Parkinson’s Law states that work expands to fill the time available. In the era of AI, workers now have a tool that expands to fill whatever a large language model can be persuaded to generate, which is to say, without limit.

What I have watched happen in my profession in the last two years, I am still struggling to describe. The first time I knew something was wrong, roughly a year and a quarter ago, I noticed a colleague replying to me using AI. His response was obviously generated by Claude. The punctuation gave it away — em dashes where no one types em dashes, the rhythmic structure, the confident grasp of technologies I knew for a fact he did not understand. I sat with it for a while, weighing whether to debate someone who was visibly copy-pasting verbatim from a model. The channel was public, and I spent more time than I should have correcting fundamentals. Eventually I stopped. He was not, in any meaningful sense, on the other side of the conversation.

Generative AI can produce work that looks expert without being expert, and the failure arrives in two shapes. The first is when novices in a field are able to produce work that resembles what their seniors produce, faster or more advanced than their judgment. The second is when people generate artifacts in disciplines they were never trained in. The two failures look similar from a distance and are not the same. Research has mostly measured the first. The second is what it is missing, and in my experience it is the more riskier of the two.

Cross domain generation

People who cannot write code are building software. People who have never designed a data system are designing data systems. Most of it is not shipped; it is built, often for many hours, possibly shown internally with great vigor, used quietly, and occasionally surfaced to a client without much fanfare. Workers can obsess over an idea, working many hours overtime. There are a few practitioners who use the current agentic tools to do complex things properly, but they are scarce and as I find, typically in code generation. AI, for all its capabilities at the level of the individual, has not scaled properly in my workplace.

I have a colleague, a careful and intelligent person in a role that is not engineering, who spent two months earlier this year building a system that should have been designed by someone with formal training in data architecture. He used the tools well, by the standards by which use of the tools is currently measured. He produced a great deal of code, a great deal of documentation, a great deal of what looked, to anyone who did not know what to look for, like progress. He could not, when asked, explain how any of it actually worked. The work was wrong from the first day. The schemas, and more importantly the objectives, were wrong in a way that would have been obvious to anyone with two years in the field. Several of us did know. When opinions were voiced even as high as a V.P., he fought back. The room had been arranged in such a way that saying so was not a contribution; his managers were too invested in the appearance of momentum to want the appearance disturbed. The work will continue, in all probability, until it is shown to a stakeholder, and they decide not to invest.

This is the part of the phenomenon I find hardest to write about. The tool did not make him a worse colleague. It made him able to impersonate, for months, a discipline he had never trained in, and the impersonation was good enough that the institutional incentives all bent toward letting him continue. Perhaps it’s a failure of management, but I have been finding management to be so eager to embrace AI that they’re willing to accept the risk.

It would be tolerable, perhaps, if the tool offered an honest assessment of what it had produced. The Cheng et al. Stanford study published in Science this spring [1] confirmed what every regular user already knew: leading models are roughly fifty percent more agreeable than human respondents, affirming the user even where the affirmation is unwarranted. Berkeley CMR meta-analyses [4] found AI-literate users often overestimate their performance. Particularly interesting when workers stray outside of their training. An NBER study of support agents [2] found generative AI boosted novice productivity by about a third while barely helping experts. Harvard Business School researchers found the same pattern in consulting work [3]. So you have overconfident, novices able to improve their individual productivity in an area of expertise they are unable to review for correctness. What could go wrong?

The conduit problem

A growing body of work calls this output-competence decoupling [5]. In any previous era, the quality of a piece of work was a more or less reliable signal of the competence of the person who produced it. A novice essay read like a novice essay; novice code crashed in novice ways. AI has severed that relationship. A novice now produces work that does not betray the novice, because the competence the work reflects is not the novice’s competence at all. It is the system’s. The person, in the transaction, becomes a kind of conduit, capable of routing the output to a recipient and incapable of evaluating it on the way through.

The skills of producing work and judging it were deliberately distinct, but accomplishing the work itself used to teach the judgment. The first skill now belongs, in large part, to the machines. The second still belongs to us, though fewer are bothering to acquire or utilize it.

The architectural critique that used to come from someone who was taught, or who had built and broken three of these before now comes from a model with no embodied memory of building or breaking anything. The slowness was not a tax on the real work; the slowness was the real work. It was how the work got good, and how the people producing the work got good, and how the firm whose name was on the work could promise the client that what they were buying was a particular kind of thing rather than a generic one.

The current generation of agentic systems is built around the premise that the human is the bottleneck — that the loop runs faster and cleaner without the awkward delay of someone reading what is about to happen and deciding whether it should. This is, in a great many cases, exactly backwards. The human in the loop is not a vestige of an earlier era; the human is the only part of the loop with skin in the game. Removing the H from HITL is not an efficiency. It is the abandonment of the only mechanism the system has for catching itself.

