What is AI slop?
AI slop is content made with AI that takes the reader more effort to use than it took the writer to produce. In a BetterUp Labs and Stanford survey, 40% of US desk workers had received this kind of work from a colleague in the past month. Software teams catch bad AI output by testing it against failures they have seen before, and an office can do the same with a short list of yes or no questions.
Questions as you read? Ask, or take the assessment with you.
Where should we send it?
Sent ✓ Download it now →
AI slop in everyday work
Shopify’s chief executive Tobi Lütke told Fortune in September 2026 that employees at his company were tossing unreviewed AI messages at each other, and he called them “slop grenades.” His company had made AI use a baseline expectation the year before.
People call this kind of work AI slop.
What is AI slop?
AI slop is content made with generative AI that the reader did not ask for and cannot use. Merriam-Webster, which picked “slop” as its 2025 Word of the Year, defines it as digital content of low quality produced usually in quantity by AI. Australia’s Macquarie Dictionary picked “AI slop” the same year and defined it as low-quality AI content, often with errors, that the user did not ask for.
The developer and writer Simon Willison helped spread it in May 2024. He compared it to spam: once there was a short name for unwanted email, sending it became something people saw as bad behavior. He credited an online account called deepfates with using the word this way first. He said not every piece of AI content counts.
What is workslop?
Workslop is the office version. Researchers at BetterUp Labs and the Stanford Social Media Lab named it in Harvard Business Review in September 2025: AI-generated work that looks like good work but lacks the substance to move a task forward. Merriam-Webster’s own list of 2025 slop includes “workslop.”
In a follow-up in January 2026, the same researchers described workslop as work that leaves the thinking to whoever receives it.
The AI Briefing
Tuesdays. 500+ leaders. No hype, just what works.
How do you spot AI slop at work?
| Document | What you notice |
|---|---|
| A client proposal | Could go to any client |
| A status update | Items listed with no sign of which needs a decision this week |
| A meeting summary | Lists what was said in order, with the one change that matters in paragraph four |
| An email to a colleague | Three paragraphs where one sentence would do, with the question at the end |
| A LinkedIn post | A hook, then a list of five, like everyone’s feed |
| A report with numbers | The prose and the table disagree, and nobody checked which one is right |
The Human Domain.
The assessment, five scored dimensions, and the interval controls that keep judgment yours.
On its way. Check your inbox, or download it now →
In a Northeastern University study, copy editors called an AI-written text slop mostly when it didn’t matter to the reader or used too many words for what it said, more than for any style habit.
Can you tell if a colleague used AI?
Not reliably. In a 2025 Pew Research Center survey, 76% of Americans said it was very important to know whether something was made by AI, and 53% weren’t confident they could tell. AI models do no better: in the Northeastern study, the models asked to spot slop matched the editors’ calls about as often as guessing would.
So judge the work: did it give this reader what they needed, in as few words as it could? Use a list of tells on your own drafts and leave colleagues’ work out of it, because the Northeastern editors flagged human writing as slop too.
How much does AI slop cost a company?
In the BetterUp Labs and Stanford survey of US desk workers, people who received workslop said each instance took them nearly two hours to deal with. From those time estimates, BetterUp’s researchers put the cost at about $186 per affected employee per month.
Matthias Holweg and Thomas Davenport warned in Harvard Business Review in June 2026 that slop can end up in a company’s processes, where a thin summary becomes the record of a decision.
What does it do to your reputation?
In the BetterUp research, about half of the people who received workslop thought less of their colleague’s creativity, ability and reliability, and nearly one in three were less likely to want to work with that colleague again.
In a University of Florida study of professionals, most readers found a supervisor’s congratulations sincere when AI helped only a little, and about half or fewer did when AI wrote most of it.
In a Cornell and Stanford study published in Scientific Reports, people who believed their chat partner was using AI-suggested replies rated that partner as less cooperative, even though partners who actually used the suggestions came across as more cooperative.
Who sends it?
In the BetterUp Labs research, more than half admitted that at least some of their own work might be workslop. In the same report, managers received it more often than the people who report to them. The University of Florida study found people go easier on their own AI use than on everyone else’s.
What about the slop a company publishes?
Google can show it to fewer people. Google’s spam policies treat mass-produced, low-value pages as spam “no matter how it’s created,” and name generative AI first among the examples. After its March 2024 update, Google reported 45% less low-quality, unoriginal content in its results. Under its monetization policy, YouTube pays nothing for mass-produced videos made from generic templates.
Why does AI produce slop?
AI has no way to know who will read the work or what good looks like for them unless the prompt says so.
AI can raise the quality of work a lot. In a Harvard Business School working paper on a Boston Consulting Group experiment, consultants using AI on tasks it handles well produced work rated more than 40% higher and finished about a quarter faster. On a task outside what it handles well, they were less likely to get the right answer than colleagues working without it.
