Information Has a Peculiar Economic Property. Then Came AI.

Author’s Note

I began thinking about this essay from a very simple observation: information behaves differently from most physical goods. It can cost years and millions of dollars to discover something, yet almost nothing for somebody else to reproduce the knowledge once it exists.

That asymmetry sits behind much of our intellectual-property system.

AI made me wonder whether the balance is changing again.

While researching the subject, I found that several parts of the argument are already visible in practice: publishers are licensing material to AI companies; websites are restricting AI crawlers; search users click original sources less often when AI-generated answers appear; coding tools are reducing some software-development costs; and AI systems are beginning to participate directly in product discovery and purchasing.

Other parts remain hypotheses.

We do not yet know whether AI will ultimately increase or reduce incentives to invest in original intellectual work. AI reduces the cost of creating as well as the cost of imitating. We do not yet know which effect will dominate.

Nor do we know how far consumers will delegate decisions to AI systems. That distinction matters to me.

The purpose of this essay is not to predict that intellectual property, marketing or the open web will disappear. It is to ask what happens when three costs decline simultaneously:

the cost of learning from what others have discovered,

the cost of reproducing what others have built,

and the cost of filtering the market before a consumer makes a choice.

The sources behind those observations are mentioned at the end of this essay.

Information has a peculiar economic property.

It can be extremely expensive to create and extraordinarily cheap to reproduce.

A scientist may spend years discovering something. An engineer may work through hundreds of failed designs before finding one that works. A pharmaceutical company may spend vast sums developing a medicine that ultimately becomes a few lines on a prescription. A writer may spend three years producing a book that can be copied in seconds.

The first copy can be expensive.
The next copy can cost almost nothing.

This imbalance is one of the reasons societies created intellectual-property systems.

Patents, copyright, trade secrets and related protections are very different legal instruments, but behind them sits a similar economic concern: if somebody bears the cost of discovering or creating something and everybody else can immediately reproduce its value without bearing that cost, why would anyone finance the original effort?

The system has never been perfect.

Intellectual property can protect too much or too little. It can reward genuine invention, protect incumbents, create monopolies, encourage disclosure or inhibit competition depending on how it is designed and used.

But the underlying problem is real.

Creation has a cost.
Imitation often costs less.

For a long time, the difference between those two costs helped shape the economics of innovation.

Then came AI.

AI Is Built from Two Very Different Kinds of Capital

When we talk about artificial intelligence, we tend to focus on the machine.

The chips.
The data centers.
The mathematics.
The increasingly sophisticated model architectures.

And the investment is enormous.

But another kind of capital sits underneath the system.

Humanity’s accumulated information.

Books. Scientific papers. Software. Images. Newspapers. Manuals. Legal writing. Discussion forums. Educational material. Music. Design. Public records. Research. Websites. And billions upon billions of ordinary human explanations, arguments, mistakes, corrections and observations.

Not all of this material is protected by intellectual property. Some is public domain. Some is openly licensed. Some is factual information that copyright does not protect. Some is proprietary. Some is copyrighted. Some was published precisely because its creator expected people to read it freely.

But taken together, it represents something extraordinary:

The accumulated intellectual output of civilization in machine-readable form.

AI companies did not create that historical corpus.
Human societies did.

Governments financed some of it. Universities financed some of it. Companies financed some of it. Readers bought books. Subscribers paid newspapers. Advertisers funded websites. Customers bought software. Investors funded research. Taxpayers financed science.

For centuries, we developed different mechanisms for paying people and institutions to produce knowledge.

Now another industry has discovered how to turn enormous quantities of that accumulated knowledge into a new productive asset.

That is where the argument becomes more complicated than simply asking whether AI companies should pay writers.

Because training is only the beginning.

The First Question: Who Pays for the Input?

The most visible conflict today concerns training data.

Can an AI company train a commercial model on copyrighted books? On journalism? On software repositories? On images? Does training constitute copying? Is it transformative? Does it compete with the original market? Does it matter whether the material was legally obtained?

