Newly unsealed court filings in The New York Times' copyright case against OpenAI and Microsoft have drawn attention to an exchange in which an OpenAI researcher described a method for bypassing the newspaper's paywall and company president Greg Brockman responded approvingly. The episode is striking on its own, but it also points to a larger problem in the emerging AI economy: machines have become some of the largest consumers of human knowledge, and there is still no working system for them to pay for it.
OpenAI and Microsoft argue that their use of copyrighted material is protected by fair use. Publishers disagree. The courts will ultimately decide the legal question, but the outcome of that litigation will not by itself produce a functioning economic model. What is needed, according to proponents of a new approach, is an architecture that preserves three things the internet has struggled to hold together at scale: rights, credit, and compensation.
One of the most practical pieces of that architecture may be an old idea that has finally found its use case. Micropayments, powered by agentic payments, could allow AI systems to pay small sums for individual pieces of content. An AI agent may need only one paragraph from a newspaper, a single data point from a research service, one photograph, or one court opinion. It may need that material once, for a few seconds. A monthly subscription is the wrong instrument for that kind of access, and free access is not a sustainable answer either.
Micropayments have existed for decades but saw little practical use. Transaction costs were part of the problem, but the bigger obstacle was human. People do not want to make hundreds of tiny purchasing decisions a day, and the operational overhead rarely justified the revenue. AI agents do not share that limitation. They can make thousands of such decisions programmatically, within parameters set by the people or businesses behind them.
That does not mean giving agents unlimited access to a user's money. More likely, delegation will be narrow and specific: defined mandates, spending limits, approved categories, and explicit conditions. An agent might be authorized to buy an article if it costs less than ten cents, or to purchase computing power according to actual usage, or to renew a service only if the price remains within a specified range. Agentic payments can operate across several financial rails, including cards, account-to-account payments supported by open banking, bank transfers, stablecoins, prepaid balances, and new machine-native protocols. For some cross-border micropayments, stablecoins and other programmable rails may make very small transactions faster and cheaper than traditional methods.
The important development is not any single rail but the creation of an ecosystem in which software can identify a transaction, verify authority, select an appropriate payment method, and execute it within defined parameters. That could make previously impractical economic models commercially viable: paying fractions of a cent for data, a few cents for an article, charging by token or API call, or automatically splitting revenue among multiple rights holders.
Moving money is only part of the challenge. Counterparties need to know which agent they are dealing with, who authorized it, the scope of its mandate, the user's intent, and its spending limits. Concepts such as «know your agent,» verifiable intent, permissioning, fraud prevention, audit trails, and dispute mechanisms are becoming part of a new trust architecture for delegated machine action. Those same capabilities fit the copyright problem closely. Copyright is not merely a payment claim. It recognizes that creative work has an author, and that principle should not disappear because the reader is now a machine.
In one vision of how this could work, content would carry machine-readable information identifying its creator, rights holder, permitted uses, attribution requirements, and price. An agent would request access and check those terms against its mandate. If the use is permitted, it would pay automatically and carry the provenance forward. If attribution is required, it would stay attached. If the use is prohibited, payment would not make it permissible. Smart contracts could distribute proceeds automatically among relevant rights holders.
The building blocks are already appearing. Mastercard's Agent Pay for Machines is designed for continuous machine-to-machine transactions, including microtransactions worth fractions of a cent. Visa's Trusted Agent Protocol focuses on verifying agent identity and authorization. Stripe and Tempo's Machine Payments Protocol provides an open standard for programmatic, machine-native transactions.
None of this would resolve disputes over historical training data, nor should it replace copyright law. But it would give copyright something it has never had at internet scale: an operational layer built for machines. The stakes go beyond publishers' revenues. If AI systems increasingly substitute for visits to original sources while returning little to the people who produced that material, the incentive to create and maintain reliable knowledge may erode. A payments layer designed for agents could help ensure that the machines consuming human knowledge also help pay for it.