• NATURAL 20
  • Posts
  • Sam Altman and Dario Amodei walk back their AI job apocalypse predictions

Sam Altman and Dario Amodei walk back their AI job apocalypse predictions

Author

Wes Roth
May 28, 2026

The AI jobs apocalypse story is starting to crack.

For years, the loudest message from Silicon Valley was that AI was coming for white-collar work. Not just routine factory work. Not just call centers. The claim was that AI was about to hit the laptop class: junior analysts, coders, support reps, writers, assistants, designers, and entry-level office workers.

Sam Altman warned that entire job categories could disappear. Dario Amodei warned that AI could wipe out up to half of entry-level white-collar jobs and push unemployment sharply higher. Executives, investors, and workers all began preparing for a world where the next model release might suddenly make millions of people economically redundant.

Now the same story is… well it’s being reversed:

Altman says he is "delighted to be wrong" about how much AI has eliminated entry-level white-collar work so far. Amodei is being framed less as a job-apocalypse prophet and more as someone describing AI as a productivity multiplier. Meanwhile, Dan Shipper at Every has published one of the clearest explanations for why the simple apocalypse story may be wrong: the more his company automates, the more human work appears.

The real story is not "AI does nothing to jobs."

That is obviously false.

Companies are cutting jobs. AI is being used to justify layoffs. Routine junior work is under pressure. Entire workflows are being rewritten.

But AI will CHANGE the shape of work, not destroy it.

It collapses the middle of tasks, floods the world with cheap competence, creates a mountain of generic output, and makes human judgment, taste, context, accountability, and review more valuable.

The job does not disappear all at once.

The job mutates.

The Headline Shift

Sam Altman just made a major tone change.

At a Commonwealth Bank of Australia conference in Sydney, Altman said:

  • "I don't think we're going to have the kind of jobs apocalypse that some of the companies in our space advocate or talk about."

  • "I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened."

  • "I'm delighted to be wrong about that."

That is striking because Altman has spent years warning that AI would dramatically transform employment.

He has previously said AI would "probably replace most of the jobs people do today." He has said entire job categories could be "totally, totally gone." Business Insider quoted him saying his high-level scorecard is that OpenAI and the AI industry have been roughly right on technological predictions, but "pretty wrong on the social and economic implications."

That line matters.

It suggests the frontier labs may have been better at predicting model capability than predicting the economy around those models.

That is the key distinction.

Benchmarks are not labor markets.

A model can crush a test and still fail to replace a job because a job is not just a task. A job is trust, coordination, responsibility, context, compliance, customer relationships, internal politics, brand judgment, taste, and accountability wrapped around a bundle of tasks.

AI can automate the middle of a task.

But someone still has to decide what task matters, whether the answer is good, who owns the mistake, and what happens next.

Dario Amodei's Shift

Dario Amodei gave one of the scariest mainstream AI labor warnings.

He said AI could eliminate around 50 percent of entry-level white-collar jobs. He warned unemployment could rise to 10 to 20 percent. That warning mattered because Amodei is not a random hype merchant. He runs Anthropic, one of the most important frontier AI labs, and he has built a reputation around taking AI risk seriously.

But recent coverage says Amodei is now framing the impact less as mass replacement and more as a productivity transformation.

The Decoder summarized the newer framing this way:

  • If you automate 90 percent of the job, then everyone does the remaining 10 percent.

  • Then that 10 percent scales back into a full job because output expands and demand rises.

That is a very different story from "half of all jobs vanish."

It is closer to the Jevons paradox version of AI work:

  • Make a task cheaper.

  • More people do the task.

  • The volume of work explodes.

  • The bottleneck moves from production to judgment.

In that world, AI does not end work.

It makes work faster, cheaper, weirder, more abundant, and more dependent on the humans who can decide what good looks like.

Why The Walkback Matters

This is not just another AI jobs take.

It is a narrative reversal from the two most important AI labs in the world.

For years, the public has been told two incompatible stories:

  • AI will make everyone richer.

  • AI will destroy your job.

That tension has created fear, backlash, and confusion. It has also made AI companies politically vulnerable. If your product is marketed as the thing that wipes out the middle class, do not be surprised when workers, regulators, and voters turn against you.

Now OpenAI and Anthropic are heading toward enormous capital needs, possible public-market scrutiny, and trillion-dollar valuation narratives. In that environment, "our product will destroy jobs" becomes a terrible investor story, customer story, and regulatory story.

So there are two ways to read the shift.

The generous reading

  • The CEOs updated because the data changed.

  • The labor market has not shown a clean AI unemployment shock yet.

  • Adoption is slower than model progress.

