AI slop and workslop

What is AI slop and workslop?

AI slop is content that a generative AI tool produced in bulk and that nobody checked before it went out: blog posts written to fill a site, product descriptions copied across hundreds of pages, images and videos that look almost right, summaries that say nothing. The content exists because producing it cost nothing, and nobody stopped to ask whether anyone needed it.

Merriam-Webster picked "slop" as its word of the year for 2025 and defines it as digital content of low quality, usually produced in quantity by means of artificial intelligence. The dictionary's own explanation of the choice lists absurd videos, fake news that looks real, junky AI-written books and workslop reports that waste colleagues' time.

Workslop is the version you meet inside a company. The word comes from a September 2025 article in Harvard Business Review by researchers at BetterUp Labs and the Stanford Social Media Lab. They describe it as AI-generated work that looks like good work but lacks the substance to move the task forward. The receiver has to figure out what it means, check it, and often redo it. The effort did not disappear, it moved from sender to receiver.

The test is simple: did a person look at this and decide it was good enough to send? If not, it is slop, no matter how polished it reads.

What workslop looks like in an SME

Three situations that come up again and again.

The ten-page memo nobody asked for
A colleague is asked for a recommendation on a new supplier. Instead of two paragraphs with a preference and a reason, the team gets ten pages with a market overview, a SWOT and a risk matrix. The recommendation is buried on page eight and reads as "both options have merits". Everyone who reads it loses half an hour and still has to ask the original question again.

The generated report with wrong totals
Someone pastes a sales export into a chatbot and asks for a quarterly summary. The result looks like a finished report, with headings and a conclusion. Two of the totals do not match the underlying rows, because the model estimated them instead of calculating. The report goes to the management meeting, a director notices that the regional figures do not add up to the total, and from then on every number in that report is suspect.

The quote the customer recognises
A sales rep lets an assistant draft a quote from a rough prompt. The text is fluent and generic, mentions a delivery term that does not match the company's actual conditions and uses phrasing the customer has already seen in three other AI-written emails that week. The customer replies asking whether anyone actually read their request.

The common thread: the output is presented as finished, but the checking, the deciding and the fixing all still have to happen, and they now happen on someone else's desk.

How workslop arises

Generative AI made producing text cheap and left checking as expensive as it always was. Writing a ten-page memo used to take a day, and that day forced the writer to decide what mattered. Now it takes two minutes, and the reader gets the day's work instead.

Incentives push the same way. In many teams visible output counts: the long document, the fast reply, the deck with forty slides. Nobody measures the hours a receiver spent decoding it. Pressure from management to "use AI more" comes on top. The BetterUp and Stanford researchers followed up in January 2026 with a second HBR piece on why people send workslop. Their split is useful: some people use AI as a pilot, to think better and work with colleagues, and others use it as a passenger, to get a task off their plate. The passengers produce most of the slop.

There is also a knowledge gap. Someone who does not know that a language model estimates rather than calculates, or that it fills gaps with plausible text, has no reason to check the totals. That is the practical link with AI literacy: the habit of checking only exists if you know what can go wrong.

What it costs

Rework
In the survey behind the HBR article, about four in ten US desk workers said they had received workslop in the past month, and they spent about two hours per instance sorting it out. Multiply that across a team and the time the tool saved disappears into the checking.

Trust between colleagues
The same survey found that a large share of receivers thought less of the sender afterwards: less capable, less reliable, less trustworthy, and a good number said they would rather not work with that person again. That is a high price for saving twenty minutes on a memo.

Customer trust
Customers notice. A quote with a wrong condition, a support answer that misses the actual question, a newsletter that reads like every other newsletter: each one tells the customer that nobody on your side paid attention. That impression lasts longer than the content.

Search ranking
For published content, Google's Search Central guidance says that using automation, including AI generation, to produce content primarily to manipulate search rankings violates its spam policies. Those spam policies name "scaled content abuse" explicitly, with "using generative AI tools to generate many pages without adding value for users" as the first example. Google does not punish AI content as such. It punishes content made for the ranking rather than for the reader, and bulk-generated pages are exactly that.

AI slop versus hallucination

People sometimes use the two words for the same thing, but they describe different failures.

A hallucination is a model stating something that is false: an invented source, a wrong date, a total that does not exist. The problem sits in the content. A careful reader with the right sources can catch it.

Slop describes the process, whatever the truth of any single sentence. A piece of workslop can be entirely accurate and still be slop, because nobody decided it was needed, nobody cut it to what mattered, and nobody checked it before sending. The receiver cannot tell which parts were considered and which parts were generated, so they have to treat all of it as unchecked.

Put differently: a hallucination is wrong, slop is unreviewed. A hallucination inside a document that someone reviewed and missed is an error. The same hallucination inside a document that nobody read before sending is slop, and the reader would be right to distrust the rest of it too.

Rules for a team

1. Say when AI was used. A line like "first draft from Copilot, I checked the figures against the export" tells the receiver what has been verified and what has not. Hidden AI use is what turns a normal mistake into a trust problem.

2. The sender checks before sending. Whoever sends it owns it. Numbers are recomputed against the source, names and conditions are checked, and the sender can defend every claim in the meeting. If you cannot explain a paragraph, cut it.

3. Short beats long. Ask what the receiver needs in order to decide, then send that. A recommendation with one reason beats a ten-page analysis every time. Length is no longer a sign of effort, so stop rewarding it.

4. AI output is a draft, not a deliverable. Use the tool for the first version, the structure, an alternative phrasing. The version that leaves your hands has been read, cut and corrected by you.

5. Managers go first. A manager who forwards unread AI summaries teaches the team that this is fine. A manager who sends a bloated memo back with "what do you recommend, in five lines?" teaches the opposite.

These rules fit on one page and belong in the same document as your AI literacy basics: which tools are approved, which data stays out, and who checks what before it leaves the company.