These questions are still being fought over, and different cases have already produced different answers.

That alone tells us something important.

There is no simple established rule saying that AI companies can take everything for free.

But there is also no simple established rule saying that every piece of training data requires an individual licence and royalty.

Meanwhile, markets are moving faster than courts.

Large publishers and platforms have begun signing licensing agreements with AI companies.

That is significant. It means data is already acquiring an explicit market price.

But it also introduces an asymmetry.

A large publisher controlling decades of articles can negotiate. A platform controlling millions of discussions can negotiate. A major image archive can negotiate. A scientific publisher can negotiate.

What does an individual writer negotiate?
What does an independent blogger negotiate?
What does the person who spent ten years answering specialized questions on a forum negotiate?
What does the programmer whose explanations, documentation and examples became part of the general technical culture negotiate?

The question is therefore not merely whether licensing exists.

It is who possesses enough bargaining power to participate in it.

This Is Already Different from the Original Internet

The old internet had an exchange.

Google crawled your page. But Google also sent people back to it.
YouTube hosted your video. But it could provide an audience and, for some creators, advertising revenue.
Facebook distributed your article. Instagram distributed your photograph. Amazon listed your product.

The relationship between platforms and creators was frequently contentious, but there was usually some form of exchange.

You produced something.
The platform aggregated or indexed it.
The platform gained content.
You gained distribution.

It was not necessarily an equal bargain. But it was recognizable.

Give us material, and we help humans find you.

Generative AI introduces another possibility.

The machine can consume information from multiple sources, synthesize it, and answer the user directly.

The human may never reach the original source.

Ask how to solve a technical problem. The answer appears.
Ask for the meaning of a scientific concept. The explanation appears.
Ask for a summary of a historical event. The summary appears.
Ask which camera to buy. The shortlist appears.

The user has obtained value.

The original producer of some of the information underlying that value may receive neither money nor even a visit.

The relationship can move from:

crawl → index → referral
toward:
crawl → synthesis → answer

That is a fundamentally different economic structure.

And we are already seeing evidence that AI-generated answers reduce some click-through to original websites.

That does not mean websites disappear tomorrow. It means one of the economic assumptions behind the open web has begun to weaken.

The Web May Respond Rationally

For thirty years, publishing openly was usually advantageous.

You wanted search engines to find you. You wanted people to link to you. You wanted your documentation indexed. You wanted your articles discoverable. Being visible was part of the business model.

But suppose machine visibility and human visibility begin separating.

Suppose allowing an AI to read everything you publish makes the AI better at answering questions without sending the user to you.

Then openness has a different cost.

The rational response may be to restrict access.

And that has already begun.

Publishers are blocking some AI crawlers. Some are licensing access. Technology providers now allow website owners to distinguish between ordinary search crawling and AI crawling. Experiments are emerging in which automated systems may eventually have to pay for machine access.

This is the beginning of what I would call the permissioned web.

Not the end of the internet. Not everything disappearing behind a paywall. But a gradual distinction between information available for humans to read and information available for machines to ingest at scale.

That distinction barely mattered fifteen years ago. It may become foundational.

And here is the irony:

AI became powerful partly because humanity put so much knowledge onto an open, machine-readable internet. AI’s success may now create an incentive to remove the most valuable knowledge from it.

But Training Data Is Only the First Problem

Suppose we somehow solve it.

Publishers negotiate licences. Writers receive royalties through collecting societies. Websites establish machine-access terms. Governments establish clearer rules. AI companies pay for commercially valuable data.

We still have a second problem.

AI does not merely consume intellectual work.

It can reduce the cost of reproducing its useful consequences.

Consider a train-booking application.

A company may spend years building one. To the customer, it looks simple.

Paris. Milan. Tuesday. Two passengers. Search.

Yet behind that simplicity may be years of experimentation.

