You open the document. It is well formatted, the headings are confident, the summary reads smoothly, and somewhere around the second paragraph you realise it does not say anything. You cannot use it, you cannot quite reject it, and you now have to work out what the sender meant before you can do the thing the document was supposed to do.
Since September 2025 that experience has had a name. Workslop is AI-generated work content that looks like good work but lacks the substance to move a task forward, so the effort transfers from the person who produced it to the person who received it. The term was coined by researchers at BetterUp Labs and the Stanford Social Media Lab in a Harvard Business Review article, and it spread fast because the experience was already universal. The numbers attached to it deserve a closer look than they usually get, because they come from one survey, run by a company that sells a fix, and they have not been replicated.
Where the word came from
The HBR piece, by Kate Niederhoffer, Gabriella Rosen Kellerman, Angela Lee, Alex Liebscher, Kristina Rapuano and Jeffrey T. Hancock, opened with a contradiction: AI use at work had doubled since 2023 and the number of companies with fully AI-led processes had nearly doubled in a year, yet an MIT Media Lab report found 95% of organisations see no measurable return on their investment. The authors proposed workslop as one explanation.
Their definition is the one to use: AI-generated work content that masquerades as good work but lacks the substance to meaningfully advance a given task. The distinguishing feature is not that the content is wrong, though it can be. It is that it shifts the burden downstream, requiring the receiver to interpret, correct or redo it. On social media the same phenomenon is called AI slop; workslop is the version that arrives with your name in the “to” field.
The survey numbers, and their limits
The figures everyone quotes come from an online survey of full-time US desk workers run by BetterUp in September 2025 with the Stanford lab. The HBR article and BetterUp’s research page give the sample as 1,150 respondents. Of those, 40% reported receiving workslop in the previous month (HBR’s summary rounds it to 41%). Respondents who had received it estimated that 15.4% of the content they receive at work qualified. It mostly flowed between peers (40%), with 18% sent upward from direct reports to managers.
The time cost is the headline. Respondents reported spending an average of one hour and 56 minutes dealing with each instance. Combined with self-reported salaries, the researchers put the cost at $186 per employee per month, and for a 10,000-person organisation at over $9 million a year in lost productivity.
The social cost came from the same survey. Over half of respondents, 53%, said receiving workslop left them annoyed, and 22% said offended. Around 42% said they saw the sender as less trustworthy afterwards, and about half viewed senders as less creative, capable and reliable.
Now the limits. Every figure is self-reported, including the time estimate that drives the dollar figure; nobody timed anyone. The survey defined the phenomenon and asked people whether they recognised it, which reliably produces high recognition rates for a vivid description. The sample size is reported inconsistently: BetterUp’s own blog describes the September 2025 survey as 1,004 desk workers, while HBR and the BetterUp research page say 1,150, which suggests the survey was still running when the article went out. And the organisation that ran it sells coaching for the problem it identified. None of this makes the numbers wrong. It makes them a single vendor-run survey rather than a measurement, and as of this writing no independent study has reproduced them.
Why it happens
The mechanism is simpler than the name. Generating a plausible document now costs seconds, and reviewing one still costs the same as it always did. When the effort to produce something falls to near zero and the effort to check it does not, the rational move for a rushed sender is to generate and forward, and the cost lands on whoever reads it next.
The best experimental evidence for that review burden comes from software rather than documents. A 2025 randomised controlled trial by METR had 16 experienced open-source developers complete 246 real tasks in projects they knew well, each task randomly assigned to allow or forbid AI tools. Developers had forecast that AI would cut their completion time by 24%. When allowed to use it, they took 19% longer. The authors are careful to call this a snapshot of early-2025 tools in one setting, but the direction fits the workslop account: generated output that looks finished still has to be understood and verified by someone, and that someone was not saved any time.
The other ingredient is mandate without review. The HBR authors argue that leaders pushing indiscriminate AI adoption with little guidance on quality standards are the ones creating the conditions for workslop. Speed is measured; the cleanup is invisible.
Trust and the productivity paradox
The 95% figure that anchors the debate is worth placing accurately. It comes from a 2025 MIT report on enterprise AI based on 150 executive interviews, a survey of 350 employees and an analysis of 300 public deployments, and it refers to organisations seeing no measurable return, which is not the same as individuals seeing no benefit. Workslop is one candidate explanation for the gap between the two: individual time saved on production, collective time lost on reception.
The trust finding is the part most likely to outlast the numbers. If 42% of people rate a colleague as less trustworthy after one piece of workslop, the durable cost is not the two hours. It is that the next document from that person gets read with suspicion, which is its own reviewer burden, applied to everything they send.
What good AI use looks like by contrast
Nothing in the workslop research argues against using the tools. It argues against forwarding their output unread. The distinction the HBR authors draw is between employees who use AI to polish work they have already thought through and employees who use it to produce content in place of thinking. The receiving end tells them apart within a paragraph.
The practical test follows from the definition. Before sending, ask whether the document advances the task for the person receiving it, or whether it hands them the job of finding out. If you could not answer a question about its contents without reopening the chat window that made it, it is not finished. The time you would save by skipping that check is the time your colleague will spend, plus the trust you will not get back.
Common questions
What does workslop mean?
Workslop is AI-generated work content that appears polished and complete but lacks the substance to advance the task, so that the receiver has to interpret, correct or redo it. The term was coined in a September 2025 Harvard Business Review article by researchers from BetterUp Labs and the Stanford Social Media Lab.
How much time does workslop cost?
The originating survey of US desk workers reported an average of one hour and 56 minutes per instance, which the researchers converted to an invisible tax of $186 per employee per month and over $9 million a year for a 10,000-person company. All of those are self-reported estimates from a single vendor-run survey and have not been independently measured.
How common is workslop?
In the BetterUp and Stanford survey, 40% of respondents said they had received it in the previous month, and those who had estimated that 15.4% of what they receive qualifies. The sample is reported as 1,150 in HBR and 1,004 on BetterUp’s own blog. No second survey has yet reproduced the figures.
Does AI actually make people less productive?
Sometimes, in measured settings. A 2025 randomised trial of 16 experienced developers on 246 real tasks found they took 19% longer with AI tools despite forecasting a 24% speed-up. That is one context with early-2025 tools, but it shows the review cost of generated output is real and easy to underestimate, which is the mechanism behind workslop.
The bottom line
Workslop is a precise name for a real thing: output that looks like work and transfers the work to you. The definition is sound, the mechanism is well supported by the one controlled experiment on review burden, and the survey numbers should be quoted with their provenance attached, because they are self-reported, inconsistently sized and unreplicated. The response that follows from the evidence is not to use AI less. It is to never send what you have not read, because the receiver will know within a paragraph, and the cost of being known for it lasts longer than the two hours.
Did this answer your question?
Thanks. That helps us decide what to fix.
Leave a comment