AI Workslop: The Work That Looks Right and Isn’t

The first paragraph is sharp. The second is plausible. By the fourth you realise there is nothing underneath it. Researchers named this workslop — AI-generated output that passes as competent work and fails to advance the task. It is not a fringe irritation: 40% of US desk workers received it in a single month, each instance costing nearly two hours to untangle. Every figure below is attributed, dated, and rated for confidence.

Thank you for reading this post, don't forget to subscribe!

The gap between what was promised and what arrives

Researchers at BetterUp Labs and Stanford’s Social Media Lab gave the phenomenon a name in September 2025: workslop, defined as “AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task.” The definition is precise about the failure mode. This is not obviously bad work — obviously bad work is cheap to reject. Workslop is expensive precisely because it is plausible. It clears the first read and fails the second, which means the cost lands on the recipient rather than the sender.

That cost transfer is the whole mechanism. In the survey of 1,150 US full-time workers, resolving a single incident took an average of 1 hour 56 minutes — roughly twenty minutes longer than if the sender had simply written it themselves. The producer books a productivity gain; the reader absorbs a larger loss. At an organisational level the ledger nets out negative while every individual dashboard shows improvement.

Set that against the capital being deployed. The five largest US hyperscalers have committed $660–690 billion of capex for 2026, nearly double 2025, with global data-centre spending passing $1 trillion. Expenditure at that scale requires a correspondingly large promise, and a large promise creates pressure to demonstrate returns quickly. Workslop is one visible form that pressure takes at desk level: the appearance of AI-driven output, produced fast, reviewed lightly, and passed along. BetterUp’s researchers draw the connection explicitly — workslop is a candidate explanation for the 95% of organisations reporting no measurable return on generative AI.

• Plausibility is the problem, not incompetence. Work that fails obviously costs nothing to reject; work that fails on inspection costs two hours.
• The cost moves, it doesn’t disappear. Sender saves time, recipient loses more — invisible in per-person productivity metrics.
• The durable damage is reputational. 42% trusted the sender less; about half rated them less capable. That doesn’t reset next quarter.

The numbers

Source: BetterUp Labs with Stanford Social Media Lab, published in Harvard Business Review, September 2025. Where secondary coverage reports slightly different values (40 vs 41 percent; 1h51m vs 1h56m) we give the range rather than pick one. The $9M annualised figure is the researchers’ extrapolation from the per-worker monthly cost, not an observed total.

The cost nobody puts on the invoice

The dollar figure is the easy part. The survey’s more consequential finding is what receiving workslop does to how colleagues are judged — and these are not transient reactions.

Two hours is recoverable. A colleague quietly reclassifying you as someone whose work needs checking is not — and it compounds, because the rational response to an unreliable sender is to review everything they send. The organisation pays twice: once for the fix, then permanently in added verification.

Why it happens

Workslop is usually read as a discipline failure — someone was lazy. The research base points somewhere less comfortable: the incentives produce it.

• Output is measured; substance isn’t. Volume of work produced is legible to a manager. Whether it advanced the task is legible only to the person who has to use it.
• Adoption mandates without task redesign. Where teams are told to demonstrate AI usage but given no change in what “done” means, generation substitutes for thinking.
• Fluency reads as competence. Language models are optimised to produce confident, well-formed prose — precisely the surface signals humans use as shortcuts for quality.
• The failure is invisible to the producer. Whoever generates it rarely learns it collapsed; the cost is discovered downstream, often silently.

This lines up with the broader failure literature. RAND’s analysis of AI project failure finds root causes that are systemic and organisational rather than technical — misunderstood problem definition, inadequate data, technology-first thinking. Between 70% and 85% of AI projects miss their original objectives, and McKinsey finds 67% of failures cite organisational resistance as the primary barrier. Workslop is the same pathology at the level of a single document.

Slop outside, workslop inside

Workslop is the enterprise cousin of a public phenomenon we track separately: AI slop, the synthetic flood degrading search results, news and music. The mechanism is identical — generation cost collapsed to near zero while the cost of evaluating what was generated stayed exactly where it was. Whether the output lands in a search index or a colleague’s inbox, someone downstream absorbs the verification burden. The difference is that on the open internet the reader can leave. Inside an organisation, they have to fix it.

What actually reduces it

Nothing here requires new tooling. It requires changing what gets rewarded:

• Make the sender accountable for the outcome, not the artefact. The person who forwards it owns whether it was right.
• Ask for provenance, not prohibition. Bans push AI use underground. “What did you verify?” is answerable; “did you use AI?” invites a lie.
• Measure the receiving end. If rework time isn’t counted somewhere, the productivity gain will keep looking real.
• Reserve generation for drafts, not deliverables. The research shows the failure surfaces on second read — so build in a second read before it leaves.

FAQ

What is workslop?

AI-generated work content that masquerades as good work but lacks the substance to advance the task. The term was coined by BetterUp Labs and Stanford Social Media Lab in 2025.

How common is it?

40% of US desk workers reported receiving it in a single month. Managers were markedly more exposed at 54%, against 38.5% of individual contributors.

What does it cost?

About $186 per worker per month, from roughly two hours spent resolving each incident — extrapolating to over $9 million a year in a 10,000-person organisation.

Does workslop explain why AI pilots show no return?

It is one candidate explanation, and the researchers propose it as such. MIT found ~95% of generative-AI pilots delivered no measurable P&L impact; if gains are being offset by downstream rework, the aggregate would look exactly like that. This is a plausible mechanism, not a demonstrated cause.

Is the answer to ban AI at work?

The evidence doesn’t support that. The failure is in review and accountability, not in the tool — and prohibition tends to move usage out of sight rather than end it.

Primary source: BetterUp Labs with Stanford Social Media Lab, “AI-Generated ‘Workslop’ Is Destroying Productivity,” Harvard Business Review, September 2025 — survey of 1,150 US full-time desk workers. Secondary coverage reviewed for consistency: Axios, CNBC, TechCrunch. Capex figures from Dell’Oro Group and hyperscaler guidance for 2026. Project-failure context from RAND, McKinsey, and MIT’s NANDA initiative. HIGH marks figures stated directly in the published study; MEDIUM marks values where secondary reports differ slightly or which rest on self-reported time estimates; DERIVED marks the annualised organisational total, which is an extrapolation rather than an observed figure. Self-reported time costs should be read as indicative — people are poor at estimating how long tasks took. Corrections: see our methodology and corrections policy.

AI Slop: The Synthetic Flood · Enterprise AI Statistics 2026 · AI Bubble Tracker · AI Infrastructure & Compute · The Dark Side of AI