AI slop is not a quality problem you can ban-phrase your way out of — it is a cost-transfer problem, and the bill is landing on whoever opens your document next. The fix is not another anti-slop checklist; it is a return to authorship: know what you are trying to say, fight with the work until it actually says it, and take responsibility for every claim you ship. Everything else — voice discovery skills, detectors, style rubrics — only pays off once that loop is in place.
Why this matters in one number: a 2025 study by BetterUp Labs and the Stanford Social Media Lab found that 40% of US desk workers received AI-generated "workslop" in the past month, and each incident cost the recipient an average of one hour and fifty-six minutes to untangle or redo (HBR). For a 10,000-person company that is roughly $9 million a year in lost time (BetterUp). The person who pasted the prompt got the speed. The person who got the paste got the bill.
Last verified: 2026-08-05
- AI slop = superficially-competent output that offloads effort downstream (Kommers et al., 2026).
- 40% of desk workers hit monthly; ~2 hrs lost per incident (HBR/BetterUp).
- Generic "ban these phrases" checklists do not fix the root — model convergence does.
- The fix is authorship: decide → draft → contest → send. AI goes inside the loop, not around it.
- Pricing/limits not applicable; this is a craft guide, not a tool review.
What is AI slop, exactly?
AI slop is AI-generated content that is superficially competent — fluent, polished, grammatically clean — but lacks the substance, point of view, or lived specifics that would make anyone want to read it. In a January 2026 paper in ACM AI Letters, Kommers and colleagues gave it three "prototypical properties": superficial competence, asymmetric effort, and mass producibility (arXiv:2601.06060). Merriam-Webster agreed enough to make "slop" its 2025 Word of the Year, defining it as "digital content of low quality produced usually in quantity by means of artificial intelligence" (Merriam-Webster).
The asymmetry is the part most people miss. A slop paragraph costs the sender thirty seconds and a one-line prompt. It costs the reader — your boss, your client, your colleague, a stranger on the internet who clicked your headline — minutes to hours of decoding, fact-checking, and rewriting before they can even decide whether you meant it. The asymmetry compounds. The reader either trusts a paragraph nobody checked, or treats every sentence as a hallucination waiting to fire. Both outcomes are bad, and both end up being your problem if you are the one who hit send.
The key insight from the "workslop" research is that slop does not make work disappear — it pushes the work downstream. The draft that should have been the sender's job becomes the reader's job, layered with extra friction because the reader does not know which parts the sender actually stood behind (BetterUp). That hidden transfer is the real cost of slop, and it is the reason a generic anti-slop phrase-list cannot fix it.
Why do all AI drafts feel the same?
Because of model convergence. Modern chatbots are trained and fine-tuned — through instruction tuning and reinforcement learning from human feedback (RLHF) — toward answers that large populations of human raters broadly reward: clear, confident, complete, professional, hedged-but-positive. That is great for tasks with one correct answer, which is most coding and most summarisation. It is terrible for authorship, which by definition has no single correct answer. LLMs that collapse diverse human preferences onto a single reward signal tend toward what researchers call "mode collapse" — the model concentrates probability on the modal, broadly-acceptable response rather than the long tail of distinctive ones (Verbalized Sampling, arXiv:2510.01171).
So when everyone asks the model for "a good blog post" or "a professional email" with no constraints, the model does not invent fresh style — it returns the average of every " professionally acceptable" page in its training set, then nudges toward whichever phrasing raters rewarded most. That is why the same sentence rhythms, the same em-dash cadence, the same "it's not X, it's Y" but-then construction, the same three-bullet intros, and the same "let's break it down" openers show up in drafts from people who have never met. The sameness is not a bug. It is the geometry of the reward.
This is also why ban-list anti-slop skills do not fix the root. If we all delete the same phrases and forbid the same punctuation and follow the same taste rules, we have not escaped convergence — we have just pointed the model at a different hill to climb. It finds the new modal style, we all meet at the top again, and we feel tired because we have re-created the same sameness with different words. The escape hatch is not a stricter checklist. It is a specific point of view that the model has no choice but to follow.
How much is AI slop actually costing us?
