# Looks Amazing! But Is It Really? AI makes it easy to send more: longer recaps, bigger plans, polished docs nobody checked. What the research says it costs, and how to send less but better. By Steven Cappiello, Founder of JobKick · Published 2026-10-22 Canonical: https://jobkick.com/blog/ai-recap-nobody-reads/ Topics: workslop, AI meeting notes, AI generated documents, information overload, AI at work, team communication The meeting ends at 10:30. At 10:31 the recap lands in everyone's inbox. It's long: a summary, eleven action items, a "risks and mitigations" section. Half the action items have no owner. One of the risks never came up. Nobody corrects anything, because nobody reads past the first screen. That afternoon a teammate sends a plan. You asked for a paragraph; you got twelve pages, with a timeline and a table. It looks amazing. Then you notice page 3 says the launch is in March and page 9 says the second quarter. When you ask which is right, they aren't sure. They didn't write either one. Nobody here is lazy on purpose. AI made it free to produce more, so we produce more. We all do it, me included. But more words isn't more said, and somebody downstream pays for the difference. ## It has a name now: workslop In September 2025, researchers at BetterUp Labs and the Stanford Social Media Lab gave this a name in *Harvard Business Review*. They call it **workslop**: "AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task." The key word is *masquerades*. Workslop doesn't look sloppy. It looks finished. And "it transfers the effort from creator to receiver": the sender saves an hour, and the reader spends two working out what it means, checking it, and often redoing it. They surveyed 1,150 full-time US employees across industries: - **40%** received workslop in the last month (source: HBR · BetterUp Labs & Stanford, 2025 · 1,150 US workers, https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity) - **1 hr 56 min** spent dealing with each instance, on average (source: HBR · BetterUp Labs & Stanford, 2025, https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity) - **15.4%** of the content they receive at work, by their own estimate (source: HBR · BetterUp Labs & Stanford, 2025, https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity) The researchers put the cost at about $186 per employee per month, more than $9 million a year for a company of 10,000. That's built from people's own guesses about time and salary, so take it as a size, not a bill. And time isn't the worst part. Trust is. **Chart (bars): What receiving workslop does** — Share of people who had received workslop who said each of these. Hover a bar for the detail. ```json {"unit":"%","highlight":"Trust the sender less","rows":[{"label":"Felt annoyed","value":53},{"label":"Trust the sender less","value":42,"note":"Saw the sender as less trustworthy than before"},{"label":"Felt confused","value":38},{"label":"See the sender as less intelligent","value":37},{"label":"Told a teammate or manager about it","value":34},{"label":"Less likely to want to work with them again","value":32},{"label":"Felt offended","value":22}]} ``` Source: Niederhoffer et al., "AI-Generated 'Workslop' Is Destroying Productivity," Harvard Business Review (Sept. 2025) (https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity); ongoing survey of 1,150 US full-time employees. You can save an hour with AI and spend a reputation doing it. ## We've been warned. We don't check anyway. Every AI tool tells you to check its work. Most of us skim it, think "looks amazing," and hit send. The University of Melbourne and KPMG surveyed 48,340 people in 47 countries in early 2025. Here's what employees told them: **Chart (bars): How employees say they use AI at work** — Share of employees in the 2025 global study, as reported by KPMG. ```json {"unit":"%","highlight":"Rely on AI output without checking its accuracy","rows":[{"label":"Rely on AI output without checking its accuracy","value":66},{"label":"Intentionally use AI at work","value":58},{"label":"Hide their AI use and present its work as their own","value":57},{"label":"Have made mistakes in their work because of AI","value":56}]} ``` Source: KPMG & University of Melbourne, Trust, attitudes and use of AI: a global study (2025) (https://kpmg.com/xx/en/media/press-releases/2025/04/trust-of-ai-remains-a-critical-challenge.html); 48,340 people, 47 countries. If most people pass AI work along as their own, the reader can't tell which parts to double-check. The warning label is on the tool, not on the document. Why does checking slip? Researchers at Microsoft and Carnegie Mellon surveyed 319 knowledge workers about 936 real tasks. The more people trusted the AI, the less critical thinking they reported. The more they trusted *their own* skill, the more they checked. So AI gets checked least where people know the subject least, which is where it needs checking most. At scale, it looks like this: in 2025, Deloitte Australia agreed to refund part of its A $440,000 fee to the Australian government after a 237-page report turned out to cite research that doesn't exist and a court quote the judgment doesn't contain. A University of Sydney researcher, Chris Rudge, called it "full of fabricated references." The revised version disclosed that Azure OpenAI was used. Courts see it too: Damien Charlotin's public database lists 2,202 legal decisions where a court found or suggested a party relied on AI-hallucinated material. ## "I didn't really say that" The part I keep coming back to: documents that say things the sender wouldn't have said, and doesn't fully understand. ![A man in a navy suit and polka-dot tie adjusts his jacket. His face is out of frame.](/blog/ai-recap-nobody-reads/confident-suit.webp) *Polished, buttoned up and confident. It looks amazing. So does workslop. Photo: rawpixel, CC0 public domain (via [Openverse](https://openverse.org/image/28d70b02-f845-4d98-aa97-f8b8acf1733c)).