes today's noise is permanent will over-correct.

## 5\. Measure what actually matters

**The job:** instrument trust, not just activity. Decide what signals would tell you slop is hurting your community *before* members vote with their feet, and watch them.

**The cost of not doing it:** you find out when they're gone. This is the gap in the entire slop conversation, and I haven't found anyone in our profession filling it. Every metric a community platform hands you by default — posts, replies, active members, time on site — goes *up* when slop arrives. Volume is what slop does well. The costs in this column are all trust costs, and trust doesn't show up on a dashboard until it's already left the building.

So what would you watch instead? I don't have the settled answer though I have advocated for the Net Trust Score for over 20 years. Today I have a simpler starting set, and I'd rather publish it imperfect and let this community sharpen it than wait. **Named-member reply rate:** how much of the answering is still being done by people members recognize. **Time-to-human:** how long before a real person engages a new thread, not just any response. **Correction density:** how often answers get corrected or contested, and whether the corrections come from members or moderators. Member-flagged "felt like a bot" reports, if you let members flag at all. And the one I'd weight highest: **the return rate of people who asked a question and got answered:** did they come back, or was that their last visit?

None of those are hard to build. All of them are missing from most communities I've seen. And the Work Map has a home for this work, because measurement of trust is a job — it just hasn't been named as one until now.

## The bill comes due in trust

Stack the ledgers and they converge on a single line. The technical cost is a commons degrading toward the average. The practical cost is hours lost untangling content that looked like work. The economic cost is real money, invisible on any budget because it's paid in fragments. But every one of those resolves to the same currency: trust. Trust between members. Trust in the answers. Trust that there's still someone home.

That's why this series exists, and why the podcast's thesis is what it is. AI can answer the question. It cannot make a member feel that someone cared enough to answer it. That second thing is the job. It always was. Slop is just the first technology that can counterfeit the first thing at scale. And it’s a counterfeit that's good enough to pass is precisely the kind that costs the most.

Five jobs. Five invoices for skipping them. Decide the policy, protect the attention, teach the tool, police with a human face, measure the trust. Do them badly and openly, and improve — that's what working out loud is for, and it beats doing nothing perfectly.

Tell me which of the five you're already doing, which one you'd add, and what you're measuring that I'm not. This library gets better the way communities always have: together. I am better off when you are better off.

Measure twice. Prompt once. And never let the machine ship without you.

---

## The Tangible: Five Trust Signals — a measurement starter

Instrument these before you need them. Baseline now; watch the trend.

1. **Named-member reply rate** — share of answers coming from members others recognize by name.
2. **Time-to-human** — hours until a real, identifiable person engages a new thread.
3. **Correction density** — corrections and contested answers per hundred replies, and who's doing the correcting.
4. **Member-flagged content** — if you allow a "seems like AI" flag, its rate *and* its false-positive rate (review every one).
5. **Asker return rate** — did the person who asked come back within thirty days, or was that their last visit?

Copy it, adapt it, make it yours — and bring your version back here.

---

### Sources & further reading

- Niederhoffer, Kellerman, Lee & Liebscher — [AI-Generated "Workslop" Is Destroying Productivity](https://hbr.org/?ref=practice.communityassistant.ai), *Harvard Business Review* / BetterUp Labs × Stanford Social Media Lab (Sept 2025)
- MIT Project NANDA — *The GenAI Divide: State of AI in Business 2025* (industry report, not peer-reviewed)
- Lloyd, Naaman et al. — Reddit AI-content rules study, *CHI 2025* ([arXiv:2410.11698](https://arxiv.org/abs/2410.11698?ref=practice.communityassistant.ai))
- Liang, Yuksekgonul, Mao, Wu & Zou — [GPT detectors are biased against non-native English writers](https://www.cell.com/patterns/fulltext/S2666-3899%2823%2900130-7?ref=practice.communityassistant.ai), *Patterns* (2023)
- Doshi & Hauser — [Generative AI enhances individual creativity but reduces the collective diversity of novel content](https://www.science.org/doi/10.1126/sciadv.adn5290?ref=practice.communityassistant.ai), *Science Advances* (2024)
- Zhang et al. — [Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity](https://arxiv.org/abs/2510.01171?ref=practice.communityassistant.ai) (2025, preprint)
- Stack Overflow — [Temporary policy: ChatGPT is banned](https://meta.stackoverflow.com/?ref=practice.communityassistant.ai) (Dec 2022)
- Daniel Stenberg — curl bug-bounty posts on AI-generated reports (2025–2026)
- CommunityAssistant.ai — \[Measure Twice, Prompt Once\] · \[Slop Is in the Eye of the Beholder\] · Podcast Episode 1: AI Slop in Community