Arguments about AI writing usually run on anecdotes. This page sticks to numbers. Every statistic below comes from a named study with a link, and each one has its own anchor, so you can cite a single stat without sending someone the whole page. We update it as new research is published.
The productivity bill
#40% of US desk workers received workslop from a colleague in the month before Stanford and BetterUp surveyed 1,150 of them in September 2025. The researchers define workslop as AI content that masquerades as good work but lacks the substance to move a task forward.
#1h 56m is the average time a recipient spends dealing with each piece of it, per the same study, in Harvard Business Review.
#85% of employees say AI saves them one to seven hours a week, per Workday’s global research from January 2026.
#~40% of that saved time goes straight back into fixing and verifying machine output.
#14% of employees consistently come out ahead on AI once rework is counted. The other 86% break even or lose time.
#19% slower is how experienced open-source developers performed with AI assistance in METR’s 2025 controlled study. They had expected to be 24% faster. Feeling faster is not the same as being faster.
#66% of 31,476 developers told Stack Overflow’s 2025 survey that “almost right” AI answers are their top frustration, and 45% said debugging AI-generated code is more time-consuming.
The credibility price
#42% of people who receive workslop trust the sender less afterward.
#~50% of recipients see the sender as less capable than before.
#1 in 3 recipients, nearly, would rather not work with that person again.
The scale of it
#13.5% of 2024’s biomedical abstracts, at minimum, carry the stylistic fingerprints of LLM assistance, according to Dmitry Kobak’s team, which analyzed more than 15 million abstracts in Science Advances.
#77 / 100 is the average slop score of an unedited AI draft on the open-source Zero Slop meter, where known human writing scores 9 to 21 and readers start noticing at 25.
#24.8% of 7,627 documents read as slop in Zero Slop’s own public RAID-derived benchmark, roughly one passage in four, across four current model families. Our measurement, on public data anyone can rerun.
#6 million generations in the RAID benchmark showed that AI-text detectors collapse under light paraphrasing. You cannot reliably prove who wrote a text, but you can measure how it reads.
The platforms respond
#1 million people used LinkedIn’s “seems like AI slop” report button within weeks of its July 2026 launch.
#100,000s of automated comment attempts get blocked by LinkedIn every day, by the company’s own account.
What the tools still deliver
#65% of employees at companies that adopted AI report positive impact, per Gallup’s May 2026 data. The tools work when people edit the output instead of shipping the first draft.
#15.4 is the mean slop score after the same 18 AI drafts went through the Zero Slop editorial pipeline, down from 76.3, with every name, number, quote, and link kept. That gap, 76.3 to 15.4, is the measurable difference between shipping the first draft and editing it.
Cite anything here. Each stat’s anchor link copies from the section marker, and the sources are linked in place. If a newer study changes a number, tell us through the feedback form and we ship the correction.