Workslop: The AI Work Nobody Reads

Reading Time: 5 minutes

Sach batao: when did you last read a ten-page document a model wrote?

“Workslop” is AI-generated content that looks good but lacks substance, in the words of BetterUp, the company behind the research that named it. The writer saves an hour. The reader loses two. The cost has not disappeared; it has moved to the other side of the inbox.

This post covers what the research found, what five companies are doing about it, and a three-question test you can run before you hit send.

🧾 Where the Word Came From

The term comes from a Harvard Business Review article published on 22 September 2025, written by researchers at BetterUp Labs and the Stanford Social Media Lab. (Harvard Business Review) I could read the title, authors and date of that article. I could not read the full text, so the figures below come from BetterUp’s own page on the research.

According to BetterUp, the team surveyed 1,150 full-time US desk workers in September 2025. (BetterUp)

  • 40% said they had received workslop in the last month.
  • It took about 2 hours on average to deal with each incident.
  • BetterUp puts the cost at $186 per employee per month.
  • It also gives $9 million a year for a 10,000-person company.

Two cautions. First, this was an online survey, so the time and cost figures are what people reported, not what anyone measured. Second, the $9 million does not follow from the $186 on its face: $186 a month across 10,000 people for 12 months is about $22 million. The page does not explain the gap. One reading is that the $9 million counts only the 40% who received workslop, which would land near that figure, but that is my guess, not BetterUp’s statement. I would treat the $186 as the number to remember and the $9 million as unreconciled.

πŸ“Œ Five Companies Writing Rules

πŸ“Œ Real-world case: A Business Insider piece dated 5 November 2026 describes how several small companies have started writing down what they expect. I read it as reprinted by DNYUZ, so everything below is as reported there. (Business Insider, via DNYUZ)

  • Twill, a ten-person tech recruiting firm. Its founder, Michelle Volberg, noticed staff pasting model output into Slack and email, with messages that lacked client context and contained errors. Her rule: use AI to brainstorm, but write internal messages in your own voice.
  • Clay. Cofounder Varun Anand posted a writing policy in August. It is not a ban. Staff must stand behind every sentence, and making a document should take longer than reading it.
  • ZocDoc. Guidance issued in August tells people not to ship their slop and to be open about when a bot is involved.
  • Spellbook. CEO Scott Stevenson told his team in early September that he would rather read five rough bullets than five polished pages, and that customer messages should be typed by hand.
  • FischTank PR. Founder Eric Fischgrund bans AI-written content for reporters and clients and uses an AI detector to check. He says he let go of one worker who kept breaking the rule.

Look at the pattern. None of these is a ban on AI. Each rule is about ownership: if you send it, it is yours. That is a smaller, more sensible ask than “do not use the tool.”

πŸ” Why It Happens

Generating text costs almost nothing. Reading it costs a person’s time. When one side of that trade is free, volume goes up.

There is a second push. In Episode 3, Tokenmaxxing Inc. I wrote about what happens when a company measures AI use and rewards it: people produce more of the thing being counted. Mandate the usage and you may get more output without more work. Workslop is the reader’s side of that same story.

BetterUp’s recommendations point the same way. It suggests clear guardrails for AI use, leaders who show what thoughtful use looks like, and a “pilot” mindset where AI helps people collaborate instead of helping them avoid the work. I am not claiming any particular model or vendor causes this. The research is about how people use the tools.

πŸͺ™ Token Truth

Output is cheap. Attention is not.

Tokens price the writing. A vendor sends an invoice for every one. Nobody sends an invoice for the reading, and that is where BetterUp’s two hours and $186 sit. (For more on how tokens are priced and counted, start with Episode 1 of The Token Diaries.)

Toh phir, if a cost is real but never billed, who notices it? Usually the person reading.

πŸ§ͺ What It Looks Like: A Made-Up Example

Here is an illustration I invented, not a quote from any company. Say a teammate asks whether to renew a software contract.

The workslop version opens with three paragraphs on why software renewals matter, lists seven general “factors to consider,” and ends with “ultimately, the decision depends on your needs.” It is polite, well formatted and 400 words long. It answers nothing.

The useful version is three lines: “Renew, but only for one year. We used 40 of the 100 seats last quarter. Ask them to cut the price; the call is due Friday.”

The second one took a person five minutes of thinking. The first took a model five seconds, and costs the reader far more than five minutes to dig through. That is the whole problem in two paragraphs.

Notice what the useful version has: a decision, a fact from your own work, and a next step. A model cannot supply the fact from your own work unless you give it, and it cannot decide for you. That is the part that has to be yours.

βœ… The Three-Question Test Before You Send

I would run these on anything an AI helped write, before it leaves your outbox.

1. Could the reader act on this without asking me a question?
If the document needs a follow-up call to explain what it means, it moved work onto the reader. A good summary says what you need from them, by when.

2. Would I say it this way to their face?
This is Twill’s idea in my words. If the tone would sound odd across a table, it will sound odd in a message. Odd messages are the first sign that nobody wrote them.

3. Can I stand behind every claim, including the ones I did not write?
Models can state things that are not true with a calm voice. If a number or a name is in there, you are the one who sent it. Check it, or cut it.

If a draft fails any of the three, shorten it, fix it or delete it. A short note you wrote beats a long one you did not read.

πŸ§‘β€πŸ’Ό For Managers: Two Lines to Borrow

You do not need a 20-page AI policy. Two lines, adapted from the examples above, cover a lot:

  • Internal messages are in your own voice. AI can help you think, not speak for you.
  • Short first. If it takes longer to read than it took to make, it is too long.

Put them where people will see them, and say that you follow them too. Seedhi baat: a rule the manager ignores is a rule nobody follows.

🚧 What I Could Not Confirm

I could not read the full HBR article or open the Business Insider original, so the company examples are as reported in a reprint. The survey is self-reported. And I have no evidence on how effective any of these rules are; they are early and small-company, and I have not seen results.

None of this is HR or legal advice. It is a reading of one study and one news report.

Think of the last long AI-written message you received. Did you read all of it, and if not, what did you do instead?

πŸ“š References

What’s your Reaction?
+1
0
+1
0
+1
0
+1
0
+1
0
+1
0
+1
0