Reading Obsolete: ethics and craft in an AI labour race

Garrison Lovely’s Obsolete: The AI Industry’s Trillion-Dollar Race to Replace Us—and How to Stop It (OR Books) is one of the clearer popular books of this moment. It refuses the false binary of “harmless bubble” versus “inevitable god-machine,” and keeps the eye on a simpler, uglier programme: the best-funded attempt in history to automate as much human labour as possible.

I write Magento and systems software with AI tools every day. That does not make me an industry cheerleader. It makes me a stakeholder who has to hold two truths at once: the tools are already reshaping craft; the political economy around them is not neutral.

Ethics first

The ethical core of Lovely’s argument is not “never use a model.” It is closer to: stop pretending the race is a weather event. Firms are choosing to pour capital into substitution. Workers, customers, and polities can choose whether that race proceeds uncontested.

• Obsolescence is a project, not a fate. Treating replacement as destiny is how consent is manufactured.
• Do not shame people for using AI at work while the workplace and the market are being redesigned around it. Shame is a poor substitute for organisation, standards, and democratic control over high-stakes systems.
• Present harms and future risk both count. Fraud, surveillance, deskilling, energy cost, and labour leverage sit beside longer-horizon safety questions. Ignoring either side is a rhetorical trick.
• Accountability stays human. When generated code or copy ships, a named person and an organisation own the consequences — liability, reputation, outages, bias.

Technical craft (what I actually do)

Ethics without practice becomes piety — moral talk with no craft behind it. Practice without ethics becomes capture — skill absorbed by someone else’s extractive agenda (labour replacement, dark patterns, liability laundering) instead of serving the people the work claims to help. On the technical side, the habits that keep AI usable for serious work look like this:

• Amplification, not abdication. Models accelerate what the operator can already evaluate. If you cannot read the output, speed is a defect factory.
• Constraints over vibes. Guardrails, runbooks, tests, and endpoint smoke checks turn “sounds right” into “passes a gate.”
• Strategy stays human. Choosing the problem, the trade-off, and the stop condition is not something I outsource. Execution can be assisted.
• Open, inspectable stacks help. When dependencies are readable, assistants can reason over behaviour instead of guessing at opaque boundaries — and so can reviewers.
• Refuse drive-by paste. The public failure mode of AI coding is unreviewed fluency. The fix is process, not nostalgia for typing every character by hand.

Where I land

I will keep using Cursor and related tools where they make careful work faster. I will also keep saying that labour substitution at scale is a political choice, not a natural law — and that practitioners who benefit from the tools owe honesty about that choice.

If Obsolete does its job, it does not tell developers to unplug their editors. It asks a harder question: who decides what “done” means when machines can produce plausible work at industrial volume — and how we keep human judgment, solidarity, and public control in that loop.

Further reading: Obsolete at OR Books and Lovely’s notes at obsolete.pub.

Written with the help of an AI assistant.