Author : Matthias Fahn, Jin Li
Generative AI can boost human productivity when used well. It helps teams explore strategies, draft marketing campaigns, and expand both the scope and speed of what one person can produce. But alongside these benefits, a new challenge is emerging inside organizations: workslop —content that appears polished and plausible at first glance but is thin on substance. A recent Harvard Business Review piece introduced this term, capturing what many leaders now encounter in daily work: well-packaged drafts that end up creating additional work for others.
Before the age of AI, it was often relatively easy to spot weak work because appearance and substance were at least loosely connected. Length, polish, and coherence were imperfect but informative signals of effort. With AI, that link is broken. It is now cheap to generate fluent, formally correct-looking memos, decks, code, and analyses. But their truth, rigor, and relevance remain uncertain until someone invests the time to check them carefully.
This shift creates an evaluation tax with organizational, not just technical, consequences. Attention, a scarce resource for any firm, has to move from production to verification. That does not mean slowing everything down. It means recognizing that for an organization’s most valuable contributors, the core task increasingly shifts from “who can draft the first version” to “who can judge whether the claims actually hold up.”
Because attention is scarce and decision-makers are already overloaded, the temptation is to rubber-stamp polished output. That shortcut can be costly. In October 2025, Deloitte Australia agreed to partially refund the government after a 237-page report was found to contain apparent AI-generated errors, including citations to non-existent papers and a misattributed legal quote. Even if the broad conclusions were defensible, the episode triggered remediation work, eroded trust, and created contractual friction. The problem was not AI use per se, but unchecked AI use.
If AI continues to accelerate drafting, the managerial question becomes unavoidable: how do organizations get the upside of scale and scope without drowning in workslop?
A natural response to the evaluation problem is to ask whether AI itself can take on the task of checking. If AI generates content, why not let AI verify it as well? In some domains, this works extremely well. Mathematics is the clearest example. Tools such as Lean can formally verify proofs with precision. In these settings, humans explore ideas and machines check correctness. Verification is formal, unambiguous, and scalable.
But most organizational work does not look like mathematics. Strategy, research synthesis, policy analysis, consulting recommendations, and much of software design are not formally verifiable. They depend on judgment, context, and interpretation. In these domains, verification cannot be fully automated and still requires human effort.
Even worse, AI-based checking can actively undermine human verification rather than complement it. Evidence from algorithm-assisted evaluation shows that when an algorithm performs an initial check, humans reduce their own scrutiny. The presence of an automated checker changes behavior: once people believe “the system has already checked,” the perceived return to careful human verification falls.
This brings us to the deeper issue: incentives. AI lowers the cost of production, but it raises the cost of motivation. Even as it becomes easier for people to do the right thing, it becomes much easier to do the wrong thing. When producing a plausible output is cheap, output no longer reliably reflects effort or judgment.
This is not a failure of work ethic. It is a rational response to how work is rewarded. People are evaluated on visible deliverables—documents produced, slides submitted, analyses completed. AI makes it easy to meet those criteria with minimal effort. Verification, by contrast, remains costly, time-consuming, and largely invisible. When incentives remain tied to production rather than scrutiny, effort shifts accordingly, and verification is systematically underprovided.
Facing this verification deficit, organizations have two primary ways to intervene. The first is to redesign the process. AI is not just a tool that speeds up existing workflows; it changes what kinds of workflows are viable. In software development, for example, line-by-line review made sense when code was written by humans at human speed. With AI-generated code, that approach no longer scales. The response in high-functioning teams has not been “review more carefully,” but to redesign the process itself: moving verification upstream, toward specifications, constraints, tests, and interfaces. The scarce skill shifts from generating outputs to defining what counts as correct.
The second is to redesign the incentives. For judgment-heavy work that cannot be formally specified, firms must make the invisible work of verification visible. This requires a radical shift in how labor is valued. Organizations should treat “catching a subtle error,” “nuancing a complex argument,” or “challenging a plausible but wrong conclusion” as paths to promotion. If firms want humans to do the hard work of judgment, they must stop rewarding volume and speed and start rewarding the detection of error. One immediate consequence is a reallocation of expertise from drafting to verification; firms will increasingly pull scarce talent away from producing content and towards reviewing and signing off on the pieces of work that really matter.
If firms fail to do either, they will resort to a third, more destructive option: standardization. Firms will shift toward work that is easier to specify, measure, and check. Roles will narrow, discretion will shrink, and complex tasks will be decomposed into standardized pieces. The result is a drift toward “bad jobs”, roles optimized for compliance and throughput rather than judgment.
This drift threatens to hollow out the human experience of work. By stripping away judgment to make outputs easier to measure, we inadvertently design roles that demand compliance rather than intelligence. We risk creating a workforce confined to the shallow end of the cognitive pool—executing standardized tasks that are safe to audit but soul-crushing to perform.
To avoid this future, we must recognize that the AI challenge is not primarily technological; it is organizational. For a future where humans can thrive, we must redesign our processes and incentives to solve the evaluation problem.