Machine Learning as a Shaping Force
Every technology reshapes the world it enters, but personal automobiles offer a particularly useful lens for thinking about what large language models will do to society. Cars were fast and convenient — everybody knew that immediately. But their true legacy is in how they remade streets, killed off urban horses and the transit systems that once served American cities, decentralized populations into sprawling suburbs, bulldozed Black communities for highways, poisoned a generation with leaded gasoline, and became a leading cause of death for young people. Even now, a third of Americans don't drive, yet most of the country is structured as if everyone does.
This is the sense in which we should think about "AI": not as a tool with convenient features, but as a technology whose second-order effects will determine the shape of daily life for decades. The space of possible futures in 2026 feels much broader than it did in 2022 — and most of those futures feel bad.
The Present Slop
Much of the bad future is already here. LLM-generated material floods search results, doctor's offices, and customer service queues. Engineers and contractors use chatbots to produce confident falsehoods. Electric utilities blame data center demand for rate hikes. Scrapers operated by model vendors degrade the small websites many of us rely on, while synthetic video of suffering animals and generated pages misrepresenting police brutality circulate freely. There's LLM spam in inboxes, synthetic CSAM on moderation dashboards, and people outsourcing meals, travel plans, art, and relationships to ChatGPT.
Professionally, the pressure is visible everywhere: pull requests full of generated code, vaporware presented for analysis, clients asking Claude to do work they might once have hired a human for. The generated code is still bad, mostly — but that could change. The fear isn't hypothetical. Reading, thinking, and writing sit squarely in the blast radius of large language models, and retraining into another profession only postpones the problem until machine learning eats that field too.
The Case for Slowing Down
The most defensible response is to stop: refuse to outsource thinking and writing, and refuse to consume what has been outsourced by others. ML assistance demonstrably reduces performance and persistence, and it denies the practitioner both the muscle memory and the deep theory-building that come from working through a problem by hand — what James C. Scott would call metis. Writing by hand, reading cookbooks by human authors, and talking through problems with friends are not Luddite gestures; they are how people stay grounded in their own capabilities.
Concretely, that means calling out colleagues who send slop, flagging ML hazards at work, canceling ChatGPT subscriptions, and resisting employer mandates to adopt Copilot — worth noting that Microsoft itself describes Copilot as being "for entertainment purposes only," not for serious use. It means unionizing and pushing back on management demands, calling members of Congress to demand regulation that holds ML companies responsible for their carbon and digital emissions, and opposing tax breaks for ML datacenters. For people working at Anthropic, xAI, or similar organizations, it means thinking seriously about their role in constructing this future — and possibly quitting.
Slowing the advancement of ML is not about stopping it entirely. There remain plenty of people committed to building it. But today's models are already capable, and the effects of existing technology will take years to fully unfold. Every delay buys time to adapt: time to manage technical debt and errors in legal filings, to prepare for synthetic CSAM, sophisticated fraud, obscure software vulnerabilities, and whatever comes next, and time for workers to find new jobs. For anyone who leaves an otherwise good job on ethical grounds, there is also the quiet benefit of a clear conscience. If the caution turns out to be misplaced, the technology can always be built later.
And Yet…
Despite the bitter distaste this generation of ML systems inspires, they do seem useful, and the temptation to use them is real. Consider a constrained problem: a set of color-changing lights speaking an unfamiliar protocol. Working it out from scratch could take a month of digging through manuals — or an LLM could write a client library in minutes. The security consequences are minimal. The use case is narrow and verifiable by hand. No technical debt is pushed onto anyone else. The harm, in such a case, is not obvious.
The question is whether it is possible to use these tools in isolated, verifiable cases without the practice seeping into everything else — and whether the cumulative effect of millions of such individual compromises is what shapes the future most of all.