  • Companies are discovering that AI is expensive, uneven, and hard to integrate.

  • The job is more complicated than the task.

The cynical reading

  • The narrative became politically dangerous.

  • Public opinion on AI is negative.

  • Regulators are watching.

  • Workers are scared.

  • Investors want growth, not social backlash.

  • AI companies need the story to become "productivity and abundance," not "mass unemployment."

Both can be true.

The walkback can be partly sincere and partly strategic.

But either way, it creates a new opening: we can stop talking about AI jobs as a cartoon apocalypse and start talking about the actual mechanics of work.

That is where Dan Shipper's piece matters.

Dan Shipper's Every Case Study

Dan Shipper is the CEO of Every, an AI-native media and software company that has pushed farther into agentic workflows than most businesses.

Every uses Codex and Claude Code across coding, writing, design, customer service, and internal operations. They alpha-test models from OpenAI, Anthropic, and Google. They use Slack agents, embedded support agents, AI email workflows, and collaborative coding agents.

If the simple automation story were true, Every should be the place where the humans disappeared first.

That is not what happened.

Shipper says Every has automated everything it can, but there is more human work than ever.

They still hire engineers. They still hire writers and editors. They still use human customer support people. They have not ditched SaaS tools for disposable vibe-coded replacements. They have not fired the team and replaced it with agents.

The work looks different, though.

  • They do not write code by hand in the old way.

  • Managers commit code like individual contributors.

  • Engineers talk directly to customers.

  • If someone is mentioned in Slack, it might be a human or an agent.

  • AI answers 95 percent of Shipper's work emails.

  • He is almost always at inbox zero.

  • But he still reviews the work.

That last point is the whole story.

AI did not remove the human.

It moved the human.

The human is no longer always typing the first draft. The human is framing the work, reviewing the result, judging quality, handling exceptions, and deciding what should happen next.

The Human Sandwich

Every has a useful phrase for this: the human sandwich.

The structure is simple:

  • Human sets the frame.

  • AI collapses the task.

  • Human judges and extends the result.

This is one of the cleanest mental models for the current era of AI work.

AI does the middle.

Humans still own the beginning and the end.

The beginning matters because the agent needs the right goal, context, constraints, and success criteria. The end matters because someone has to judge whether the output is correct, useful, safe, on-brand, and worth shipping.

That is why the human does not vanish from the workflow.

If anything, the human role becomes more leveraged and more responsible.

The person who knows what to ask, what to accept, what to reject, what to fix, and what to do next becomes more valuable, not less.

The Two Modes Of AI Work

Shipper says work with agents is settling into two modes.

1. Agents as employees

These are agents you delegate work to.

They live in Slack. They have names. They have jobs. You can tag them and ask them to do things. They can also be embedded inside workflows, like customer support.

Every uses agents like:

  • Claudie, which helps the consulting team write proposals, draft training decks, and track project todos.

  • Andy, which helps the editorial team collect interesting internal ideas and turn them into newsletter inputs.

  • Viktor, which gathers growth metrics, analyzes surveys, and turns messy discussions into memos.

  • Fin, which handles customer support conversations.

Fin is the clean support example.

In one recent week, Fin participated in 65 percent of 202 Every support conversations. It closed 81 conversations without a human, which was 40.1 percent of all actionable conversations.

That is real automation.

But the support job did not disappear.

The human support manager now spends less time answering basic tickets and more time handling complex cases, improving the system, supervising the AI layer, and building better workflows.

The pattern is clear:

  • AI absorbs repeatable work.

  • Humans move into exceptions, supervision, system design, and quality control.

2. Human-agent collaboration

This is the mode Shipper thinks is more important.

Tools like Codex, Claude Code, and Claude Cowork are not just chatbots. They are becoming operating systems for knowledge work.

Humans and agents work in the same environment. Multiple agents can run in parallel. The agents can access files, data, code, and project context. The human can interrupt, redirect, review, and decide what happens next.

This is not simply asking ChatGPT a question.

It is more like managing a team inside your computer.

That means the worker's job changes from "do the task manually" to:

  • Define the work.

  • Assign the work.

  • Monitor the work.

  • Review the work.

  • Improve the system.

  • Decide the next move.

That is management and it’s not going to be automated.

Cheap Competence Creates More Work

The central economic mechanism is simple: AI makes competence cheap.

Models are trained on the visible residue of human expertise:

  • Code.

  • Essays.

  • Emails.

  • Designs.

  • Support tickets.

  • Product specs.

  • Documentation.

  • Sales decks.

That means AI can reproduce the average patterns of existing work. A person who could not write code can now generate a pull request. A marketer can create thumbnail ideas. An engineer can draft product documentation. A support person can produce internal process docs.

