AI agents and micropayments are being presented as a possible answer to a growing copyright dispute, after newly unsealed court filings revealed an exchange in which OpenAI president Greg Brockman appeared to welcome a way around a New York Times paywall.
An OpenAI researcher told Mr Brockman about “a hack to get around nytimes paywall”, to which he replied: “ah nice”. The exchange appears in the newspaper’s copyright case against OpenAI and Microsoft.
OpenAI and Microsoft argue that their use of material is protected by fair use. Publishers dispute that position, leaving the courts to decide the legal arguments. But the wider question is how artificial intelligence systems should pay for the human knowledge they increasingly consume.
One proposed answer is to create a payment system designed for machines rather than people. Under this model, an agent could buy a single newspaper paragraph, research data point, photograph or court opinion for a small sum, using a payment method selected and authorised within limits set by its owner.
Traditional subscriptions may be unsuitable for such uses. An agent might need access to one piece of material only once, and for a matter of seconds, while unrestricted free access would provide no sustainable source of income for publishers and other rights holders.
Micropayments have existed for decades but have struggled to gain widespread practical use. The difficulty was not only the cost of processing small transactions. People are also reluctant to make hundreds of minor purchasing decisions each day, while the administration involved has often outweighed the revenue.
Automated systems would not face the same obstacle. They could make thousands of small purchasing decisions according to instructions, spending caps and conditions established by individuals or businesses.
Payments within defined limits
The emerging model of agentic payments would allow software to identify a transaction, confirm that it has authority to act, choose a suitable payment method and complete the purchase within an approved mandate.
That mandate could be tightly restricted. An individual might instruct an agent to buy an article for less than 10 cents, purchase an item below a specified price, or renew a service only if its cost remained within an agreed range.
Potential payment routes include cards, bank transfers, open-banking payments, prepaid balances, stablecoins and machine-focused protocols. The relevant technology could vary according to the transaction, with programmable payment systems potentially helping to handle some small cross-border payments.
The same infrastructure could support charges based on data use, tokens or API calls, as well as the automatic division of revenue between several rights holders.
Agents may also transact directly with one another. An algorithm representing a buyer could assess price, timing and terms against a system acting for a seller, completing the transaction within parameters set by the people or organisations behind them.
Such arrangements would require more than a mechanism for moving money. A counterparty would need to establish which agent it was dealing with, who had authorised it, what it was allowed to do and how much it could spend.
Identity checks, verifiable intent, permission controls, fraud prevention, audit trails and ways to resolve disputes would therefore form part of what has been described as a new trust architecture for delegated machine activity.
Copyright and machine-readable rights
Those safeguards could also be applied to copyrighted material. Content could carry machine-readable details identifying its creator and rights holder, setting out permitted uses, attribution requirements and a price.
An agent seeking access would check those terms against its mandate. Where use was allowed, it could pay automatically and preserve information about the material’s origin. If attribution was required, it would remain attached to the content. Payment would not make prohibited use lawful.
Smart contracts could also distribute revenue automatically among the relevant rights holders. The proposed system would not settle disputes about material already used to train artificial intelligence models, and would not replace copyright law. It would instead provide an operational layer for machine transactions.
Several companies are developing components of that infrastructure. Mastercard’s Agent Pay for Machines is designed for continuing machine-to-machine transactions, including payments worth fractions of a cent. Visa’s Trusted Agent Protocol is focused on verifying an agent’s identity and authority, while Stripe and Tempo’s Machine Payments Protocol is intended to support programmatic transactions.
The debate extends beyond the income received by publishers. If artificial intelligence systems increasingly provide answers without sending users to original sources, while offering little payment or recognition to the people who produced the material, the incentive to create costly original work could weaken.
At the same time, the internet is accumulating synthetic material, with future systems increasingly learning from content produced by earlier models.
Locking knowledge behind barriers is unlikely to provide a complete answer, but treating human work as free raw material for artificial intelligence would not resolve the problem either.
Supporters of agentic payments argue that copyright can establish the rights, provenance can preserve authorship and credit, and automated micropayments can make compensation sufficiently precise to operate at machine scale.
As artificial intelligence agents become larger consumers of human knowledge, the proposal is that they should also become paying customers: recognising the creators of the material they use, complying with the rights attached to it and paying for its value.