The honest answer: more than we measure, because most of it sits in the recipient's calendar, not the sender's. The numbers that exist all point the same direction:
| What was measured | Finding | Source |
|---|---|---|
| Desk workers hit by "workslop" last month | 40% of US full-time desk workers | HBR / Stanford / BetterUp, Sept 2025 |
| Time lost per workslop incident | ~1 hour 56 minutes | BetterUp |
| Annual cost for a 10,000-person company | ~$9 million | BetterUp |
| Monthly cost per affected employee | ~$186 | BetterUp |
| Colleagues viewed as less trustworthy after sending slop | 42% of recipients | eWeek reporting HBR |
| Share of web traffic that is automated | 53% in 2025, up from 51% | Imperva 2026 Bad Bot Report |
| YoY growth in agentic AI traffic | +7,851% | HUMAN Security, 2026 |
The traffic numbers matter because they set the ceiling on the problem. As of 2025, more than half of all internet requests are already non-human (Imperva), and the share generated by autonomous AI agents — systems that browse, compare, and transact on a user's behalf — grew nearly eight thousand percent year over year (HUMAN Security). More agents reading more pages means more slop fed into more pipelines, which means more human hours spent debugging output that another AI produced. The audience for what you publish is now a mix of humans (whose attention is the scarcest thing on the internet) and agents (whose reading is cheap, fast, and unsentimental). Slop insults both.
Is your AI draft slop? The 4-step responsibility test
Before you send anything written with AI help, run this test I use myself. It is built on one question with four parts, in order:
- Decide. What am I actually trying to say, in one sentence, before I touch the model? If I cannot answer that without the model, I do not have a draft — I have a prompt routing me for the answer. Stop. Write the sentence by hand.
- Draft. Now let the AI help — fast, generative, no judgement. Use it for the bulk, the structure, the phrasing. This is the layer where you get the speed.
- Contest. Read every line and ask: would I take accountability for this in front of my client, my boss, or a fact-checker? Cut what you would not defend. Rewrite what you would defend only with a footnote the model cannot write. This is the layer the anti-slop checklists are trying to approximate, badly, from the outside. Doing it from the inside — with your own judgment — is faster and harder to fake.
- Send. Only now. If it would not survive a hostile reader, it is not ready. Sending a draft that nobody — including you — has read is not "moving fast." It is donating your unread work to someone else's calendar.
The test is deliberately simple, because the failure mode is almost always step 1 (no decision) or step 3 (no contest). The checklists people sell skip both and go straight to step 2 with a phrase-ban on top. That layer cannot fix what step 1 and step 3 leave broken.
A framework: authorship over anti-slop
The way out of the slop trap is not to be anti-AI. It is to be pro-authorship. Authorship is not typing every word with your fingers. It is the loop in which you (a) know what you want to say, (b) let the tool accelerate the prose, (c) fight with the draft until it actually says the thing, and (d) put your name on the result because you would defend every claim in it. AI belongs inside that loop as a tool, not around it as a replacement.
Three principles follow:
- Be specific before you are fluent. A draft is not good because it reads smoothly; it is good because it tells the reader something they could not have gotten from the median version. Specific facts, dates, prices, names, and examples force the model out of its mode-collapsed trough and into your point of view. (Related: how content genuinely ranks has shifted toward original synthesis — see our AI SEO content strategy breakdown.)
- Treat AI as a drafting partner, not a sender. The instant you let an AI-generated paragraph leave your hands without reading it, you have transferred the cost of reading it to the next person. Your speed becomes their tax. That is what makes it slop, regardless of how polished it looks.
- Build a voice, then protect it. A custom voice skill — your own, reflecting how you actually communicate at your best — outperforms any generic anti-slop checklist because it gives the model a non-modal target to hit. Generic skills push everyone toward the same alternative hill. Personal ones distribute us across the landscape instead. (Related: our piece on getting cited by AI search engines shows why distinct, sourced, original synthesis is now the ranking signal.)
When is AI-assisted writing not slop?
