When the Office Falls for the Machine: A Blog Post About AI Delusion and What It Reveals

A recent personal essay titled "our workplace LLM mass delusion", published by a blogger writing under the name Ava, has been circulating in the corners of the web where people who are tired of artificial intelligence gather to compare notes. The piece is not a technical critique. It contains no benchmarks, no discussion of context windows or retrieval pipelines, no argument about whether transformer models can reason. What it offers instead is something rarer and, in its way, more uncomfortable: a ground-level report on how a large language model deployment actually felt to one employee inside an organization that could not afford it, and a meditation on what that experience revealed about the place she works.

The core argument is simple to state and harder to dismiss. Ava claims that her employer, an organization she describes as chronically underfunded, has redirected money it swore it did not have toward AI tooling, consultants, and seminars, all while cancelling employee bonuses, refusing to fill open positions, and telling overworked departments to "find a way to deal with it." The contradiction is the whole point. The same leadership that demands annual justification for a single database her team needs found, almost overnight, the budget for enterprise licenses to both ChatGPT and Copilot, plus years of workshops. The technology is incidental. What she is really describing is a revelation about institutional will.

The evidence she marshals

The essay's persuasive force comes from accumulation rather than abstraction. Ava reports attending every company-wide meeting where teams presented their LLM pilot projects, and her tally is stark: across hundreds of people, many teams, and months of enthusiastic effort, not one project succeeded. Every pilot, she writes, concluded the same way, that the tool did not save time, over-complicated the work, or simply could not be made reliable. The recurring complaints she lists are familiar to anyone who has watched these deployments closely: hallucinated content, an inability to correctly fill in or edit documents, the model ignoring source material it was given, and the long tail of verification and correction that erased whatever speed the tool seemed to promise.

What elevates the piece from a list of grievances is her eye for the absurd particular. She describes a presentation where an employee demonstrated, in earnest, that you could ask the chatbot how it was feeling that day. She describes another where a colleague uploaded a single-page cafeteria menu, a tidy little timetable, and asked the model what was for lunch on Wednesday, only for the bot to produce an answer longer than the sheet itself. And she describes her own head of IT advising staff that, when unsure whether an email is a phishing attempt, they should save the suspicious file to their desktop and upload it to ChatGPT. That last anecdote is the one that should make security professionals wince, because it inverts basic incident-response hygiene: the recommended workaround involves downloading and retaining exactly the artifact you would normally want quarantined or deleted, and routing potentially sensitive material to a third-party service. The detail lands because it shows enthusiasm overriding competence in someone who, by title, ought to know better.

The implication that actually stings

If the essay only catalogued failed pilots, it would join a large and growing pile of disillusionment posts. Surveys through 2025 and into 2026 have repeatedly found that the gap between AI investment and measurable return remains wide, and that a large share of corporate pilots never reach production. Ava's contribution is to draw a conclusion most of those reports leave implicit. The speed of her organization's AI adoption, she argues, exposes a lie at the heart of how the place has always operated.

Here is the reasoning, and it is worth following carefully because it is the strongest thing in the piece. Public-sector and large bureaucratic institutions habitually explain their inertia as structural. Change takes years. Budgets are fixed. The infrastructure cannot be touched quickly. Employees internalize this as a law of nature, the way the tides are a law of nature. Then a technology arrives wrapped in enough hype, and suddenly the same institution stands up new licenses, organizes recurring company-wide meetings, contracts external trainers, and reorganizes people's time, all within months. The obstacle was never structural. It was a matter of priority. As Ava puts it, the moment is "mask-off," because it proves that the friction employees were told to accept as inevitable was, in fact, a choice that leadership could have made differently at any point. That realization, she writes, is "completely trust-shattering," and it is easy to see why. Discovering that your organization could always have moved fast for the things it cared about reframes every previous refusal as a statement of what it did not care about.

This is the deeper pattern the essay reaches toward, even if it does not name it in these terms. Technologies do not arrive into neutral environments. They arrive into existing structures of power and budget and attention, and the way an organization receives a new tool tells you more about the organization than about the tool. The LLM here functions less as a productivity instrument than as a kind of contrast dye, lighting up the circulatory system of institutional priority that was invisible before.

The psychology she diagnoses

Ava also offers a behavioral explanation for why intelligent, responsible colleagues became, in her words, shills. She invokes the Dunning-Kruger effect, the cognitive bias in which limited competence in a domain coincides with inflated confidence, and argues that the chatbot is unusually good at amplifying it. The mechanism she sketches is plausible. A fluent, confident, well-formatted response makes a trivial task feel substantial. The tool dresses up mundane work in the costume of innovation, letting a person feel they are doing something important while summarizing a lunch menu. The fluency of the output gets mistaken for the significance of the work.

There is a real insight buried here, and it generalizes beyond her office. Language models are optimized to produce text that reads as competent and authoritative regardless of whether the underlying task warranted any tool at all. That surface polish can short-circuit the ordinary judgment by which a professional would otherwise ask whether a task needed automating. When every output looks like progress, the question of whether it is progress stops getting asked. Her phrase for this, "packaging every fart in a nice bow that makes it seem deep," is cruder than a research paper would put it, but it points at the same phenomenon that more careful observers describe as automation bias compounded by the persuasive register of generated prose.

The counter-perspective the author herself concedes

To her credit, Ava does not claim her experience is universal, and any honest reading has to hold her account against the cases where it does not apply. She closes by genuinely congratulating people whose workplaces use these tools well, in industries where the fit is real. That concession matters, because the failure she documents is specific in ways she names clearly: her field is not software development, and she estimates that ninety percent of her colleagues do simply not have the kind of work that current tools meaningfully accelerate.

That specificity is the most important caveat for any reader tempted to generalize from the piece. The contrast between her office and, say, a team of engineers using these models for code generation, where the output is immediately testable and the feedback loop is tight, is enormous. The failure she describes is partly a failure of fit, a powerful general-purpose tool pressed onto work it was never going to suit, by leadership that wanted the appearance of adoption more than any particular outcome. Defenders of the technology would reasonably argue that her organization's experience indicts its rollout, not the underlying capability, and that a tool deployed without a real problem to solve will of course fail to solve one. They would be right about that. Where Ava is also right is that this distinction offers little comfort to the employee whose bonus was cancelled to fund the cosplay.

The two readings are not actually in conflict, which is what makes the essay worth sitting with. A technology can be genuinely useful in the contexts that fit it and simultaneously function, elsewhere, as an expensive ritual of looking modern. The same set of facts supports both claims. What Ava has written is finally less an argument about whether language models work than a document of what it does to people to be told there is no money, repeatedly, for years, and then to watch the money appear the moment leadership decides the story is worth telling. She calls it her "second Covid," a shared event that quietly rearranged her sense of how the institutions around her actually function. Whether or not the bubble she anticipates ever bursts, that particular disillusionment will not un-happen, and her essay is a careful record of the moment it set in.