Which departure information matters most? How should flexible tickets be displayed? When should seat selection appear? How should a cancellation be explained? Which payment failures are common? How do you prevent fraud? How do you handle missed connections? How much information can you show before the customer abandons the booking?

The company accumulates answers through research, failed designs, customer complaints, engineering work and millions of interactions.

Eventually it develops something that feels obvious.

That is what successful design often looks like. After somebody discovers the answer, it looks obvious.

What exactly is protected?

This is where intellectual property becomes less powerful than many people assume.

Copyright can protect the company’s software code, text, illustrations and other original expression. A trademark can protect its identity. Patents may protect particular inventions if the company has patentable technology and valid claims. Trade secrets can protect information the company actually keeps secret.

But copyright generally does not give the company ownership of the abstract idea:

“Here is a good way to organize a train-booking experience.”

Nor should it necessarily.

Imagine if the first company to invent an online shopping basket owned the concept permanently. Or the first person to use tabs in an application. Or pull-to-refresh. Or a search box.

Innovation also depends on imitation, improvement and recombination.

The balance has always been delicate.

What AI changes is the cost of that recombination.

A competitor can observe the train application. It can examine several competitors. It can read thousands of customer reviews. It can identify recurring complaints. It can study public documentation. It can inspect accessible interfaces. It can ask an AI coding system to propose architectures and generate substantial portions of the implementation. It can combine features from different services.

The AI does not need to steal the original company’s private source code.

It can help somebody recreate a great deal of the product’s visible functional value independently.

That distinction is crucial.

The disruption may occur without traditional infringement.

The Imitation Gap

For an innovator, time matters.

Imagine that developing something genuinely new costs $10 million. The company spends that money because it believes its innovation will give it an advantage. Perhaps competitors need four years to catch up. During those four years, the innovator earns enough to recover the original investment.

Now imagine that AI reduces the time competitors need to reach a comparable product from four years to eighteen months.

The original company still owns its code. Nobody necessarily violated its copyright. Its patent rights, if any, remain. Its trademark remains. Its product still exists.

But its economics have changed.

Its imitation gap has become shorter.

And that matters because intellectual investment is made today against expectations about tomorrow’s returns.

This gives us a more useful way to think about AI and innovation.

AI changes two costs simultaneously.

It reduces the cost of creation.
And it reduces the cost of imitation.

The first effect is widely celebrated.

A small company can build software that previously required a large engineering team. A scientist can analyze information more quickly. A designer can explore more alternatives. A writer can research faster. An entrepreneur can prototype ideas cheaply.

All of that can increase innovation.

But the second effect receives much less attention.

The same tools available to the innovator are available to the follower.

So the real economic question becomes:

Which cost falls faster?

If AI makes invention much cheaper while imitation becomes only slightly cheaper, we could see an explosion of innovation.

If AI makes copying the useful consequences of innovation cheaper even faster, the incentive to bear the cost and risk of going first may weaken in some industries.

We do not yet know the answer.

But it is a more interesting question than whether AI is simply “good” or “bad” for creativity.

Patents Do Not Solve the Whole Problem

Patents are particularly interesting because they already assume that competitors will read the invention.

A patent is not a secret. That is the bargain. The inventor discloses the invention publicly and receives a limited period during which others cannot practice the protected claims without permission. Afterward, society gets the invention.

AI does not automatically destroy that model.

If an AI reads a patent and proposes something that still falls within the patent’s claims, the patent may still matter.

But AI could make searching the patent landscape easier. It could make prior-art research easier. And it could make designing alternative implementations around certain patents less expensive.

That does not make patents obsolete. It changes the economics surrounding them.

For some businesses, this could make trade secrets more attractive when secrecy is practical. For others, patents remain essential precisely because software or machines become easier to reverse-engineer.

Again, the direction is not simply “IP disappears.”

It may be the opposite.

As imitation becomes cheaper, control over the things that cannot legally or technically be imitated may become more valuable.

Then Comes the Open-Source Problem

Suppose society concludes that high-quality training data deserves compensation.