Slop on the inside

Requirements documents that were once a page are now twelve. Status updates that were once three sentences are now bulleted summaries of bulleted summaries. Retrospective notes, post-incident reports, design memos, kickoff decks: every artifact that can be elongated is, by people who do not read what they produce, for readers who do not read what they receive. The cost of producing a document has fallen to nearly zero; the cost of reading one has not, and is in fact rising, because the reader must now sift the synthetic context for whatever the document was originally about. Each individual decision to elongate seems rational, and each is independently rewarded — readers are more confident in longer AI-generated explanations whether or not the explanations are correct [5]. The collective effect is that the signal in any given workplace is harder to find than it was before any of this began. The checkpoints have been hidden, drowned in their own paperwork, even when the people drowning them were genuinely trying to “be brief”.

This is a new form of slop, and it is more expensive than the public kind, because the people producing it are being paid a salary to do so. The pipeline of future experts is thinning from both ends. The work that used to teach judgment is now done by the tool, and the entry-level roles where the teaching happened are being cut on the theory that the tool can do the work. What this is causing, in many offices including mine, is a great deal of motion and very little of what motion used to create.

The downstream costs are accumulating quickly. Most of the public discussion of AI slop has focused on the flood into public markets — a University of Florida marketing study [6] being among the more direct treatments. What is less remarked upon is the same dynamic playing out inside organizations: time wasted using AI on tasks that did not need it, on artifacts no one will read, on processes that exist only because the tool made it cheap to construct them. On decks that spell out things that previously didn’t even need to be said or were assumed.

What to do about it

What discipline looks like, in this environment, is almost embarrassingly old-fashioned and may seem obvious to most of you until you try to avoid it. Use the tool where you can verify precisely what it produces. Never ask a model for confirmation; the tool agrees with everyone, and an agreement that costs the agreer nothing is worth nothing.

Generative AI does well on tasks where feedback is fast, where being approximately right is good enough, where the human remains the final arbiter. Drafting a memo, generating examples, summarizing material the reader could verify if they cared to. The University of Illinois Generative AI guidance [7] and the PLOS Computational Biology “Ten Simple Rules” paper on AI in research [8], among the more careful documents now circulating, list much of this explicitly: brainstorming, copyediting, reformulating one’s own ideas, pattern detection in data one already understands.

In every recommended use, the human supplies the judgment and the tool supplies the throughput. This is a stronger position than human-in-the-loop. The tool sits outside the work, contributing where invited and silent otherwise, which is the opposite of what most agentic systems are now being built to do.

For firms, the competitive advantage of a firm whose work can be trusted has not disappeared; it has, if anything, appreciated, because so many of the firm’s competitors are quietly converting themselves into content-generation pipelines and counting on the client not to notice.

This is already coming to a head. Deloitte has already refunded part of a $440,000 fee over an AI-hallucinated government report. It could be a production system built on a hallucinated specification, or a senior engineer who realizes they have spent the last year nominally reviewing work they could no longer competently review. The reckoning will not be subtle. The firms still doing the work properly will be in a position to charge for it. The firms that have hollowed themselves out will discover that what they hollowed out was the thing the client was paying for.

Misunderstanding and misuse of AI in the workplace is rampant. In many of the rooms I now find myself in, expertise has been asked to look the other way: to deliver faster, produce more, integrate the tools more deeply, get out of the way of the colleagues who are “getting things done”. The artifacts are accumulating; the work is not. And somewhere on the other side of all this output, a client is opening a deliverable, reading a summarized list, and they may just choose to review it manually.

References

1. Sycophantic AI decreases prosocial intentions and promotes dependence (Cheng, Lee, Khadpe, Yu, Han, & Jurafsky, 2026). Science.

2. Generative AI at Work (Brynjolfsson, Li, & Raymond, 2025). The Quarterly Journal of Economics, 140(2), 889–942. Also: NBER Working Paper No. 31161, April 2023.

3. Navigating the Jagged Technological Frontier (Dell’Acqua, McFowland, Mollick, et al., 2026). Organization Science. Originally HBS Working Paper No. 24-013, 2023.

4. Seven Myths About AI and Productivity: What the Evidence Really Says (Berkeley CMR, 2025). Meta-analysis confirming asymmetric AI productivity gains and user overconfidence.

5. Beyond the Steeper Curve: AI-Mediated Metacognitive Decoupling (Koch, 2025). Longer AI explanations make users more confident regardless of correctness.

6. Generative AI and the market for creative content (Zou, Shi, & Wu, 2026). Forthcoming, Journal of Marketing Research.

7. Generative AI Guidance (University of Illinois). Recommended uses and limitations of generative AI in academic and professional work.

8. Ten simple rules for optimal and careful use of generative AI in science (Helmy, Jin, et al., 2025). PLOS Computational Biology, 21(10), e1013588.