Several recent Harvard Business Review (HBR) articles warn that firms that have aggressively adopted generative artificial intelligence are increasingly facing a phenomenon labeled "knowledge decay." The authors — Matthias Holweg, professor of operations management at the University of Oxford, and Thomas Davenport, professor at Babson College — argue that low‑quality AI output gradually corrodes the information base companies use for decision making. The June 2026 HBR analysis describes a feedback loop in which faulty or low‑value AI outputs become inputs for subsequent work phases, increasing the time teams spend on checking and fixing content.
Defining and measuring "workslop"
Researchers at BetterUp Labs and the Stanford Social Media Lab coined the term "workslop" in an HBR piece published in September 2025 to describe AI‑generated work that appears serviceable but lacks sufficient value to advance tasks. In a survey of 1,150 full‑time U.S. employees, 41% reported receiving at least one workslop item in the prior month; each incident required on average 1 hour 56 minutes of extra work to uncover and correct errors.
Financial and organizational costs
Using respondents' reported wages and time, the BetterUp–Stanford team estimated that workslop costs an average of $186 per employee per month. For a company with 10,000 employees this would translate to more than $9 million in lost work time annually. That figure excludes broader impacts on morale, trust, and collaboration: among those who received workslop, 53% felt annoyed, 42% trusted the sender less, nearly half judged the sender as less creative or competent, and one in three were less willing to collaborate with that colleague again.
Productivity promises vs. measured returns
Broader evidence on productivity is similarly sobering. The MIT Media Lab reported in July 2025 that 95% of organizations could not demonstrate measurable returns from generative AI investments despite large global spending. Goldman Sachs reached a comparable conclusion in March 2026: at the macro level there is no clear link between AI adoption and productivity gains, even as more than 70% of S&P 500 executives mentioned AI in quarterly earnings reports.
How knowledge decay differs from hallucinations
Knowledge decay is distinct from the familiar notion of AI hallucinations. A hallucination is a discrete factual error in an AI response; knowledge decay describes what happens when such errors and generally low‑effort outputs accumulate across months in an organization. Employees lose confidence in internal documents, processes built on unreliable information produce unreliable results, and organizational knowledge atrophies as staff rely more on AI than on developing their own expertise.
Particularly affected areas: hiring and employee resistance
Holweg and Davenport highlight hiring as especially vulnerable: AI‑generated resumes flood recruiters, AI‑written job ads mislead applicants, and automated filters can exclude genuinely suitable candidates. HBR characterizes trust in hiring processes as having fallen to a "historic low" among both applicants and recruiters. Worker pushback is also measurable: a 2026 survey of 2,400 U.S., U.K. and European employees found 29% admitted actively undermining their employer's AI strategy; among Generation Z respondents that share was 44%, largely driven by fear of job loss.
AI‑related layoffs and the reality behind them
This phenomenon coincides with waves of AI‑cited layoffs: in the tech sector in 2026 more than 95,000 jobs were cut across 247 distinct layoff events, and roughly half of those companies cited AI as the rationale. Analysts have questioned whether many of those firms actually had mature AI systems capable of taking over the tasks of the laid‑off workers.
The irony: more human work required
Paradoxically, addressing workslop and knowledge decay requires more human labor and new processes: leaders must implement oversight, quality standards, and human review layers to ensure AI outputs meet expectations. That, in turn, undercuts some of the efficiency arguments that motivated rapid AI adoption.
Not all AI use is equal
Both HBR articles stress the distinction between undifferentiated AI usage and targeted deployment. The June analysis suggests that models trained on company‑specific data can create genuine value, while public large language models (LLMs) often produce generic, template‑style text that is more error‑prone. Organizations that froze hiring citing AI gains may now find those gains illusory if work quality erodes faster than headcount.
Conclusion
The concept of knowledge decay reframes the productivity debate around AI: the question goes beyond whether AI speeds up individual tasks to whether widespread use improves or degrades an organization's overall decision‑making capability. HBR's authors conclude that for firms that deployed AI without adequate quality controls the net effect is more likely negative. It is important to note that the knowledge‑decay hypothesis so far rests largely on synthesis of existing studies and self‑reported survey data rather than controlled empirical trials. Still, multiple independent analyses — including work from Goldman Sachs, the MIT, BCG, and two HBR research teams — point in a similar direction: many companies are not receiving the promised returns from generative AI, and those that adopted it fastest and most enthusiastically may be paying the highest hidden price.
