And here I am, using Codex to research this piece, refine my notes, and edit my words into a decent-sounding article. This is the thing I am trying to name.
Like many of us, I’ve been watching how AI CEOs talk about work. Our work! Not just what they predict, but how the pitch has moved through the new gold rush: from reassurance, to inevitability, to turning their products into the place where work gets done.
The labor story is not only that AI may replace work. It is that AI products are being sold as places where work happens, while people remain responsible for judgment, liability, and cleanup.
They do not have to replace our labor to become indispensable to it. We pay them to spin our wheels.
By September 2024, the rhetoric was already split. In March 2021, Altman could sound blunt about AI making labor cheaper and capital stronger. By May 2024, even the softer public version was admitting that many current jobs, and some job classes, would disappear.
Anthropic, led by Dario Amodei, was less soothing. Its early safety writing treated the automation of knowledge work as a real possibility, not just a paperwork problem. So the split was already there before the agent era took over: softer public reassurance on one side, and broad economic warnings about labor, knowledge work, and power on the other.
That is what the rhetoric has been drifting toward.
-
Mar 2021-Mar 2023
The abstract baseline
Before the workplace pitch, it was a warning.
Before the pitch became a job tool, the public claims were broad: cheaper labor, capital power, automatable knowledge work, and a society that needed time to adapt.
Altman argues that software which can think and learn will do more of the work people do now, shifting power from labor to capital. OpenAI later frames job displacement as a deployment problem society has to learn through. Anthropic says most or all knowledge work may become automatable.
Moore's Law for Everything Planning for AGI and beyond Core views on AI safety -
May 2023-Sep 2024
The reassurance layer
Take it easy: tasks, not jobs.
First, the soft version: a tool for tasks. Some jobs may go, but better jobs will appear. Humans will still have things to do.
At the Senate hearing, Altman says labor disruption is one of his greatest fears, but also frames GPT-4 as good at tasks, not jobs. By May 2024, he is more explicit that many current jobs and some job classes will go away. By September, he is still saying most jobs will change more slowly than people think.
Senate hearing coverage MIT Sloan conversation The Intelligence Age -
Jan-Feb 2025
The coworker arrives
The labor market may change, but humans stay central.
Then: actually, the labor market may change. Some tasks will go away. Some jobs will be reshaped. The balance between capital and labor could get "messed up." But humans still have agency, taste, judgment, and willfulness.
OpenAI moves from tools people use to agents that materially change company output, then imagines junior virtual coworkers multiplied across knowledge work.
Reflections Three Observations -
Jun 2025
The visible crack
Coding stops being special human terrain.
Then: coding is the first visible crack. The abstract labor story becomes concrete because software work is both economically important and easy for the industry to inspect.
Altman writes that 2025 has brought agents doing real cognitive work and that writing software will never be the same.
The Gentle Singularity -
Aug 2025-Jun 2026
The loop leaves coding
The loop moves into law, recruiting, and finance.
Then: the pitch leaves coding and moves into professional loops. Legal work becomes the clean example. OpenAI's Codex research gives the broader internal signal: Legal, Finance, Recruiting, and Engineering all crossed into Codex as their primary AI tool.
Harvey describes GPT-5 as powering a legal coworker that plans, reasons, executes, branches, loops, and self-corrects while keeping users in the driver's seat. Amodei's warning also hardens: entry-level white-collar work is exposed.
OpenAI Harvey case study Claude Legal Solutions Building a Legal Coworker with GPT-5 The Adolescence of Technology How agents are transforming work
This is the strange bargain we are accepting: they do not have to replace our labor to become indispensable to it. The loop can be sold as progress, while we remain responsible for deciding whether all the shenanigans mean progress.
They do not have to know exactly how every lawyer, recruiter, finance analyst, or operator does their job. Or even how digital products are built, now that Coding Is Solved™. The product becomes a workflow machine: bring your documents, judgment, liability, and domain context. We supply the accountability layer. They supply the wheel.
Once that loop is in, the job is no longer just doing the work. It is deciding what to delegate, when to trust the outcome, when to intervene, and who is responsible when the loop produces something plausible and wrong.
That is the distinction I want to keep intact: a tool helps us do the work. A loop makes us responsible for work happening inside someone else’s system.
AI companies have succeeded at inserting themselves into our workflows and our lives. The next skill is tuning our awareness: noticing when the tool helps us think, and when it becomes the place where thinking has to happen.
That is the line I want to keep watching. How many decisions, and how much knowledge, are we surrendering before we notice? And when the loop feels smooth, can we still explain the work without the system that now contains it?