This is a huge shift.

But when something becomes cheap, people use more of it.

Cheaper code means more software gets written. Cheaper writing means more writing gets produced. Cheaper design means more variations get tested. Cheaper support means more interactions get handled. Cheaper research means more questions get asked.

The first-order effect is not that everyone stops working.

The first-order effect is that everyone produces more stuff.

Then the organization has a new problem: what do you do with all that stuff?

The Slop Problem

Shipper's definition of slop is useful because it avoids the lazy version of the argument.

Slop is not one specific mistake.

It is not an overused phrase. It is not a certain sentence rhythm. It is not a design cliche. It is not just bad punctuation or generic AI phrasing.

Slop is sameness at scale.

It is what happens when many people, in many contexts, use the same models trained on the same corpus without enough human judgment.

The problem is not that the output is always bad.

The problem is that it becomes interchangeable.

If everyone can make a decent report, a decent app, a decent thumbnail, a decent landing page, or a decent product brief, then decent becomes a commodity.

That is where humans come back in.

The market stops rewarding generic competence and starts rewarding difference.

People want work that feels:

  • Specific.

  • Alive.

  • Contextual.

  • Tasteful.

  • Correct.

  • Weird in the right way.

  • Right for this company, this customer, this audience, this moment.

AI makes the first draft cheap.

That makes the final judgment more important.

Why Experts Become More Important

This is the most important reversal in the piece.

AI does not make experts irrelevant. It creates more situations where expert judgment is needed.

If operations people submit AI-generated pull requests, engineers need to review them.

If marketers create AI-generated thumbnails, designers need to sharpen them.

If engineers draft product guides, writers and editors need to make them actually good.

If support agents answer routine tickets, humans need to handle the hard edge cases and maintain the system.

Expert work moves upward.

Experts become:

  • Reviewers.

  • Editors.

  • System designers.

  • Taste-makers.

  • Quality controllers.

  • Workflow architects.

  • Accountability holders.

This is not the end of expertise.

It is the relocation of expertise from production to judgment.

The New Expert Job

The expert's job becomes less about producing every artifact from scratch.

It becomes more about building systems that turn cheap AI output into excellent final work.

That means building:

  • Review queues.

  • Evals.

  • Agent instructions.

  • Repo rules.

  • Claude and Codex instruction files.

  • Continuous integration checks.

  • Permissions.

  • Workflow design.

  • Feedback loops.

  • Final approval systems.

Automation is not magic dust.

It is infrastructure.

It has to be maintained.

It has to be governed.

It has to be reviewed.

It has to be made specific to the business.

That is why AI-native companies can automate aggressively and still need humans.

The humans are not just doing the old work slower.

They are building and supervising the machine that does the work.

Why The Job Apocalypse Has Not Arrived On Schedule

There are several reasons the simple apocalypse story has not shown up cleanly in the labor data yet.

1. Adoption is slower than capability

Models improve fast.

Organizations change slowly.

A benchmark can jump overnight. A company cannot safely rebuild its workflows overnight.

Brookings has warned that AI capability is not automatically translating into broad economic gains or meaningful adoption. Diffusion is uneven across sectors, firms, countries, and worker groups. Digitally mature companies move faster. Others lag.

2. AI is expensive and uneven

AI demos look cheap.

Production AI systems are not always cheap.

Companies have to pay for compute, tokens, monitoring, integration, security, compliance, review, and failure handling.

TIME noted that some major companies are struggling to justify AI costs. Uber's CTO reportedly burned through a Claude Code budget far faster than expected. Microsoft reportedly moved to cancel some internal Claude licenses because token costs were high and unpredictable.

The point is that using AI well is not the same as buying access to a model.

3. Jobs are bundles, not tasks

This is the biggest conceptual mistake in the job-apocalypse conversation.

A job is not one task.

A job is a bundle of tasks plus context, trust, coordination, judgment, accountability, customer knowledge, and institutional memory.

AI can automate tasks inside a job without replacing the full job.

That still changes the job.

It may reduce headcount in some areas.

But it does not imply instant mass unemployment.

4. Cheap output creates new demand

If AI makes knowledge work cheaper, people will ask for more knowledge work.

More versions. More experiments. More documentation. More analysis. More personalization. More customer interaction. More software. More content. More design.

This is the productivity paradox inside the office.

When the cost of work falls, demand often expands.

The bottleneck moves upward.

________________________________

Let’s keep an eye out for how this develops, but certainly this might be a big change in how the future looks.

Wes “back to managing agents” Roth

Sources

Reply

or to participate.