AI Made It Possible: They Tried to Replace the Worker. They May Have Empowered the Competitor.

What AI music exposed first is now spreading across the rest of the economy: when advanced capability becomes cheaper, the biggest disruption may not be machines replacing people. It may be ordinary people and smaller organizations gaining capabilities that used to belong to institutions.

I have seen a version of this fight already.

Over the last year, I have spent a lot of time following AI music: the lawsuits, the label negotiations, the platform rules, the copyright arguments, the licensing questions, the distribution fears and the constant predictions about what creators supposedly would or would not be allowed to do.

What interested me was never simply whether a machine could generate a song. The deeper question was what happened once ordinary people gained access to production capabilities that previously required far more money, equipment, people and institutional access.

That is why I no longer think the music fight was an isolated fight.

Music was not the exception. It may have been the preview.

The fight became bigger than music

Generative AI is not one product or one industry. It reaches writing, software, design, customer support, research, education, marketing, video, translation, data analysis and increasingly the operation of software itself.

That breadth matters because the economic effect is different from simply inventing another piece of software. AI can reduce the cost of obtaining many kinds of cognitive and creative capability at once.

A musician who could not afford a traditional production team can experiment with arrangement, production and iteration. A five-person business can draft campaigns, analyze information, localize material, prepare customer communication and operate across more channels. A solo creator can perform pieces of research, editing, planning and production work that previously demanded specialists.

This does not mean one person literally becomes twenty fully qualified professionals. Expertise, accountability and context do not disappear. But it can mean that a person or small team gains enough additional capacity in specific workflows to compete differently.

And that is where the economic question gets interesting.

Capability is not autonomy

One of the most important mistakes in the AI debate is treating impressive capability as proof of dependable autonomy.

Those are not the same thing.

Frontier agents are getting better quickly. They can write code, use tools, browse interfaces and sustain longer sequences of work. That deserves to be taken seriously. But current benchmarks also show why organizations should be cautious about turning that progress into claims that complex human work can simply be handed over end to end.

METR, one of the organizations measuring frontier-agent task horizons, explicitly warns that its results are frequently overinterpreted. Its benchmark work is concentrated heavily in software and research tasks. Real jobs are often poorly defined, high-context, collaborative and messy. METR also notes that a 50% success horizon does not mean everything below that horizon can safely be delegated, particularly when reliability requirements are high.

That distinction gets even sharper when the assignment is difficult to verify. A model can produce an answer that looks plausible while drifting from the real mandate. Humans do this too. The problem is that software can now do it at extraordinary speed and scale.

That is why I think the useful distinction for businesses and governments is not whether AI belongs inside critical systems. It already does, and it increasingly will.

The distinction is between AI operating inside a critical system and handing the critical system over to AI without meaningful human accountability.

The recursive-self-improvement question

This is where the discussion becomes more difficult.

Recursive self-improvement is not a ridiculous subject. AI systems already help generate code, evaluate outputs, produce training material and improve parts of AI-development workflows. Researchers and frontier labs are actively studying whether these feedback loops could eventually accelerate AI development much further.

That deserves research, testing and safeguards.

But there is an important difference between bounded systems improving pieces of a process under human-created objectives and a hypothetical system that independently determines its own objectives, redesigns itself repeatedly and develops into an autonomous superintelligence.

The first category is already visible. The second remains a forecast.

METR itself cautions against projecting current time-horizon measurements into months- or years-long autonomous work. Recent warnings from leading AI executives show that serious people believe recursive self-improvement could become dangerous. That is a legitimate counterargument to complacency. It is not, by itself, proof that the predicted system exists today.

My concern is with allowing a forecast about what AI might become to erase the evidence about what AI currently is.

We can prepare for uncertain futures without pretending uncertainty has already resolved itself.

The replacement bet

A visible part of the corporate AI conversation has focused on a straightforward equation:

That approach is understandable. Businesses have always looked for ways to reduce costs. Some tasks really will require fewer human hours. Some jobs will shrink. Some positions will disappear altogether.

But I think that equation may be too small.

Suppose you have ten people. The first question does not have to be whether AI lets you operate with seven.

The first question could be:

Could they serve more customers? Run three marketing channels instead of one? Improve documentation? Produce more variations? Localize campaigns? Analyze more data? Develop a second product line? Respond faster? Test more ideas? Raise quality while increasing frequency?

That is a very different AI implementation strategy.

It treats the human being as the scarce source of direction, context, relationships and accountability—and AI as elastic capacity around that person.

What the labor evidence actually says so far

The current evidence does not justify a victory lap for either side.

The International Labour Organization's 2026 review finds that generative-AI productivity gains are real but uneven. It also says large-scale job displacement remains limited so far. Worker-reported time savings have not automatically translated into equivalent gains in output, earnings or employment.

That last sentence matters.

Buying AI does not create productivity by itself. Organizations still have to redesign processes, train people, establish quality controls and decide what to do with the saved time.

The counterargument is equally important: limited displacement so far is not evidence that future displacement will remain limited. The ILO specifically warns about inequality, younger workers losing entry-level opportunities, and changes to job quality and worker autonomy.

Some workers will be hurt by this transition. Pretending otherwise would be as irresponsible as pretending every job is about to disappear.

The evidence points toward transformation—but transformation can create winners and losers.

The small-business opportunity is real. It is not automatic.

This is the part of the AI economy I believe deserves much more attention.

OECD research published this summer finds opportunities for smaller firms from exposure to generative AI, while also finding that firms with stronger existing capabilities retain advantages.