That sounds reasonable.

But now ask who can afford it.

A handful of giant technology companies can sign large licensing agreements with newspapers, publishers and databases.

A university laboratory may not. An independent researcher may not. An open-source project certainly cannot negotiate separately with millions of creators.

This creates another paradox.

Rules designed to compensate creators could unintentionally strengthen the largest AI companies because those companies have the money to purchase access.

The smaller competitors might increasingly depend on public-domain information, openly licensed material and voluntarily shared datasets.

That is not necessarily disastrous. There is an enormous amount of such information.

But if the most current, expert and commercially useful data becomes increasingly private, access itself becomes a competitive moat.

The future AI race may therefore depend not only on who has the best model architecture or the most computing power.

It may also depend on:

Who has permission to know what?

That is an unexpected destination for a technology frequently described as democratizing information.

Can’t AI Simply Train on AI?

There is an apparent escape from this problem.

If obtaining human-created information becomes expensive, machines can create training data themselves.

And they already do.

Synthetic data is extremely useful. Models can generate examples, solve problems, critique outputs and create simulated environments.

But synthetic data does not make the underlying information problem disappear.

Research has shown that repeatedly training on poorly controlled model-generated data can progressively degrade models and narrow the distributions they learn from.

This is sometimes described as model collapse.

That does not mean synthetic data is useless. Far from it. Carefully generated and curated synthetic data can be highly valuable.

But a system ultimately needs some connection to reality.

New scientific results have to come from somewhere. New political events happen in the world. New products are invented. New diseases appear. Languages change. Cultures change. People discover new things.

An AI cannot indefinitely obtain genuinely new information merely by asking itself what humanity discovered yesterday.

At some point, somebody still has to look at the world.

This Is the AI Incentive Paradox

AI needs a civilization that continues producing fresh, high-quality intellectual material.

At the same time, AI can reduce the scarcity from which some intellectual activity previously derived its economic value.

That gives us the AI incentive paradox:

The better AI becomes at extracting economic value from human knowledge, the more important it becomes to preserve the institutions that finance the creation of new human knowledge.

People will not stop thinking simply because AI exists.

People painted before copyright. They wrote poetry before royalties. They discovered mathematics before patents. Curiosity is older than markets.

But modern knowledge production is not powered by curiosity alone.

Drug research requires laboratories. Journalism requires reporters. Semiconductor development requires enormous capital. Scientific institutions require funding. Professional authors need time. Software companies pay engineers. Design requires experimentation.

Economic incentives matter because organized knowledge production costs money.

The relevant question is therefore not:

Will humans stop creating?
They will not.

It is:

Who finances expensive creation when the economic lifetime of its advantage becomes shorter?

Fine. Then Compete Through Marketing.

Suppose software becomes increasingly easy to reproduce. Perhaps code itself stops being the main competitive advantage.

A conventional business strategist might say:

Fine. Build a brand. Market better. Sell better. Own the customer relationship. The product may be commoditized, but distribution remains scarce.

That is a perfectly reasonable response.

Except AI is beginning to transform distribution too.

For most of modern commercial history, marketing was an exercise in persuasion.

A company had a product. It wanted the customer to choose that product. So it advertised. It hired salespeople. It bought shelf space. It optimized websites for Google. It sponsored influencers. It built a brand. It tried to occupy a larger portion of the customer’s attention than its competitors.

The customer was surrounded by competing messages and eventually made a decision.

The structure looked roughly like this:

Company → persuasion → consumer → choice

AI introduces another participant.

The Consumer No Longer Needs to Examine the Market

Imagine that I want a laptop.

I could open Google. Read reviews. Visit manufacturer websites. Compare specifications. Watch YouTube videos. Go through Reddit discussions. Look at prices. Try to decide which reviewers I trust.

Or I can ask:

I travel frequently. I don’t play games. I care about battery life and keyboard quality. I want the machine to last at least five years. My budget is $1,200. What should I buy?

The AI can process the market for me.

