A recent New York Times opinion essay argued, in effect, that Silicon Valley is preparing itself for a “permanent underclass.”

That phrase is uncomfortable. It should be.

But it is also easy to misread. If we read it as “AI will certainly make most people unemployed and society will soon split into a few AI owners and a mass of useless workers,” it becomes a fear story. Fear stories spread quickly, but they are also easy to reject.

I think the more useful reading is this:

AI may not first affect ordinary people by making them unemployed tomorrow.
It may first affect them by compressing the value of the execution layer inside their work.

That is less dramatic than “mass unemployment.” It is also more realistic.

Do not only watch unemployment

When people talk about AI and work, they usually ask:

Which jobs will disappear?
Will unemployment spike?
Can AI fully replace people?

Those questions matter. But they may not be the earliest signal.

The earlier signals may look more ordinary:

A company hires one fewer junior worker.
A small team stops adding execution-heavy staff.
Drafting, editing, updating, testing, and checking now take half a day instead of three.
The tasks that used to train new workers become scarce.
The same job has less bargaining power.
The same team is expected to produce more, faster.

None of this has to show up immediately as “a job replaced by AI.” It changes where a person sits inside the workflow.

If you only watch unemployment, you may miss where the change is actually happening.

AI first takes describable, checkable, repeatable execution

The question “Will AI replace people?” is too coarse.

The better question is:

How much of your work is describable, checkable, repeatable execution?

If a task can be described clearly, given context, mapped to inputs and outputs, judged against acceptance criteria, and retried after failure, AI has an increasing chance of taking over that stretch of execution.

That does not mean AI can own the whole job.

But it can take over layers inside a job:

organizing materials
drafting a first version
making local code changes
updating copy in bulk
executing a set of issue-driven changes
repairing based on check results
producing something reviewable

These actions used to consume a lot of time. They were also how many ordinary workers proved value inside an organization.

If those actions are compressed, the person’s value gets recalculated.

The entry-level risk is not that AI is simply smarter than juniors

I do not think the issue is that AI is simply “smarter than juniors,” therefore juniors are useless.

The deeper issue is that many juniors used to grow through the execution layer.

You organize materials, check details, fix small bugs, update docs, run tests, maintain spreadsheets, follow up on process, and handle repetitive problems. Through those tasks, you learn the business, absorb context, and build judgment.

If AI takes over those tasks first, organizations face a practical problem:

Senior people remain valuable because they define problems and judge outcomes.
Junior people become fewer because their training tasks have been compressed.

This is not a story about AI instantly replacing everyone.

It is a story about a narrower career ladder.

Without new training paths, many people may not be directly replaced by AI. They may simply lose enough chances to move from execution to judgment.

That is why the “permanent underclass” phrase, while too strong as a prediction, should not be dismissed as pure drama. Its real warning is that if entry paths are cut off, long-term mobility becomes harder.

But a permanent underclass is not a technological destiny

This is where the argument needs discipline.

AI affecting work does not mean jobs automatically disappear.

A job being exposed to AI does not mean the job will be substituted away. Research from the ILO, OECD, David Autor, and Acemoglu and Restrepo all points to a more careful view: technology can substitute for some tasks, complement others, and create new tasks. The final labor-market outcome depends on adoption costs, process redesign, demand growth, institutions, education, training, and distribution.

So I do not think a permanent underclass has been proven as the future.

It is better understood as a risk that could be produced by incentives and institutions.

If the efficiency gains from AI are captured mainly by capital and a few platforms, while ordinary people absorb the transition cost; if firms keep reducing entry-level roles without building new training paths; if public policy waits until unemployment data turns ugly, then class immobility becomes more likely.

But if we recognize the problem and redesign education, training, workflows, distribution, and organizations, the outcome is not inevitable.

Technology is not destiny.

Workflow design, institutional choice, and individual migration speed all matter.

The ordinary worker has to move positions

I previously wrote that after seeing AI complete a large set of real issue-driven project changes, I felt, “I may be much closer to replacing myself than I thought.”

That sentence was emotionally true, but not precise.

The more precise version is:

It is not that I as a person am immediately unemployed.
It is that part of the execution-layer value I used to rely on is becoming less defensible.

If I am only the person who performs the execution step by hand, the risk grows.

But if I can move to another position, the value changes:

I can define the task
I can set the boundary
I can organize the context
I can judge the result
I can design the verification path
I can manage the evidence
I can decide whether the work moves to the next stage

This is not about dressing everyone up as a “manager.”

It is more basic: a person has to move from being a single-point executor to being responsible for a workflow.

You do not necessarily have to write every line of code, draw every diagram, or draft every sentence by hand. But you do have to know why the task exists, where the boundary is, what counts as acceptable, where the risks are, how evidence is preserved, and when the work must stop.

That layer of ability becomes more important, not less.

Using AI is not enough

Many AI tutorials tell people to learn prompts, learn tools, and automate more.

That is useful.

But it is not enough.

The real questions are:

Can you turn a vague request into a task AI can execute, check, and review?
Can you judge which parts of AI's result are reliable and which parts are dangerous?
Can you turn one execution into a more stable process for the next one?

If not, you are still being pulled along by the tool.

If yes, you are beginning to turn AI’s execution capacity into your own workflow capacity.

That is why I increasingly think of AI less as a chat tool and more as a worker inside a workflow.

A chat tool answers.

A worker takes on tasks.

The first changes information access. The second changes the structure of labor.

The danger is not only that AI is strong. It is that you still live in the execution layer.

The NYT essay’s phrase is heavy.

“Permanent underclass” is uncomfortable, and it should be.

But I do not read it as a conclusion. I read it as a question:

If AI first takes over the execution layer,
what lets ordinary people keep moving upward?

I do not have a complete answer.

But one thing is becoming clear: do not only ask whether AI will replace you. That question may arrive too late.

The earlier question is:

Is my value still mainly tied to describable, checkable, repeatable execution?

If the answer is yes, it is time to move.

Move where?

Toward defining tasks, organizing context, judging results, managing evidence, designing workflows, and owning boundaries.

That is not an easier position.

But it is closer to where ordinary people can still keep agency in the AI era.

References and further reading