In one sentence
Treating “AI” as a failed technology is more analytically useful than treating it as an inevitable innovation. LLMs routinely fail to deliver promised value, yet remain widespread because capital, competitive anxiety, and institutions use them to weaken worker power and normalize compliance.
Overview
Marcotte begins by deliberately rejecting the usual language of AI progress. Calling LLMs a failure does not deny that particular uses can be helpful; it places those uses against what he sees as the technology’s dominant pattern: unreliable performance combined with substantial harm.
His case has several parts. As products, LLMs have not delivered the sweeping benefits promised by vendors. Users often dislike AI features, distrust companies that impose them, and receive systems that produce inconsistent or undesirable results. The costs are not incidental: environmental damage, copyright violations, potential real-world deaths, and the exploitation of underpaid, traumatized labor are built into the system.
The apparent contradiction is that AI remains ubiquitous. Marcotte interprets this not as evidence of success but as evidence of artificial support: investment capital, government contracts, and organizations’ fear of falling behind keep deploying a technology that would not exist at its current scale on demonstrated merit alone.
The concluding shift is political rather than technical. AI’s practical value, he argues, often lies less in improving products or productivity than in changing workplace relations—pressuring employees to comply, weakening collective power, and making managerial decisions appear technologically necessary.
Core ideas
Failure is a useful category
Calling LLMs a failure is not the same as claiming they never work. It distinguishes valuable exceptions from the technology’s broader tendency toward unreliable outputs, inflated promises, and harmful deployment.
Product failure is separate from social success
AI can fail as a useful, sustainable product while succeeding as an instrument of extraction, speculation, surveillance, and institutional power. Its continued visibility therefore does not prove consumer or social value.
Costs belong in the evaluation
A system should not be judged only by what it produces. Training and deployment involve energy and environmental burdens, unauthorized use of creative work, exploitative labor, and possible physical harms. These costs make mediocre output especially indefensible.
Adoption is being subsidized
The technology’s ubiquity may reflect investment, public contracts, and competitive panic rather than organic demand. “Everyone else is doing it” becomes a reason for adoption even when pilots and products do not perform well.
The workplace is the key site of analysis
Marcotte’s strongest claim is that AI adoption can function as a social mechanism: it imposes top-down change, demands employee compliance, and weakens worker bargaining power while presenting those political choices as neutral technological progress.
The title invokes Luddite history
“Stocking frames” recalls the machinery targeted by nineteenth-century textile workers. The reference reframes resistance to AI: opposition need not mean hostility to every machine, but can mean resisting technologies deployed to dispossess workers and protect owners’ power.
Practical takeaways
- Evaluate AI systems as products, not as promises: ask whether they solve a defined problem reliably, improve outcomes, and justify their full social and environmental costs.
- Separate a tool’s isolated usefulness from the institution’s reason for adopting it. A feature may help some users while still serving a broader program of surveillance, deskilling, or labor reduction.
- Treat mandatory AI adoption as a labor and governance question, not merely a technical upgrade. Ask who decided, who benefits, who bears the risks, and what workers can refuse.
- Be skeptical of adoption driven by fear of falling behind. Competitive pressure can sustain weak products and make organizational waste look like innovation.
- Use the language of failure to make exceptions more visible: preserve genuinely useful applications while rejecting the assumption that the entire technology deserves trust or expansion.
Caveats and counterpoints
- The essay is an argument and diagnosis, not a systematic review of all LLM applications. Its broad claim about failure leaves room for domains where users experience real benefits.
- Marcotte emphasizes harms and institutional power, so readers interested in engineering productivity or accessibility may find the treatment incomplete. A contemporary response argues that some experienced developers do find LLMs transformative, even while agreeing that the technology is overhyped and often imposed from above.
- The article refers to a recent MIT report and other evidence, but the recovered public text does not expose all footnotes or supporting details. The argument should therefore be read as a critical framework rather than a fully documented empirical case.
- The original page returned a 403 during research. The notes were reconstructed from an accessible DAIR zine reproduction and a reputable page that preserves the essay’s opening and context; the reproduction itself says it prints only the beginning and directs readers to the original for the remainder. ()
Questions worth revisiting
- When does a genuinely useful AI application remain acceptable if the surrounding business model depends on exploitative data, labor, or infrastructure?
- What evidence would distinguish AI adoption motivated by workplace control from adoption that produces durable productivity or quality gains?
- Could worker ownership, consent, and democratic governance make some LLM uses acceptable, or are the underlying costs too high?
- What would a non-AI alternative look like for the specific problems organizations are currently trying to solve?
Return to this when…
Return to this essay when an organization presents AI adoption as inevitable, when a product’s popularity is being confused with value, or when a debate about technical capability needs to be reframed around labor power, ownership, and deployment costs.