That is an important warning against simplistic democratization rhetoric.

Giving everybody access to the same model does not make everybody equally capable.

A company with better data, better processes, stronger managers, deeper expertise and more capital can often extract more value from the same technology.

At the same time, lowering the cost of capabilities matters enormously to organizations that previously could not afford those capabilities at all.

For a huge corporation, AI may make an existing marketing department more productive.

For a five-person company, AI can mean something closer to: we finally have meaningful marketing capacity.

For an independent musician: I can explore production now.

For a small publisher: I can research, edit, prepare metadata and promote more consistently.

For a local business: I can communicate across several platforms instead of abandoning all but one.

The IMF estimates the labor-cost equivalent of time currently saved through observed AI usage at about $2.7 trillion annually across its sample, or 3.4% of combined GDP. The IMF is careful to say that this is an indicative measure of the value of time saved—not a direct GDP estimate. It also finds that the gains are distributed very unevenly.

That gives us the right conclusion:

Access matters. Education matters. Process matters. Judgment matters. The ability to turn capability into something useful matters.

Then the label fight starts to look different

This is why I keep returning to music.

Once AI music became more capable, the arguments immediately moved beyond sound quality.

Who owns the output? Who licensed the training material? Who controls distribution? What human contribution matters? What evidence should a creator keep? Which platforms will accept the work? What happens to an industry whose expensive production barriers have suddenly weakened?

Those were not merely music questions.

They were early examples of what happens when productive capability becomes cheaper and more distributed.

Now similar questions are moving through software, video, writing, advertising, professional services, education and small-business operations.

The fight is not simply about whether AI can perform a task.

It is increasingly about who gets to scale.

The strongest case against my argument

If I want this argument taken seriously, I have to state the opposing case as strongly as I can.

I accept all four of those points as legitimate.

They do not eliminate my thesis. They change how carefully I should state it.

I am not arguing that AI will never replace workers.

I am arguing that replacement is not the only economic effect worth measuring.

I am not arguing that advanced autonomous intelligence is impossible.

I am arguing that forecasting it is different from demonstrating it.

I am not arguing against AI regulation.

I am arguing that we should examine who can afford to comply and whether safety rules unintentionally become competitive walls.

And I am not arguing that giving a small business ChatGPT, Claude, Gemini or another AI tool magically turns it into a multinational corporation.

I am arguing that lowering the cost of capability changes the competitive possibilities.

Governance versus a competitive moat

This distinction becomes increasingly important as governments decide how advanced AI should be governed.

The OECD describes today's AI market as dynamic but uneven. Model capabilities are improving rapidly and some prices have fallen. At the same time, essential layers such as chips, cloud infrastructure, compute, data and specialized skills can be highly concentrated. Vertical integration and gatekeeping can strengthen already-powerful firms.

Open-source AI complicates that picture because it can reduce dependence on a small number of providers, lower entry costs and increase experimentation. It can also create genuine security and misuse concerns.

So the useful question is not “regulation or no regulation?”

The useful question is:

Safety can be legitimate.

It can also become a moat if the cost of compliance freezes smaller competitors out.

Those possibilities can both be true at the same time.

Maybe the technology was not the part that was oversold

There is another reason this debate is becoming urgent: the amount of capital being committed to AI is extraordinary.

The Bank for International Settlements estimates the five largest global technology companies will invest more than $1 trillion in AI across 2025 and 2026. Reuters reports industry-wide AI spending approaching $800 billion in 2026, with projections above $1 trillion in 2027.

That does not prove an AI bubble.

History gives us an important counterargument. Railway booms, telecommunications build-outs and the dot-com era all included spectacular excess while leaving behind infrastructure that enabled genuine long-term transformation.

The technology can be transformative while some of the financing assumptions surrounding it are still wrong.

Which raises a question I think deserves far more attention:

What if the enormous productivity opportunity is real, but the fastest route to that productivity is not eliminating as many salaries as possible?

What if the better return comes from reorganizing work so that capable humans can produce, test, analyze, serve and build at a scale that previously required much larger organizations?

That model does not sound as dramatic as “the machines replace us.”

Economically, it may be more disruptive.

When does the dam actually break?

In The AI Panic Machine: When Does the Dam Break?, I questioned how much of today's public debate is being driven by demonstrated present capability and how much is being driven by predictions about a future superintelligence.

This is the other side of that argument.

Maybe the dam does not break when a machine “wakes up.”

Maybe it breaks when millions of people realize they do not need to wait for that hypothetical event.

The capabilities already available are enough to change who can create, who can operate, who can launch, who can experiment and who can compete.

A songwriter can become a producer.

A solo operator can build a real publishing system.

A small company can communicate like a much larger one.

A worker who understands the tool can extend the range of what that worker can deliver.

None of those statements requires AGI.

They require access, judgment, learning and human direction.

What I am betting on

This is where evidence ends and my own conclusion begins.

I think some of the most durable AI winners will be organizations that learn to make people more capable rather than simply racing to remove them.

I think creators and smaller businesses have been handed access to an unusual amount of productive leverage.

I think incumbent institutions have legitimate reasons to worry about safety, ownership and economic disruption—and equally legitimate reasons to worry about new competition.

I think we should take advanced-AI risks seriously without converting predictions into facts before they have been demonstrated.

And I think human direction remains the part of the system that is easiest to undervalue precisely because the machine is becoming so impressive.

That is the opportunity.

That is also the disruption.

And that is why I still come back to the same phrase.

AI made it possible.

What happens next still requires us.