The next stage of artificial intelligence may not be just faster machines. It may be machines helping build the next machines.
For most people, artificial intelligence is only a tool.
Open a screen, type a question, AI writes an answer, fixes a sentence, makes a list, summarizes an email, writes a bit of code, or explains something you did not understand.
That is the public face of AI.
But what happens behind the curtain? AI is no longer only being used by ordinary people to write blog posts, emails, recipes, business plans, or school papers. It is being used by AI companies themselves to build the next generation of AI.
Anthropic recently called attention to this in a piece titled “When AI builds itself.” The company said that, inside Anthropic, AI systems are now handling a growing share of the work involved in developing AI. That includes coding, debugging, building internal tools, and speeding up technical research. Anthropic says its engineers are shipping about eight times as much code per day as they did in 2024, largely because Claude is writing much of that code while humans direct, review, and correct it.
Anthropic is careful to say that full recursive self-improvement has not arrived yet. In plain language, recursive self-improvement would mean an AI system becomes capable of designing, developing, and training a more powerful successor with little or no human involvement. Anthropic says this is not inevitable, but it could come sooner than many institutions are ready for.
Pause over that sentence a minute.
Not inevitable.
Not here yet.
But possibly coming faster than society is prepared to handle.
The Feedback Loop
The old model of technology development was fairly easy to picture. Humans built a tool. Humans improved the tool. Humans used the improved tool to do more work.
Now the loop is changing.
Humans build AI.
AI helps humans build better AI.
At some point, if the human role keeps shrinking, the loop may no longer be mostly human. It may become machine-assisted, then machine-led, then perhaps machine-driven.
That is the feedback loop people are talking about.
This matters because AI development already moves faster than most people can follow. A model is released. A new one follows. Coding tools improve. Image tools improve. Agents become more capable. Research tools get better. Benchmarks fall. Something that looked impossible a year ago starts looking routine.
Anthropic says AI’s ability to complete tasks on its own has been doubling roughly every four months, and Reuters reported that Anthropic is now calling for major AI labs to consider a coordinated, verifiable pause if risks grow too quickly.
That should get our attention.
Because this is not only a story about technology. It is a story about speed.
And speed has a cost.
The Cost Is Not Just Money
Most AI coverage still talks about cost in financial terms.
How much did the company raise?
How much is the model worth?
How much are the chips?
How much are the subscriptions?
How much money can be made?
But the deeper cost is not only money. The deeper cost is what gets consumed, displaced, strained, or ignored while the race continues.
AI is often described as if it lives in the air. It does not.
It lives in data centers.
It lives in power grids.
It lives in cooling systems.
It lives in land deals, water withdrawals, rare minerals, server farms, transmission lines, and local zoning meetings.
The International Energy Agency says global data center electricity demand is expected to more than double by 2030 to around 945 terawatt-hours, slightly more than Japan’s current total electricity consumption. The IEA also says AI is the most important driver of this growth, along with rising demand for other digital services.
A 2026 report from the United Nations University Institute for Water, Environment and Health makes the same basic point in broader terms: AI is not only a digital technology. It is a material system with measurable costs in energy, water, carbon, land, minerals, and electronic waste.
This is the part that often gets buried.
Every cheerful answer from a chatbot sits on top of a physical system. Every new model has a footprint. Every jump in capability means more demand somewhere.
More chips.
More servers.
More cooling.
More electricity.
More water.
More land.
More pressure on communities that may have very little say in what is being built around them.
The Local People at the Edge of the Machine
When an AI company says development is accelerating, investors hear opportunity.
Engineers hear capability.
Governments hear competition.
But people may hear something else.
They may hear a new data center is coming to farmland nearby.
They may hear the electric grid needs upgrades.
They may hear water use is increasing.
They may hear their county is offering tax incentives.
They may hear jobs are coming, but not always the kind of jobs that replace what is being disrupted.
They may hear that the future is arriving, whether they asked for it or not.
That is where the AI story becomes a ground-level story.
It is not only about whether an AI system can write code. It is about where the power comes from to run that system. It is about who pays for the grid expansion. It is about what happens to water in dry places. It is about what communities are asked to sacrifice in the name of innovation.
The United Nations University report frames AI’s environmental footprint as a governance and justice issue, not only a technical one. The benefits of AI may spread across the world, but the burdens of data centers, water demand, electricity use, mineral extraction, and e-waste can fall heavily on specific communities.
That sentence ought to be written on the wall of every planning commission meeting where a data center is proposed.
The Human Cost of Acceleration
There is another cost too.
The human cost.
Anthropic’s own Economic Index has been tracking how Claude is being used across the economy. In one 2026 report, Anthropic said about 49% of jobs had seen AI usage for at least a quarter of their tasks. The company also noted that AI’s effect on work is complicated because task coverage does not automatically mean full job replacement. Still, the direction is clear: AI is moving deeper into ordinary work.
This does not mean every person loses a job tomorrow.
It means work as we have known it, is changing.
It means fewer people may be needed for some kinds of tasks.
It means one worker may be expected to produce far more than before.
It means companies may decide that what once required a team can now be done by one person with AI tools.
It means the pace of work may speed up because the machine never tires.
That is another feedback loop.
AI helps workers produce more.
Companies expect more.
The new output level becomes normal.
Then the next AI tool arrives.
The human being is not always replaced. Sometimes they are simply stretched thinner.
The Safety Question
Anthropic is not presenting recursive self-improvement as a simple success story. The company is warning that if AI systems become capable of fully building their own successors, then securing, monitoring, and shaping those systems becomes much more important.
Reuters reported that Anthropic has called for a coordinated and verifiable pause among major AI labs if development risks rise too sharply. The company argues that a pause by one lab alone would not solve much, because competitors could keep moving. A meaningful pause would require coordination among multiple well-funded frontier labs.
Sound reasonable, and also raises an uncomfortable question.
If the companies building the most powerful AI systems are now warning that society may need a way to slow them down, why did the race get this far without that system already in place?
This is the problem with acceleration. The machine gets built first. The rules come later.
The land is cleared first.
The water is promised first.
The grid is strained first.
The workers adjust first.
The public finds out later.
The Real Question
The question is not whether AI can help write code.
It can.
The question is not whether AI can speed up research.
It already is.
The question is not whether AI companies will keep using AI to build better AI.
They will.
The real question is whether society will be allowed to understand the full cost before the loop becomes too fast to govern.
Because recursive self-improvement is not just a technical phrase. It is a warning about momentum.
Once a system helps improve itself, the timeline changes. The pressure changes. The balance of power changes. The distance between one generation of technology and the next begins to shrink.
And while that happens, the physical world remains physical.
A river is still a river.
A power grid is still a power grid.
A town is still a town.
A worker is still a person.
A field cleared for a data center is no longer a field.
AI may be becoming more capable, but the question remains the same one we should have been asking all along:
What is being spent to make it so?
And who, exactly, is being asked to pay?
Source Notes to Link at Bottom
Anthropic, “When AI builds itself.”
Reuters, “Anthropic urges AI labs to pause development, warns humans risk losing control.”
International Energy Agency, Energy and AI, Executive Summary.
United Nations University Institute for Water, Environment and Health, “The Environmental Cost of Artificial Intelligence: Carbon, Water and Land Footprints.”