This is enormously useful. That should not be minimized.

Modern consumers face an absurd abundance of choices.

Hotels. Insurance. Software. Restaurants. Flights. Cameras. Financial products. Courses. Books. Health information. Products of every kind.

For much of history, information was scarce.

Now attention is scarce.

AI offers to solve that problem by filtering the information before we see it.

But filtering creates power.

Suppose there are 5,000 plausible laptops. The AI returns three.

The consumer still chooses.

But 4,997 products have already disappeared.

This is where marketing begins changing its nature.

From Persuasion to Pre-Selection

Traditional marketing wanted to persuade me after I encountered the product.

AI can affect whether I encounter the product at all.

That is a different position in the decision process.

I would describe the progression like this:

Persuasion → recommendation → pre-selection → algorithmic imposition

We are largely somewhere in the middle today.

“Imposition” is intentionally the provocative end of the spectrum. Current AI systems do not generally force consumers to select a particular product. The consumer retains agency.

But consider what happens as we delegate more of the process.

First: “Show me some laptops.”
Then: “Compare them.”
Then: “Tell me which one suits me.”
Eventually: “Buy whichever one is best for me.”

At each stage, the machine handles more of the decision.

And the user may welcome this.

That is what makes the development interesting.

This is not necessarily power taken from consumers against their will. It can be power voluntarily delegated by consumers because the machine can process more information than they can.

The better AI becomes, the more rational that delegation may appear.

The New Shelf Is the Answer

Retailers have always understood the value of placement.

A supermarket product at eye level has an advantage. A restaurant on the busiest street has an advantage. In the Google era, appearing on the first page had an advantage. In the social-media era, appearing in the recommendation feed had an advantage.

AI creates another kind of shelf.

The answer itself.

If the AI examines a thousand products and recommends three, those three occupy extraordinarily valuable economic territory.

Suddenly the important question for a company is no longer only:

How do we persuade the consumer?

It also becomes:

How do we become one of the options the machine chooses to show?

This does not kill marketing. It moves marketing upstream.

Instead of competing exclusively for human attention, companies begin competing for machine consideration.

We can already see early versions of this in the frantic discussion around optimizing information so AI systems understand and surface a company, product or website.

Some call it GEO. Some call it AEO. Some call it LLM optimization.

The terminology matters less than the structural change.

The target of persuasion is moving.

But Can the Machine Be Persuaded?

This leads to a strange possibility.

We may imagine AI recommendation as the end of marketing.

Instead, it may produce an entirely new marketing industry aimed at machines.

Companies will structure product information for agents. They will publish machine-readable specifications. They will seek favorable mentions in sources AI systems consider authoritative. They will optimize pricing feeds. They will try to appear in comparison databases. They will build APIs intended for purchasing agents. They will attempt to understand which signals lead a model to include one product and exclude another.

The advertising industry does not disappear. It adapts.

The battle simply moves away from the billboard and perhaps eventually even away from the human screen.

This could create something stranger than conventional advertising.

Two machines negotiating over us.

A merchant’s AI attempts to make an offer attractive. Our personal AI evaluates it. The seller optimizes for the buyer’s agent. The buyer’s agent negotiates against the seller.

Commerce becomes increasingly machine-to-machine.

The human sits at the end of the transaction and approves—or eventually perhaps does not even need to approve every routine purchase.

That is no longer traditional marketing.

It is an algorithmic market.

Now the Intermediary Becomes Extraordinarily Powerful

Once an AI stands between the market and the consumer, another question becomes unavoidable.

Whose interests does the AI represent?

Imagine that your personal assistant says:

“This is the hotel I recommend.”

Why?

Because it genuinely determined that this hotel best matches your preferences? Because the hotel has the best data available to the system? Because other customers rated it highly? Because it paid to appear? Because the platform running your assistant receives a commission if you book it? Because it has a commercial partnership with the booking platform?

These possibilities already exist in conventional commerce.

