Artificial intelligence was supposed to cut out busywork. Instead, a new study claims it’s generating an altogether different kind of mess: “workslop.” Over 40 percent of full-time U.S. employees experienced AI-generated, polished-looking content in the last month that failed to advance a task, according to research by the Stanford Social Media Lab and BetterUp Labs.
What lands in their inbox now, on average, fits that description 15.4% of the time, according to them. The results, published on Sunday in Harvard Business Review (HBR), describe a landscape of workplace pipelines growing increasingly clogged with smooth summaries, generic slides and snippets of code that take more fixing than they spare.
What Workslop Looks Like, on the Ground
Most workslop moves sideways, according to the Stanford team: Peers reported that about 40% of the work they receive from colleagues includes evidence of AI overreach — confident wording, thin substance and context widely missed.
And another 18% moves up the chain when direct reports send AI-composed draft materials to their manager, with additional costs for review and revision.
It’s most acute in technology and professional services, where generative AI was embraced early on and much of the work involves documents, decks and code. Examples include autogenerated status updates that don’t mention anything about key risks, slideware full of buzzwords and devoid of understanding, meeting notes that seem to hallucinate decisions being made, or boilerplate emails that leave customers wondering why they just read them.
Why the Efficient Approach to AI Can Backfire at Work
Workslop exists where AI saves money for decent first drafts but costs more when it comes to verification. Even when done at an individual level, though, such modest per-worker time savings can be canceled out by new oversight tasks like sifting through students’ homework for AI misuse or auditing plans recommended by AI for mistakes, according to research from the University of Chicago with the University of Copenhagen.
The pattern isn’t universal. Earlier work by MIT and Stanford found that for customer support agents, using AI on routine questions allowed them to complete messages 55% faster; and when developers were given a simple coding task to do with Copilot, they completed it in 55% less time than without. But for complex assignments, independent studies show that AI tooling can actually bog teams down as they spend more time prompting, validating and refactoring than solving the actual problem.
The Costs for Teams That You Don’t See from AI
Beyond time, workslop corrodes trust. Almost half of the participants in the Stanford and BetterUp Labs poll reported that coworkers who send workslop appear to be less creative, competent and dependable. And 42 percent found them to be less trustworthy, while 37 percent thought they were less intelligent. Those reputational hits multiply when AI artifacts leak through to customers or executives.
There’s also signal dilution. When inboxes are flooded with look-alike drafts, the signal gets lost in the noise. Reviewers get more skeptical, which may slow decisions and inhibit real experimentation. Meanwhile, institutional memory withers when teams revert to standard outputs rather than building the detail-rich nuance that separates good work from mere good-looking work.
How Companies Can Put the Brakes on Workslop Without Stalling AI
Set clear provenance rules. Make it mandatory for employees to reveal when AI supplemented, which model was used and what human checks were employed. Combine that with “quality gates” within collaboration tools — templates that prompt authors to cite their underlying data, note their interpretation assumptions and include a brief validation plan prior to submitting. The NIST AI Risk Management Framework and ISO/IEC 42001 provide valuable scaffolding for policy development.
Redefine done. Apply lightweight review rubrics to AI-assisted deliverables: factual accuracy, domain relevance, original reasoning and actionability. Not only measure the volume of output, but also time spent on rework and error rates. Some teams monitor a “slop tax” — the portion of AI-originated content sent back for revision — to illuminate bottlenecks and refine prompts or guidance.
Train for judgment, not just prompting. Train employees to know when AI is helpful (summarizing long documents, drafting alternatives and generating test cases) and when it’s dangerous (novel analysis, high-stakes decisions, subtle stakeholder dynamics). Promote “AI as thought partner, not final author”: model off of options and then render conclusions in human voice with source-backed evidence.
The Bottom Line on Curbing AI Workslop in Workplaces
AI can speed up real work, but only if companies hold quality to the same standard they do for speed. Stanford’s finding is a cautionary tale for our times: When leaders measure things without measuring meaning, they get more of the former and less of the latter. Rolling back AI isn’t the fix — raising the bar for what constitutes work is.
