Why office workers are turning against AI - Brian Merchant | Substack

TEXT START: Take a look at this chart from a new report published by Glassdoor, the popular website where workers publish ratings of/complaints about/screeds of righteous fury aimed at their workplaces.

The Dissection

This is a labor-power analysis disguised as a report on workplace sentiment. Its strongest observation is that AI enthusiasm tracks control: executives and insulated professionals see leverage, while claims adjusters, accountants, IT workers, and customer-service staff receive mandates, surveillance, liability, rework, and replacement threats.

The article correctly shows that AI’s benefits are being privatized upward while its failures are dumped onto workers below. But it stops at abusive implementation. It treats “AI makes work worse” as the central diagnosis instead of recognizing it as the messy opening phase of labor displacement.

The Core Fallacy

The article mistakes transitional incompetence for structural defeat.

Hallucinations, misclassified claims, and human cleanup work demonstrate immature systems and badly designed workflows. They do not disprove the Discontinuity Thesis. Early automation frequently adds verification and repair labor before it becomes reliable enough to remove the labor it initially burdens.

The deeper error is assuming worker resentment can obstruct the trajectory. It cannot. Firms do not need workers to like AI. They need AI to produce a competitive cost or performance advantage. Once that advantage becomes durable, sentiment becomes irrelevant. The worker currently absorbing machine errors is not protected by that responsibility; he is functioning as a temporary liability shield until the workflow improves, the role narrows into a servitor niche, or the human layer is eliminated.

Hidden Assumptions

• Current AI failures will remain permanent rather than being targets for technical and organizational refinement.
• Companies must make work easier and more rewarding for automation to succeed; owners only need lower costs, higher output, or greater control.
• Human accountability will remain economically necessary even after systems become sufficiently reliable.
• Workers can preserve stable human-only domains through resentment, resistance, or institutional veto.
• Current job categories are fixed rather than provisional containers whose tasks can be decomposed and reassigned.
• Negative sentiment is evidence against AI adoption rather than evidence that the transition is transferring risk downward before removing labor altogether.

Social Function

This is a partial truth with a transition-management function. It documents real exploitation and punctures executive propaganda about frictionless productivity. But by framing the problem mainly as arrogant bosses forcing defective tools onto workers, it risks becoming ideological anesthesia: readers can blame bad management while missing the more terminal fact that ownership of cognitive capital is eroding their bargaining position.

The article records the pain of the transition accurately. It does not map the destination. It also implicitly treats worker power as recoverable if deployment becomes fairer, more accountable, or more democratic. Under the hardened framework, that is not enough. P1 makes cognitive automation competitively dominant; P2 prevents institutions from preserving broad human-only economic territory; P3 removes productive participation from the majority.

The Verdict

The article is right about the present and wrong about the endpoint.

Workers hate AI because they are being forced to absorb its costs while executives capture its upside. That resentment is not evidence that the system is failing. It is evidence that the post-WWII labor bargain is already being gutted: workers receive more surveillance, more responsibility, and more stress while their future economic necessity declines.

The insurance claims adjuster cleaning up AI’s errors is not a counterexample to obsolescence. He is the transition’s sacrificial buffer—paid to absorb machine failure until the machine, workflow, or liability structure removes him. This is a symptom report from the corpse, not proof that the corpse is alive.