AI slop is machine-generated content that looks finished but gives the reader little to use. The sentences may be fluent. The piece as a whole is vague, repetitive, or empty. At work, the problem has its own name: workslop, a term from Stanford’s Social Media Lab and BetterUp Labs for AI work that leaves the person receiving it to finish the task.
Is AI slop the same as AI-assisted writing?
No. AI can help produce clear, useful writing. Slop is what remains when nobody checks the claim, removes the padding, or takes responsibility for the final draft.
The distinction is usefulness, not authorship. In Gallup’s May 2026 survey, 65% of US employees at organizations that had implemented AI reported a positive effect on their productivity and efficiency. That is a reported effect, not a timed test. An unchecked draft can leave the next person with the work.
What does AI slop look like?
Common signs include repeated binary contrasts (“It’s not X. It’s Y.”) and throat-clearing openers (“Here’s the thing”). Faux-insight setups (“what nobody tells you”) flatter the writer as the only person who noticed. A dramatic colon can promise an insight without supplying one. Describing a launch as one that “marks a pivotal moment” adds ceremony where the fact could be that it is the company’s first paid product.
Vague attribution uses “studies show” in place of a citation. Unnecessary synonym changes call the same thing an agent, an assistant, and a tool in successive sentences.
A project that collected discussions from 47 subreddits found many of the same complaints, led by em-dash overuse, flat rhythm, and formulaic contrasts. Its sample and LLM-led review make the ranking exploratory, but the complaints are useful prompts for editing. Several concern the whole document and cannot be found with a word list. We examined how our meter scored against that ranking.
Can you measure AI slop?
Some parts are measurable. AI-assisted prose can overuse certain words; researchers traced their use rising across 15 million scientific abstracts after ChatGPT launched. Sentence variety, repeated templates, and excess formatting can also be counted. Zero Slop combines those checks in a 0-to-100 score. In its shipped test sets, the human references score from 9 to 21 and deliberately obvious AI drafts average 76.3. Twenty-five is Zero Slop’s review threshold. It is a product setting, not a proven boundary for human perception.
What does AI slop cost?
In BetterUp and Stanford’s September 2025 survey of 1,150 U.S. desk workers, recipients of workslop recalled spending nearly two hours repairing each incident on average. Forty-two percent trusted the sender less afterward. These are respondents’ estimates and reactions, not timed observations.
A separate Workday study surveyed 3,200 active AI users at large employers in November 2025. It estimated that rework absorbed nearly 40% of their reported time savings across workplace tasks, not writing alone.
LinkedIn added a button for reporting it, and over a million people have already used it.
In the Stanford and BetterUp survey, half of recipients saw the sender as less capable and almost a third said they would rather not work with that person again. Those are reported reactions, not observed career outcomes, but they show why weak AI work is also a trust problem.
How do you fix AI slop?
Review the draft before sharing it. Check the claims, cut sentences that carry no information, and read the result aloud. That workflow is available as a free, open-source skill, and the sources for the figures on this page are linked from our post The Slop Tax. See also the full anti-slop toolkit and the four-editor replay and its limits.