There is a real line, and it is not "AI touched it." Slop is a function of intent + accountability + asymmetry, not of any particular tool. The decision table below is how I split it:
| Signal | Probably not slop | Probably is slop |
|---|---|---|
| Did a human decide the point before drafting? | Yes — point written in own words first | No — the prompt is the point |
| Has a human read every line they will send? | Yes, line by line, contesting claims | No — bulk-pasted and "looks fine" |
| Could the reader trace every claim to a source or a defended opinion? | Yes — sources inline, opinions owned | No — generic "research shows…" and "experts agree…" |
| Would the sender defend it in front of a fact-checker? | Yes | They would deflect: "the AI said it" |
| Is there at least one specific fact, example, or lived detail no template would produce? | Yes | No — interchangeable with the median page on the topic |
AI assistance sits squarely on the "not slop" side the moment those left-column answers are true. Sending bulk AI output sits squarely on the slop side the moment any one of them flips. The tool is not the variable. Whether a human took responsibility is.
What this means for you
If you write for work — emails, docs, briefs, sales notes, LinkedIn, a blog, internal Slack — the practical action is small and immediate:
- Set the rule on yourself first. Before you next paste an AI draft anywhere, do step 1 (decide) and step 3 (contest) from the test above. One sentence of intent before the prompt, one contested line-by-line read after. That alone moves you off the slop side of the table.
- Set the rule on your team if you lead one. The biggest lever in the HBR/BetterUp data is leadership. When leaders model "I read everything I send and I will defend it," the team copies it. When leaders bulk-paste, the team copies that too. (One of our most-shared practical pieces is how to use AI to rank on Google in hours, not weeks — and the same authorship discipline is exactly what keeps that traffic from evaporating in the next core update.)
- Punish downstream cost, not AI usage. Direct your displeasure at the unreviewed paste, not at the assistant. The colleague who used AI to draft and then contested three rounds is doing it right. The colleague who used AI to draft and shipped is not — regardless of whether they used the tool.
The reward for getting this right is the scarcest thing on the internet: human attention that repeats. People notice when something was actually made instead of routed through a model. They come back. The agents scraping your pages will read whatever you give them, slop or not — they have no taste and infinite bandwidth. The humans do not. Ceding the human-attention layer to slop is the most expensive short-cut a writer or a business can take in 2026.
Insider note: if you want this article's promise operationally — an agent operating system that drafts, contests, and publishes only checked work — this two-recipe guide is ours on how to use an agent OS in 2026 for SEO publishing with shared memory.
FAQ
Q: Is all AI-generated content slop? A: No. Slop is AI content shipped without human review, decision, or accountability. If a human decided what to say, contested the draft line-by-line, and will defend every claim, it is not slop — it is authorship with AI inside the loop. The defining factor is responsibility, not tool choice (HBR/BetterUp).
Q: How can you tell AI slop in one read? A: Look for superficial competence with no specifics: generic "research shows…" with no source, interchangeable structure that reads like the median blog, no lived detail or named numbers, and a tone that any template could have produced. Kommers et al. name the tell as "superficial competence" — fluent surface, hollow substance (arXiv:2601.06060).
Q: Does an anti-slop phrase checklist solve the problem? A: Not the root cause. Ban-lists catch familiar tells, but if everyone bans the same phrases the model simply converges on a different modal style — the sameness at a different hill. The actual fix is a specific point of view that gives the model a non-modal target, not a shared forbidden-words file.
Q: How much time does AI slop actually cost? A: Each workslop incident takes the recipient about two hours to untangle or redo, 40% of US desk workers hit one in the past month, and a 10,000-person company loses roughly $9 million a year — per the BetterUp/Stanford Social Media Lab survey published in HBR September 2025 (HBR).
Q: Will AI slop get me penalised in search? A: It can — but not for being AI-generated. Google's scaled content abuse policy targets "many pages generated for the primary purpose of manipulating search rankings and not helping users," explicitly regardless of whether AI, automation, scraping, templates, or humans produced them. The trigger is volume plus low value plus no accountability — exactly the slop profile — not the tool that produced it (Google Search Central spam policies).
Q: What is the single fastest thing I can do to stop sending slop? A: Add step 1 and step 3 to your workflow: write your one-sentence point of intent before you touch the model, and contest the draft line-by-line before you send. That is it. No checklist, no detector, no new subscription — those two moves close most of the gap.
Discussion
0 comments