AI and the post-IP internet: what’s worth something when producing another version costs less

Rights are still in force. The question for media companies, creators, and startups is which part of their business depends on controlling a single piece of content, and which part depends on something a new version can’t replace.

Rights are still in force. The question for media companies, creators, and startups is which part of their business depends on controlling a single piece of content, and which part depends on something a new version can’t replace.

On July 1, 2025, Cloudflare unveiled an experiment to charge AI crawlers for accessing content. The mechanism uses web infrastructure and payment-required responses to put a price on an automated query. It’s a scene considerably less futuristic than an internet where nobody owns anything: someone knocks on the door, and the owner tries to charge admission.

Generative AI opened up a discussion about a “post-IP” internet. Here we use IP to mean intellectual property, not internet protocol. And we use “post-IP” to describe an economic hypothesis: an environment where holding rights to a piece of content still matters, but turns out to be insufficient to ensure its producer captures the value it generates.

Our hypothesis is that, in categories where outputs are easily substituted, part of the value can shift toward access to information, the relationship with an audience, and the capacity to take responsibility for a result. That relationship includes distribution capacity, which requires more than simply having content available on a platform. If buyers keep paying for interchangeable pieces without demanding those attributes, the hypothesis loses force.

Intellectual property didn’t disappear with the prompt

The “post-IP” label bundles together problems that are worth separating. One is whether a work generated with AI can receive protection. Another is under what conditions third-party works are used to train a model. A third is who controls access to published material. Their answers aren’t interchangeable.

In January 2025, the U.S. Copyright Office held that protection of an output depends on sufficient expressive human contribution. It also clarified that using AI as assistance doesn’t exclude protection, while providing prompts alone isn’t enough. This is a position on U.S. copyright law and authorship, not a global rule governing every use of AI.

Scope matters. That conclusion doesn’t grant blanket permission to copy third-party works, nor does it resolve the training debate on its own. Nor does it make a trademark, a patent, and an editorial piece into the same kind of asset. Each right protects something different and operates under specific conditions.

That’s why “AI killed intellectual property” is a poor description of what can actually be observed. There are new production tools and disputes over specific uses. The economic question can be relevant without announcing a legal abolition that never happened.

A useful substitute can compete without being an identical copy

The internet had already lowered the cost of reproducing and distributing files. AI adds the possibility of generating a different response that fulfills part of the function a person was looking for in the original.

If someone needs an introductory explanation of a concept, they may accept a summary instead of reading a specific piece. If they’re looking for original research, the identity of the source and access to evidence can matter much more. Not all information demand behaves the same way.

This distinction allows us to talk about economic substitution without settling a legal question in advance. Two products competing for the same time or budget doesn’t prove that one infringes the other’s rights. And a use being lawful doesn’t guarantee it preserves the business model of whoever produced the information.

For a startup, the risk appears when what it sells looks too much like a reproducible output. A template library can face different pressure than a service that implements and maintains a process within an organization. Both produce files. The second also takes on work that continues after delivery. That distinction matters when deciding what to automate in content operations: speeding up production leaves open the decisions about criteria, validation, and accountability.

That doesn’t make the service invulnerable. If the implementation becomes standardized and another provider can take it on at lower cost, the advantage is back up for debate. Adaptability needs to be observed in operation, not declared in a pitch deck.

Licensing deals are evidence against the most extreme version of the thesis

In May 2024, OpenAI announced a multiyear agreement with News Corp for access to current and archival content from its publications, to be shown in responses. The announcement doesn’t reveal the full economics of the deal, but it does show that an AI platform considered it worthwhile to negotiate access with a content rights holder.

The example contradicts a maximalist reading of “post-IP”: if rights and access had lost all relevance, deals like this would be hard to explain. A more reasonable interpretation is that certain catalogs retain bargaining power, owing to their content and the terms they’re able to offer.

The simplest alternative explanation is that we’re looking at a new stage of licensing, with disputes over prices and uses, within familiar structures. That explanation fits the available evidence better than the idea of intellectual property simply vanishing.

Our hypothesis adds a further question: what happens to those who produce substitutable content and lack the volume, exclusivity, or commercial relationship needed to negotiate. Deals struck by major groups don’t prove a viable market exists for every publisher. Nor do they prove that smaller players are necessarily shut out.