Search engines have advertisements. Booking platforms receive commissions. Retailers promote sponsored products. There is nothing unprecedented about commercial influence.

What changes is the interface.

With a search page, we see the marketplace. We see ten links. We often see labels marking advertisements. We know somebody is competing for our attention.

A conversational agent feels different.

We ask it a question. It answers. The relationship resembles advice.

And advice produces a different kind of trust.

That makes transparency much more important.

When the line between adviser, search engine, marketplace and advertising platform begins to blur, the economic power of the intermediary grows.

Persuasion May Become Less Visible, Not Disappear

This is why I would not say that AI ends advertising.

It may actually make persuasion less visible.

Traditional advertising announces itself.

A television commercial interrupts the program. A banner sits beside the article. A salesperson calls. We recognize the attempt.

An AI recommendation can feel like the result of reasoning.

“This is probably best for you.”

That is psychologically different.

If commercial incentives eventually shape those answers, we may not have escaped marketing at all.

We may have moved from visible persuasion to embedded persuasion.

And perhaps that is where your phrase “algorithmic imposition” becomes useful.

Not because the algorithm literally forces us to obey.

But because increasingly powerful filtering can determine what enters the field of possible choices before we exercise our freedom to choose.

The power lies not necessarily in selecting the winner.
It may lie in selecting the candidates.

So What Remains Valuable?

Return to the train application.

Suppose AI can help somebody reproduce a comparable interface quickly.

What can it not reproduce so easily?

The railway contracts. The live inventory relationships. Regulatory approvals. Payment infrastructure. Historical fraud data. Customer support. Brand recognition. Insurance. Operational competence. The ability to compensate a traveler when something goes wrong. The legal entity standing behind the transaction.

The AI can reproduce the digital shell more easily than it can reproduce the institutional machinery beneath it.

That suggests a broader shift.

As information and code become cheaper, economic value may migrate toward things that remain scarce.

Physical infrastructure. Energy. Chips. Exclusive contracts. Proprietary data. Distribution rights. Licences. Reputation. Brands. Communities. Human relationships. Regulatory permission. Trust. Liability. The right to connect to something real.

Perhaps the strongest companies of the AI era will not be those whose software cannot be copied.

They may be those whose position cannot be copied.

And That Creates Another Closed Circle

If this happens, companies react rationally.

They hide more valuable data. They expose less proprietary information. They protect APIs. They prefer contracts to openness. They move knowledge behind authentication. They build closed ecosystems. They rely more heavily on assets requiring permission.

In other words, AI’s ability to extract value from open information may encourage valuable information to become less open.

The open internet does not disappear.

But the premium layer may become increasingly gated.

That would be a strange reversal.

The internet spent thirty years reducing the cost of access.
AI might make access so economically powerful that owners begin rebuilding the gates.

The Larger Issue Is Not Copyright

This is why I think reducing the AI debate to copyright misses the more important transformation.

Copyright matters. Patents matter. Licensing matters. Courts will matter. New legislation may matter.

But the deeper economic issue exists even if every AI company follows every rule perfectly.

Imagine that all training data is licensed. Imagine every copyrighted work is handled legally. Imagine AI-generated code never copies protected source code. Imagine every advertisement is clearly disclosed.

We still face the central problem.

• AI can make it cheaper to learn from what others discovered.

• It can make it cheaper to reproduce many of the functional advantages they built.

• And it can make it cheaper for consumers to delegate comparison and selection.

Those are three different frictions falling at once.

Historically, each friction protected some part of economic value.

Knowledge was difficult to acquire. Products were difficult to reproduce. Customers were difficult to reach.

AI attacks all three.

That does not automatically destroy innovation. It may unleash enormous innovation.

But it changes where the value goes.

One Machine, Three Positions

This is what makes AI different from many previous information technologies.

The printing press reproduced information. Search engines organized information. Social networks distributed information. E-commerce platforms organized transactions.

AI can potentially sit across all of these functions.

