The early framing of generative writing tools in the workplace treated output volume as the metric: more documents, more updates, more drafts, produced faster. That framing assumed the cost of communication sits with the sender.
Research published from late 2025 onward suggests the cost largely moved to the recipient.
The workslop findings
Researchers at BetterUp Labs and Stanford's Social Media Lab introduced the term workslop in Harvard Business Review in September 2025, defining it as AI-generated content that has the appearance of competent work without the substance to advance the task. Their survey covered 1,150 US full-time employees.[1]
| Reported finding | Figure |
|---|---|
| Received workslop in the preceding month | 41% |
| Average time to resolve one incident | 1 hour 56 minutes |
| Felt annoyed on receiving it | 53% |
| Viewed the sender as less trustworthy afterward | 42% |
| Less likely to want to work with that person again | about one third |
Two distinct costs appear here. The first is time, and it is transferred rather than saved: the sender's minutes become the recipient's two hours of verification and repair. The second is reputational, and it is the more consequential of the two, because it attaches to the person rather than the document.
Follow-up work in 2026 examined why people produce it despite these costs.[2]
Interpreting these figures carefully
These are self-reported survey responses, with the usual limits. "Workslop" requires the recipient to judge that content was AI-generated and inadequate, and recipients may misattribute poor work to AI, or fail to notice competent AI-assisted work at all. The measure is therefore closer to perceived workslop, which is the relevant construct for trust effects but not for productivity accounting.
The two-hour figure is an average of self-estimated remediation time, a category people estimate poorly.
Why the trust effect is the important number
Time costs are recoverable. A team can absorb an hour, or fix its review process.
The reputational finding is different in kind. If 42% of recipients trust the sender less, the sanction applies to every subsequent message from that person, including ones they wrote themselves. A worker who sends one inadequate AI-drafted document has changed how their next unassisted document is read.
This creates an asymmetry that is worth stating plainly: AI assistance is invisible when it works and attributed when it fails. A polished document that holds up is read as competence. A polished document that collapses under inspection is read as delegation. Nobody receives credit for the first case, and the second is charged to the person.
A related 2026 survey found that 81% of workers believe they can usually tell when a message was written by AI.[4] Whether that confidence is calibrated is a separate question, and probably it is not. But a workforce that believes it can detect AI authorship will act on that belief, which is sufficient to produce the trust effect regardless of accuracy.
The implication for spoken channels
A trust penalty attached to a channel raises the value of channels that do not carry it.
Live conversation is currently the least delegable workplace channel. A spoken answer in a meeting is produced in real time, and cannot be reviewed before delivery. Under the asymmetry above, this makes it the channel where perceived authorship is unambiguous, which is precisely what the written channel lost.
There is direct evidence that workers are already reasoning this way about their own communication: a May 2026 survey found 63% reporting that they use AI to avoid difficult conversations at work.[3] If unpleasant conversations migrate into the assisted written channel, the residue left in live conversation is the communication that people judged too important, too sensitive, or too relationship-dependent to delegate.
That is a significant reweighting. The spoken channel is carrying a smaller share of total workplace communication and a larger share of its consequential portion, which is a poor trade for anyone whose spoken performance is weaker than their written performance. Survey evidence suggests that describes a substantial share of the workforce: 51% report that AI use has made spontaneous conversation feel more difficult.[3]
Meanwhile the spoken channel is itself being recorded and summarised at scale,[5] so the reviewability that drove people toward assisted writing is arriving in conversation as well.
What is not established
- No causal evidence links AI authorship to trust loss. The surveys measure association within self-report.
- No study has measured whether disclosure removes the penalty. Disclosure norms are widely recommended and, as far as we can determine, untested.
- The detection claim is unvalidated. The 81% figure measures confidence, not accuracy. No published test compares self-reported detection ability against ground truth at scale.
- The spoken-channel displacement is inferred, not measured. The 63% avoidance figure is self-reported intent, and nobody has tracked whether difficult conversations actually moved.
The third of these is the most consequential and the easiest to test. If workers are systematically wrong about which messages were AI-written, then the trust penalty is being applied to a population that partly did not earn it, and the entire behavioural response rests on a detection ability nobody has verified.