Cloudflare is trying to intervene precisely on the question of access. As of this writing, its Pay Per Crawl documentation describes it as a closed beta. The existence of the mechanism proves a technical possibility. Its adoption, and crawlers’ willingness to pay, will determine how much commercial value it can produce.

Four assets worth distinguishing before changing your business

We propose evaluating an offering by asking which part of its value survives once someone else can produce an acceptable version of its output. It’s a diagnostic tool, not a prediction of future prices for any given industry.

AssetWhat can sustain its valueWhat weakens it
Piece or catalogIdentifiable expression, clear rights, and demand for that specific materialThe buyer accepts any sufficiently similar substitute
AccessOriginal, up-to-date information, or information hard to obtain with permissionThe same data is available from multiple equivalent sources
RelationshipA returning audience, earned trust, and direct distribution capacityThe connection depends entirely on an intermediary platform
ExecutionIntegration, maintenance, and accountability for a resultThe service is limited to delivering a generic file

A media outlet can hold all four assets. It publishes protectable pieces, obtains its own information, builds a relationship with readers, and offers services or tools. A software company can too: its code is one part of the business, alongside data, integrations, distribution, and support.

The table helps avoid a common reaction: responding to abundance by producing even more. If the problem is that your output is easily substituted, multiplying it can raise costs without changing the reason someone chooses you.

The answer depends on what you sell. For a publication, it might be research that others need to cite. For a B2B tool, authorized access to a workflow and the ability to operate it. For a consultancy, judgment applied to a context and support through implementation. No option becomes defensible simply by describing it in fancier words.

A brand matters when it represents something verifiable

The abundance of content can make a recognizable source more useful. But a brand only captures that usefulness if the experience lives up to the expectations it creates. A distinctive voice earns attention; a source whose claims can be verified gives people a reason to come back.

For a startup using the same models as its competitors, describing itself as “the most advanced AI platform” leaves the purchase question wide open. Which task does it understand better? What data can it legitimately use? What result can it demonstrate, and in what context does it fail?

Those answers need to travel together, from the product to the homepage to the sales conversation. At El Programa we work on that connection between vision, competitive context, and narrative. In a category where capabilities look alike, the strategic work consists in finding a reason to choose that the business can actually sustain.

It also helps to make content verifiable by machines. Dated data, clear definitions, and identified sources help a claim remain interpretable outside its original page. That structure improves the usefulness of content; it doesn’t guarantee a system will cite it, or that a citation will produce revenue.

What to test over the next quarter

A useful test starts with the offering, not with whatever AI model is trendy. Pick a deliverable you currently sell and ask what would happen if a buyer could get an acceptable version tomorrow for a fraction of the cost. Identify what reason would remain for them to pay you.

If the answer is “quality,” spell out how it’s evaluated. If it’s “trust,” figure out what experience builds it and how it transfers when the person delivering the work changes. If it’s “proprietary data,” review its origin, how current it is, and what permissions apply. That review keeps you from mistaking an advantage for an asset you don’t actually control.

Then test an offering that makes the remaining value explicit. For example, replace a generic library with a product updated for a specific task, or pair a recommendation with implementation and follow-up. Measure whether comparable buyers recognize that difference and are willing to pay for it. Don’t assume that adding service automatically improves the economics: it also adds costs.

The displacement hypothesis would gain strength if customers reduce their willingness to pay for generic outputs while maintaining or increasing payment for access, relationship, or execution. It would weaken if pieces retain demand and price without those complements. Both sides need to be watched.

Frequently asked questions

Does a “post-IP internet” mean all content will be free?

No. It’s a hypothesis about value capture, not a legal description. Licensing deals and paid-access mechanisms show that attempts to charge for content and permissions also exist.

Should you block all AI crawlers?

The decision depends on which bots are accessing your content, for what purpose, and what value the business gets from that exposure. Separate out crawling for search, answer retrieval, and training wherever the tools allow you to distinguish between them. Our guide on blocking AI crawlers goes into the difference between expressing a preference in robots.txt and technically restricting access. Blocking can limit exposure; allowing access doesn’t guarantee compensation.

What changes for a small team?

It’s worth identifying what you produce that isn’t just a reformulation of available information, and how that connects to revenue. A direct audience or original information can be valuable, but building it is costly. The priority is to test a source of value the team can actually sustain.

The proof of an advantage is that someone keeps choosing you even once they have access to a sufficiently similar alternative.