It can learn from producers. It can help new producers create alternatives. And it can stand between producers and consumers when the consumer asks what to buy.

The same technological layer can participate in:

Production, imitation and distribution.

That concentration of functions deserves more attention than it currently receives.

Because once the intermediary sits in all three places, bargaining power may shift dramatically.

Who Pays for the Next Idea?

None of this leads to a simple conclusion that AI is bad for creativity.

That would be too easy.

AI may lower the cost of creating a company. It may allow one researcher to accomplish what previously required a team. It may give an independent filmmaker tools that once belonged only to studios. It may let small businesses compete with corporations. It may dramatically accelerate science. It may create entirely new categories of intellectual work.

The productive possibilities are enormous.

But every technological revolution changes incentives as well as capabilities.

And capabilities receive much more attention than incentives.

We ask:

What can AI write? What can it code? What can it design? What jobs can it automate? What companies can it disrupt?

Perhaps we should add another question.

What economic system continues producing the information that makes those capabilities possible?

For thirty years, the great achievement of the internet was making human knowledge easier to access.

AI’s great achievement may be making that knowledge easier to use.

But access is not creation. Use is not discovery.

Someone still has to conduct the experiment. Someone has to report from the war zone. Someone has to spend ten years developing a molecule that fails. Someone has to discover the algorithm before it becomes obvious. Someone has to build the first train application before an AI can analyze a hundred train applications. Someone has to write the book worth summarizing. Someone has to create the next thing the model has never seen.

Artificial intelligence can make extraordinary use of what civilization already knows.

But eventually it encounters the edge of that knowledge.

Beyond that edge, somebody still has to go first.

And going first is expensive.

That may be the economic question we should be thinking about now.

Not whether AI can absorb everything humanity has already produced. It increasingly can.

The question is what happens afterward.

Who pays humanity to produce what the machines need to learn next?

Notes & Further Reading

This essay is an argument about economic incentives rather than a legal prediction. A few sources are useful for readers who want to examine the evidence behind it more closely.

On intellectual property and the economics of invention:
The World Intellectual Property Organization’s explanations of patents provide a useful introduction to the basic bargain: temporary exclusivity in exchange for disclosure, allowing inventors an opportunity to recover investment while eventually adding the invention to public knowledge.

On what copyright does — and does not — protect:
The U.S. Copyright Office emphasizes an important distinction used in the essay: copyright protects original expression, not ideas, systems, procedures or methods themselves.

On AI and software productivity:
Recent field experiments with professional developers show that AI coding tools can increase productivity, although the size of the effect varies considerably depending on the task, developer and development environment. More recent research also suggests that dramatic increases in code production do not translate proportionally into finished software releases.

On AI training and licensing:
The legal position remains unsettled. Courts have reached different conclusions depending on how copyrighted material was obtained and used, while publishers and AI companies have simultaneously begun creating a commercial licensing market for training and content access.

On the changing economics of web traffic:
Pew Research Center found that users exposed to Google’s AI summaries were substantially less likely to click traditional search results. This is one reason publishers are reconsidering the old bargain between allowing automated crawling and receiving human traffic in return.

On the permissioned web:
Cloudflare has introduced tools allowing publishers to distinguish AI crawlers from conventional search crawlers and has experimented with allowing websites to charge automated systems for access.

On AI as an economic gatekeeper:
OECD research on AI markets has highlighted the potential importance of control over data, distribution, defaults and vertically integrated AI services as these systems become intermediaries between businesses and consumers.

On synthetic data and model collapse:
Research published in Nature showed that indiscriminate recursive training on machine-generated material can cause models to progressively lose information about the original data distribution. This does not mean synthetic data is inherently harmful; carefully constructed synthetic data can remain extremely useful.

These sources do not prove the essay’s final outcome. That is precisely the point. They establish that the underlying mechanisms — cheaper production, emerging licensing markets, reduced referral traffic, machine filtering and restricted data access — are already real. What they ultimately do to incentives for human creation remains an open economic question.

Karim Spring