In an MIT Media Lab study led by Nataliya Kosmyna, essays written with ChatGPT came out similar to one another, and their writers reported feeling little ownership of them.
What makes it worse inside a company?
In our client work, it gets worse when people are told to use AI and nobody says what good output looks like.
We wrote about setting that standard in loops, not prompts. With the clients we work with, the part most often missing is a check on what goes out.
How software teams test AI output
In their guide to evals, Anthropic’s engineers describe how software teams test AI output before it ships and again after every change. They call these tests evals.
An eval is a test you run on AI output every time, built from examples of how it has failed before.
The guide says teams usually start building evals when users say the product feels worse after a change and the team is “flying blind,” and that a team with evals can judge a new model quickly.
Where does a good eval start?
In their evals FAQ, the evals practitioners Hamel Husain and Shreya Shankar call reading outputs and sorting their failures “the most important activity in evals,” and warn that generic quality scores create false confidence. Anthropic’s guide suggests starting with 20 to 50 tasks drawn from failures you have seen, beginning with the checks the team already does by hand.
Each check is a yes or no question. Husain and Shankar recommend pass or fail over a 1 to 5 rating, because a middle score hides what the rater was unsure about and means different things to different people. “Does the first paragraph tell the client what we need from them?” has a yes or no answer, and anyone checking it will give the same one.
Anthropic’s guide says checks that guard against old failures should pass close to every time, and that you only know your checks work if you read the outputs and the grades yourself.
Why isn’t a quick read enough?
In a 2023 study in Radiology, experienced radiologists reading mammograms with AI suggestions went from right about four times in five to right less than half the time when the suggestion was deliberately wrong.
Researchers at Microsoft Research and Carnegie Mellon found that office workers who trusted generative AI more thought less critically about its output. Psychologists call handing the thinking to a tool cognitive offloading.
Ask whoever reviews it a specific question, such as whether the numbers in the second paragraph match the table.
Researchers at LMU Munich found that when correcting an AI’s mistake took extra effort, people accepted more of its mistakes.
Does every piece of AI work need an eval?
For most office work, the check is reading the draft as its reader will before it goes. Work that goes out in volume, or without anyone reading it, needs more:
| Level | How it works |
|---|---|
| 0. Someone reads it | You or a named reviewer read every piece as its reader would; this suits one-off work like an email or a proposal |
| 1. Fixed rules | Checks for things that must always or never appear, such as a leftover placeholder or a price that is off the list, for anything built from a template |
| 2. Code checks | A script recomputes the numbers in a report or dashboard and compares them with the source data |
| 3. A model checks | A second model answers yes or no questions about known failures, for volume nobody can read in full |
| 4. You check the model | Someone marks a sample of the model’s answers and compares them with their own, so level 3 stays accurate |
A weekly client report that goes out with nobody reading it is the first place to add one.
How does an eval work on a client report?
We build AI systems for clients, and one of them drafts a weekly analytics report from the client’s own data. Someone on our side read every report before it went out. In July we ran an eval on that report for the first time, checking every past report we had against yes or no questions.
The first run found three problems the weekly read had missed. Internal shorthand from the client’s sales system had leaked into text the client would read. A lead count in the summary didn’t match the same count in the table. And the data file behind the report had changed shape, which had already broken the client’s dashboard once.
Each problem became a check. Does any internal term appear in client-facing text? Does every number in the summary match the table? Does the data file have the fields the report expects? A script runs all three in seconds, every week, before anyone reads the report.
Why use a second model as a checker?
Some questions need judgment, such as whether a finding is stated in terms the client uses. For those, a second model answers yes or no and quotes the line from the report that decided it, so whoever checks a sample can see the reason for each answer. We check a sample of its answers against our own, and when we disagree, we rewrite the model’s question to match our answer.
The weekly read now checks whether the report tells the client something useful to them.
What still gets through?
Thin documents get through. A check can’t tell whether a document says what this reader needed, so you or whoever knows the reader has to read it.
The Human Domain.
The assessment, five scored dimensions, and the interval controls that keep judgment yours.
On its way. Check your inbox, or download it now →
How does a company stop AI slop?
Start with ownership and build from your own failures.
What does a first checklist look like?
A first checklist is short and specific to one kind of document:
| Document | Example check |
|---|---|
| A client proposal | Does it name the goal the client gave us on the call? Is every price on the current price list? |
| A status update | Does the first line say what needs a decision this week? |
| A meeting summary | Are the decisions and their owners listed before the discussion? |
| A client email | Is the question we need answered in the first two sentences? |
| A quote or invoice | Does every line match the scope the client signed? |
| A job posting | Does it name the work the new hire will do in their first month? |
| A sales follow-up | Does it mention something the prospect told us on the call? |
| An investor update | Does every number match the latest finance report? |
Write yours from the failures you marked, and keep the list short enough that someone runs it every time.
