Workslop: the hidden cost of untrained AI use

40% of US desk workers say they received "workslop" in the last month: AI-generated content that looks good but lacks substance. Each instance took about two hours to sort out (BetterUp Labs & Stanford, 2025). Workslop is what untrained AI use produces. It moves the work from the person who made it to the person who receives it, and sometimes all the way to a client or a court. This post looks at what it costs and how to make sure you are not the one producing it.

What workslop is

The term comes from researchers at BetterUp Labs and Stanford, who define it as "AI-generated content that looks good, but lacks substance" (BetterUp Labs & Stanford, 2025).

As an illustration: a report with neat headings and confident paragraphs that never quite answers the question. A slide deck whose numbers do not add up. A summary of a meeting that invents a decision nobody made. Each looks done. Each has to be redone by someone else.

1. 1.AI draft, unchecked

   Neat headings, confident paragraphs.

2. 2.Sent on as finished

   It looks done, so it goes out.

3. 3.Someone else redoes it

   The work moves to whoever receives it.

4. 4.Or it reaches a client

   Sometimes all the way to a court.

The researchers put a price on it. About $186 per employee per month in lost time, or roughly $9 million a year for a company of 10,000 people (BetterUp Labs & Stanford, 2025). Those are US figures, but the pattern travels.

Why it happens

Workslop is not a failure of the tools. It is a failure of habit. The KPMG and University of Melbourne study of more than 48,000 people found that 66% rely on AI output without evaluating it, and 56% have made mistakes in their work because of AI (KPMG & University of Melbourne, 2025). 57% hide their AI use, so nobody knows to look more closely.

Confidence makes it worse. A Microsoft Research and Carnegie Mellon survey of 319 knowledge workers found that higher confidence in generative AI was associated with less critical thinking (Microsoft Research & CMU, 2025). The better the output looks, the less we check it.

When it reaches a client

Deloitte Australia refunded more than A$97,000 on a A$440,000 report for the Australian government. The report contained a fabricated court quote and references to papers that did not exist (CFO Dive, 2025).

This was a large, well-resourced firm, on a paid government contract. The problem was not access to good tools. It was that invented material got through review and reached the client.

When it reaches a court

The legal world has seen this most often, because citations are easy to check and the consequences are public.

In India, a Bengaluru bench of the Income Tax Appellate Tribunal withdrew an order after it was found to cite judgments that did not exist (Taxscan, 2025).

On 27 February 2026, the Supreme Court of India said that relying on fake AI-generated judgments "would be a misconduct" (MediaNama, 2026).

It is not a handful of cases. A public database tracking court filings with AI-invented content listed 2,079 such cases worldwide as of 25 September 2026 (AI Hallucination Cases database).

When it leaks data

Workslop has a cousin: careless input rather than careless output. Samsung staff exposed internal company data by pasting it into a public AI chatbot. In May 2023 Samsung banned generative AI tools on company devices, and about 65% of staff it surveyed said such tools posed a serious security risk (CSO Online, May 2023).

Nobody in that story set out to leak anything. They were trying to get work done faster, with no clear rule about what could go into a public tool.

Why so many AI projects disappoint

You may have seen the claim that 95% of company AI pilots fail. It comes from an MIT NANDA report (Fortune, 2025). Treat it with care: it rests on 52 interviews and a narrow definition of success, and it has been criticised for both (Marketing AI Institute). It is a warning sign, not a measurement.

The better-evidenced point is quieter. People can feel more productive while their organisation sees little change. In Denmark, workers reported productivity benefits from AI but null effects on earnings and hours (NBER, Humlum & Vestergaard). Workslop is one reason: time saved by one person gets spent by another.

A checklist for not producing workslop

Before anything AI helped with leaves your hands, run through this.

1. Did I give it real context? The actual brief, audience, data and constraints, not a one-line request.
2. Does it answer the question that was asked? Read the first and last paragraph. If the point is not there, it is not done.
3. Have I checked every fact? Every number, name, date, quote and citation, against its original source.
4. Can I explain every line? If a colleague asks "why does it say this?", you should have an answer that is not "the AI wrote it".
5. Is it shorter than it needs to be, or longer? AI pads. Cut anything that adds words without adding meaning.
6. Did I paste anything I should not have? Client data, personal details, unreleased numbers, source code. Follow your organisation's rules; if there are none, ask.
7. Have I said where AI helped? Being open lets the reviewer check the right parts.

As an illustration: an HR executive asks AI for a new leave policy. Workslop is sending the draft as is. Good work is giving it the current policy and the relevant rules, checking every entitlement against the actual law and company handbook, trimming the generic paragraphs, and telling the manager which sections AI drafted.

1. 1.Give it real context

   The current policy and the relevant rules.

2. 2.Check every entitlement

   Against the actual law and company handbook.

3. 3.Trim the generic parts

   Cut what adds words without meaning.

4. 4.Say where AI helped

   Tell the manager which sections AI drafted.

What to do next

1. Adopt the checklist above for one month and note how often it catches something.
2. Pick one high-risk output in your work, such as client reports, legal references or financial figures, and make verification a fixed step.
3. Find your organisation's AI data policy this week. If there is none, raise it with your manager.
4. Tell your team when you use AI. Open use gets reviewed; hidden use does not.
5. Practise checking. Take an AI answer on a topic you know well and mark every error you can find.

AI Fluency teaches this checking habit from the first session, live and in small cohorts, with a project you keep each month. See the programme.