TL;DR
HBR says companies that went all-in on AI face “knowledge decay” as low-quality outputs pile up, erode trust, and cost $9M a year in rework.
Companies that pushed hardest to adopt generative AI are now contending with a problem the technology was supposed to prevent: their work is getting worse. Two articles published by Harvard Business Review this month describe a feedback loop in which AI-generated low-quality output degrades the information companies rely on to make decisions, a phenomenon the authors call “knowledge decay.”
The June 2026 HBR article, written by Oxford operations management professor Matthias Holweg and Babson College professor Thomas Davenport, argues that the damage goes beyond individual errors. When employees use AI to produce work that looks polished but contains mistakes or lacks substance, colleagues downstream waste time verifying, correcting, or redoing it. As those errors compound across teams and departments, the organisation’s collective knowledge base deteriorates.
The term for this low-quality AI output already has a name. BetterUp Labs and Stanford’s Social Media Lab coined “workslop” in a September 2025 HBR article to describe AI-generated content that masquerades as good work but lacks the substance to advance a task. Their survey of 1,150 US full-time workers found that 41 percent had received workslop in the preceding month, with each incident requiring an average of one hour and 56 minutes to sort out.
The financial cost is significant. Using respondents’ self-reported salaries and time estimates, the researchers calculated that workslop costs roughly $186 per worker per month. For a company of 10,000 employees, that translates to more than $9 million annually in lost productivity, a figure that does not account for the downstream effects on morale and trust.
Those social costs may matter more than the financial ones. In the BetterUp-Stanford survey, 53 percent of workers who received workslop said they were annoyed, 42 percent viewed the sender as less trustworthy, and roughly half considered the colleague less creative, capable, or reliable than before. A third said they were less likely to want to work with that person again.
The broader productivity picture is no more encouraging. A July 2025 MIT Media Lab report found that 95 percent of organisations saw no measurable return on their generative AI investments, despite billions in spending. Goldman Sachs reached a similar conclusion in March 2026, finding no meaningful relationship between AI adoption and productivity gains at the economy-wide level, even as 70 percent of S&P 500 management teams discussed AI on earnings calls.
The knowledge decay problem is distinct from the familiar complaint that AI hallucinates. Hallucinations are factual errors in AI output. Knowledge decay describes what happens to an organisation when those errors, and the broader pattern of low-effort AI-generated work, accumulate over months.
Workers stop trusting internal documents. Processes built on unreliable information produce unreliable results. Institutional memory thins as employees lean on AI rather than developing expertise themselves.
Holweg and Davenport warn that the hiring process has been particularly damaged. AI-generated resumes flood recruiters, AI-generated job listings mislead candidates, and AI-powered screening tools filter out qualified applicants. The result, as HBR puts it, is that trust in the hiring process has sunk to “all-time lows for both job seekers and recruiters.”
The worker backlash is already measurable. A 2026 survey of 2,400 workers across the US, UK, and Europe found that 29 percent admit to actively sabotaging their employer’s AI strategy by ignoring guidelines, refusing training, or deliberately skewing performance data. Among Gen Z workers, that figure rises to 44 percent, driven largely by fear of job displacement.
This resistance sits alongside a broader pattern of AI-justified layoffs that often lack clear evidence that AI systems actually replaced the eliminated roles. The tech sector recorded more than 95,000 job cuts across 247 events in 2026, with nearly half attributed to AI, even though analysts have questioned whether many of those companies had mature AI implementations capable of absorbing the work.
The irony is that fixing the workslop problem requires exactly the kind of labour AI was supposed to reduce. Business leaders must now invest in verification processes, quality standards, and human oversight to ensure AI-generated content meets the bar, work that consumes the time of actual employees. HBR’s prescription amounts to building a new layer of human checking around AI output, which undermines the efficiency argument that justified adoption in the first place.
Both HBR articles draw a distinction between indiscriminate AI mandates and targeted use. The June article notes that proprietary models trained on company-specific data can add genuine value, while public LLMs applied to tasks they are poorly suited for produce “generic prose that often contains mistakes.” Companies that froze hiring citing AI productivity gains are now discovering that the gains may be illusory if the quality of the work degrades faster than the headcount shrinks.
The knowledge decay concept reframes the AI productivity debate. The question is no longer just whether AI makes individual tasks faster, but whether the cumulative effect of widespread AI use makes an organisation’s decision-making better or worse. HBR’s answer, for companies that adopted AI without quality controls, is that it makes it worse.
Holweg and Davenport’s credentials lend the argument weight, but it is worth noting that the knowledge decay framework has not yet been tested through controlled empirical studies. The concept synthesises existing evidence rather than presenting new data, and the BetterUp-Stanford workslop survey relies on self-reported estimates of time lost. How accurately workers gauge time spent on rework is an open question.
Still, the pattern is consistent across multiple sources. Goldman Sachs, MIT, BCG, and now two separate HBR articles from different research teams arrive at variations of the same conclusion: most companies are not getting what they expected from generative AI, and the ones that pushed hardest may be paying the highest hidden cost.