What goes into an AI writing policy?
The sales software company Clay rolled out an AI writing policy across the company in August 2026, written by Sophie Alpert. Its central rule is that more time should go into writing a document than into reading it. It allows AI for brainstorming and drafting, and it says the author has to be able to defend every sentence.
A policy for your team can fit on one page. Say what AI may do, such as drafting and summarizing, and what may not go out without someone reading every line, such as anything with a price or a contract term in it. Name who owns the checklist for each kind of document. Say what happens when a document fails a check: it goes back to whoever sent it, and the failure goes on the list.
Keep it next to your general AI policy, which covers which tools people may use and what data may go into them. Our AI acceptable use policy template covers that part.
Who reads which draft, at which step, is covered in our guide to human in the loop AI.
How many examples do you need to start?
In their evals FAQ, Husain and Shankar suggest reading about 100 outputs and marking the failures in at least the first 30 yourself. In our client work, a month of one kind of document has been enough to show the pattern.
Which work should get checks first?
Start with work that goes out without anyone reading it, such as a proposal with a price in it or an assistant that talks to clients.
How do you know the checks are working?
Count how often each check fails. A check that fails every week points to a problem to fix at the source, in the prompt or the template. Keep a check that hasn’t failed in months as a guard, since Anthropic’s guide expects checks against old failures to pass close to every time.
Track one more number: the time your team spends fixing documents it receives. BetterUp’s survey measured workslop the same way, by asking people how long each instance took them.
What happens when the AI model changes?
Your output can change without notice. In a paper on how ChatGPT’s behavior changes, researchers at Stanford and Berkeley found that GPT-4 went from right most of the time on one task to right about half the time between its March and June 2023 versions. The researchers called for continuous monitoring of language models in use.
In 2025, Anthropic published a postmortem on three bugs that made some Claude responses worse for weeks, and said its own evals had not caught them. At the worst hour, about one in six requests to one Claude model went to the wrong servers.
Before a new model version touches client work, run last month’s questions on last month’s documents. We cover the rest of a model change in switching AI models.
Should you ban AI writing instead?
A ban fits work where your own voice is the point, such as a personal letter to a client.
When does it help to get outside help?
Get outside help when your team would be grading its own work.
An outside reader can go through the first batch of your team’s documents with you and leave you the list and the questions. Your team should end up owning the checks and running them without the outside firm. Our AI training work includes this.
Is all AI-written content slop?
AI can produce good work when you bring the context and do the final read. The Northeastern editors labeled some human-written text slop as well.
In a series of experiments by Oliver Schilke and Martin Reimann at the University of Arizona, people who said they used AI were trusted less, and people who got exposed by someone else lost more trust still. In October 2026, KTVU reported that Dartmouth’s provost had resigned after a faculty committee found he had used AI to change his writing in a meaningful way, including a newspaper column, and had not been forthcoming about it.
We call this approach Humans First AI: someone reads the draft and signs it before it goes.
Start with one document
Pick one document your team sent this week that AI helped with. Mark each place where you would have skimmed or had to ask a question; each mark becomes your first check. Run next week’s drafts against it and add every new catch.
Build Your Team’s Slop Checklist From Your Own Drafts
Paste three to five recent documents of one kind that AI helped write, such as client updates or proposals. You get back a short checklist drawn from your own failures.
Context: I run a team of [number] people. We use AI to help write [type of document], which goes to [who reads it] so they can [what the reader needs to do with it]. Below are [number] recent examples. Read them as that reader would, someone busy who needs to act on this.
Step 1. Mark the failures: For each document, list every place where the reader would skim or need to ask a question. Quote the exact sentence. Look hardest at relevance (does it matter to this reader?) and length (could it be said in fewer words?), then at style habits that make it sound machine-written.
Step 2. Group them: Sort the failures into types and count how often each type appears across the documents. Show the most common first.
Step 3. Write the checks: Turn each type into one yes or no question that two different people would answer the same way. Mark each question “person” if only someone who knows the reader can answer it, or “rule” if a text check could catch it.
Step 4. Test the checklist: Run it on two documents you did not use in step 1. Flag any question that two readers might answer differently, and rewrite it until they would agree.
Output: A checklist of five to ten yes or no questions, each with one quoted example from the documents and its “person” or “rule” label.