* In a small MIT Media Lab study, 54 people wrote essays with ChatGPT, a search engine, or nothing. The ChatGPT group "struggled to accurately quote their own work" and felt the least ownership of it. It's a small preprint, so treat it as a signal, not a verdict. But it matches what you see when you ask someone about their AI-written plan: they know roughly what it says, not what it commits them to. Ideas suffer too. In a Boston Consulting Group experiment with 758 of its consultants, those using GPT-4 on a creative task scored about 40% higher, but their ideas were 41% less diverse, and on a business problem the AI handled poorly they did 23% *worse* than people working alone. The polish was the same either way. Only the substance changed. ## So what about the meeting recap? The recap is where this hits the [Kickoff Cliff](/blog/the-kickoff-cliff/). The promise is that everyone stays in the loop. In practice it often goes the other way. - **Nobody reads it.** Microsoft's telemetry says the average worker gets 117 emails a day, most skimmed in under a minute. A long recap is one more. - **People skip the meeting because of it.** In a survey of 1,000 full-time workers by the review site Software Finder, nearly 3 in 10 had skipped a meeting counting on the AI summary, and 41% had missed context it didn't capture. Few details on method, so take it as a rough read. - **It gets "who does what" wrong.** When The Oasis Group tested six AI notetakers on one scripted meeting, most transcribed it perfectly, but action-item accuracy ranged from 62% to 87%. Hearing every word isn't knowing who owns what. Nobody owns the recap, nobody corrects it, and three weeks later two people are working from two versions of what was decided. It's the cc pile, with better formatting. ## In fairness None of this means AI is bad at work. The same BCG study found real gains where AI is strong, and a clean recap for someone who missed the meeting is a gift. As the workslop researchers put it, "There has always been sloppy work." AI didn't invent the padded report. It made it free. A lot of it is pressure, not laziness: "indiscriminate imperatives yield indiscriminate usage," the HBR authors write. Tell people to use AI for everything and they will, including what it's bad at. ## What to do instead 1. **Only send what you'd sign.** If you can't explain a sentence, or wouldn't say it out loud, cut it. You're the author, whoever typed it. 2. **Use AI to make things shorter.** If the AI version is longer than yours would be, ask whether the length helps the reader or only looks thorough. 3. **Check what people act on.** Names, numbers, dates, quotes, sources, owners. A wrong adjective costs nothing. A wrong date costs a lot. 4. **Turn the recap into decisions.** What was decided, who owns each next step, by when. Post it where the project lives, not in twelve inboxes. 5. **Say how you used it.** "AI drafted this, I checked the numbers" tells the reader where to look. With 57% hiding their AI use, it stands out. 6. **When you get workslop, ask one kind question.** "What do you need me to do with this?" usually gets you the paragraph that should have been sent. ## What's the longest AI document you've received this year? Did anyone read it? Have you sent one yourself? Tell me in the comments, especially if your team has a rule for AI recaps that works. ## Sources - Niederhoffer, Rosen Kellerman, Lee, Liebscher, Rapuano & Hancock, "[AI-Generated 'Workslop' Is Destroying Productivity](https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity)," *Harvard Business Review*, Sept. 2025 (BetterUp Labs & Stanford Social Media Lab; 1,150 US full-time employees). - KPMG & University of Melbourne, [Trust, attitudes and use of AI: a global study 2025](https://kpmg.com/xx/en/media/press-releases/2025/04/trust-of-ai-remains-a-critical-challenge.html) (48,340 people, 47 countries, Nov 2024 – Jan 2025). - Lee et al., "[The Impact of Generative AI on Critical Thinking](https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/)," CHI 2025 (Microsoft Research & Carnegie Mellon; 319 knowledge workers, 936 examples). - Associated Press, "[Deloitte to partially refund Australian government for report with apparent AI-generated errors](https://www.bnnbloomberg.ca/business/artificial-intelligence/2025/10/07/deloitte-to-partially-refund-australian-government-for-report-with-apparent-ai-generated-errors/)," Oct. 2025. - Damien Charlotin, [AI Hallucination Cases database](https://www.damiencharlotin.com/hallucinations/) (2,202 cases as of Oct. 21, 2026). - Kosmyna et al., "[Your Brain on ChatGPT](https://arxiv.org/abs/2506.08872)," MIT Media Lab preprint, 2025 (54 participants). - Boston Consulting Group, "[How People Can Create and Destroy Value with Generative AI](https://www.bcg.com/publications/2023/how-people-create-and-destroy-value-with-gen-ai)," 2023 (758 consultants). - Microsoft WorkLab, "[Breaking down the infinite workday](https://www.microsoft.com/en-us/worklab/work-trend-index/breaking-down-infinite-workday)," Work Trend Index, June 2025. - Software Finder, "[AI note-taking at work](https://softwarefinder.com/resources/ai-note-taking-tools)," survey of 1,000 full-time professionals (site survey; method not detailed). - Kitces, "[The Risks of AI Meeting Notetakers](https://www.kitces.com/blog/artificial-intelligence-ai-meeting-notetakers-note-tools-client-communication-accuracy-domentation/)," Aug. 2025, reporting The Oasis Group's test of six notetakers. - Photo: "Confident businessman," [rawpixel](https://www.rawpixel.com/image/9080003/confident-businessman-desktop-wallpaper), CC0 1.0 public domain.