Deciding which work needs someone reading every piece, and keeping the checks current when your AI model changes, depends on what each workflow is worth to you. See where you stand →
Sources
- 2025 Word of the Year: Slop (Merriam-Webster, December 2025): word of the year, definition, “workslop” reports
- Word of the Year 2025: AI slop (Macquarie Dictionary, November 2025)
- Slop is the new name for unwanted AI-generated content (Simon Willison, May 2024)
- AI-Generated “Workslop” Is Destroying Productivity (Harvard Business Review, September 2025)
- Why People Create AI “Workslop” and How to Stop It (Niederhoffer, Robichaux and Hancock, Harvard Business Review, January 2026)
- Don’t Let AI Slop Muck Up Your Company’s Processes (Holweg and Davenport, Harvard Business Review, June 2026)
- Workslop (BetterUp Labs and Stanford Social Media Lab, September 2025): 40% prevalence, $186 per month, survey of 1,150
- The hidden costs of workslop (BetterUp, September 2025): trust and reputation findings, who sends and receives
- Workslop (Thomas Otter, September 2025): critique of the survey method
- Smart replies and perceptions of conversation partners (Hohenstein, Kizilcec, DiFranzo, Hancock, Jung et al., Scientific Reports, April 2023)
- Measuring AI “Slop” in Text (Shaib, Chakrabarty, Garcia-Olano and Wallace, arXiv, 2025, revised January 2026)
- Using AI to write at work could hurt your credibility (University of Florida, August 2025): Coman and Cardon, International Journal of Business Communication
- Being honest about using AI at work makes people trust you less (University of Arizona, May 2025): Schilke and Reimann, Organizational Behavior and Human Decision Processes
- How Americans View AI and Its Impact on People and Society (Pew Research Center, September 2025)
- Navigating the Jagged Technological Frontier (Dell’Acqua et al., Harvard Business School Working Paper 24-013, 2023)
- Spam policies for Google web search (Google Search Central): scaled content abuse
- New ways we’re tackling spammy, low-quality content on Search (Google, March 2024, results update April 2024)
- YouTube channel monetization policies (YouTube Help, updated July 2025): inauthentic content
- AI suggestions and radiologist performance in mammography (RSNA, May 2023): Dratsch et al., Radiology
- The Impact of Generative AI on Critical Thinking (Lee, Sarkar et al., Microsoft Research and Carnegie Mellon, CHI 2025)
- Correction costs and acceptance of AI suggestions (Beck, Eckman, Kern and Kreuter, LMU Munich, arXiv, 2025)
- Your Brain on ChatGPT (Kosmyna et al., MIT Media Lab, 2025, preprint)
- Shopify CEO Tobi Lütke on AI “slop grenades” (Fortune, September 2026)
- Dartmouth provost resigns after AI detector flags his writing (KTVU, October 2026)
- Clay’s AI writing policy (Clay, August 2026)
- Demystifying evals for AI agents (Anthropic, January 2026)
- Evals FAQ (Hamel Husain and Shreya Shankar, updated September 2026), including why binary pass or fail beats 1 to 5 ratings
- How Is ChatGPT’s Behavior Changing over Time? (Chen, Zaharia and Zou, arXiv, revised October 2023)
- A postmortem of three recent issues (Anthropic, September 2025)
- OpenSEO keyword data for “ai slop”, US monthly searches, pulled November 2026
Frequently Asked Questions
Where does the term AI slop come from?
Merriam-Webster says “slop” originally meant soft mud and later food waste. In 2024 the developer Simon Willison helped popularize it for unwanted AI content, crediting an online account called deepfates, and in 2025 Merriam-Webster and Australia’s Macquarie Dictionary both chose it as their word of the year.
What is the difference between AI slop and workslop?
AI slop is the general term for low-quality AI content. Workslop is the workplace version, named by BetterUp Labs and Stanford researchers in 2025: AI-generated work that looks finished but lacks the substance to move a task forward, so a colleague has to redo it.
Can AI detectors spot AI slop?
In a Northeastern University study, AI models asked to judge slop matched human editors’ calls about as often as guessing would.
Should you tell clients when AI helped with their work?
Research by Schilke and Reimann found that people who disclose AI use are trusted less, and that people exposed by someone else lose more trust than people who disclose it themselves.
What is an AI eval, in plain terms?
An eval is a repeatable check on AI output. It uses examples of past work, a written list of the ways the output has failed before, and a yes or no question for each failure, run every time the work is produced or the setup changes.
How many examples do you need to start checking AI output?
Anthropic suggests starting an eval set with 20 to 50 tasks drawn from failures you have seen, and the practitioners Hamel Husain and Shreya Shankar suggest reading about 100 outputs and marking the failures in the first 30 yourself.
Will a better AI model fix slop?
A better model still has to be told who the reader is and what good looks like, and what goes out still needs a check.
Does a small team need an AI writing policy?
A one-page policy is enough for most small teams. It should say what AI may draft, what needs a full read before it goes out, and who keeps the checklist for each kind of document. Clay’s policy adds one rule: more time should go into writing a document than into reading it.
Who should own AI quality on a team?
Whoever sends a piece owns it. For recurring work, name an owner to keep the checklist current and rerun it when the model or the prompt changes.


