AI Sells Answers by the Token—Creators Are Testing a Meter of Their Own

AI access is still sold as a metered service, while the right to train commercial systems on copyrighted work remains legally unsettled. What has changed since May 2026 is the appearance of a second meter: an experiment designed to compensate participating publishers when their material contributes to an AI search result.

That development makes the “knowledge arbitrage” argument more precise, not obsolete. Model providers charge for computation and access, but creators may now gain mechanisms to price the use of their work as well. The central conflict is therefore moving from a simple accusation of extraction toward a contest over measurement, permission and who receives each payment.

The AI utility has two possible meters

The familiar meter sits at the output end of the system. Developers and businesses pay for model access according to usage, although the exact unit and rate vary by product. OpenAI’s current token guidance explains that API billing distinguishes input, output and cached tokens, turning both the material submitted to a model and the response it produces into measurable consumption.

This is a utility-like commercial structure, but it does not mean that customers are merely buying back facts they already owned. The fee also pays for model development, computing infrastructure, inference, software, safety systems and availability. None of those costs resolves the separate question of whether copyrighted training material was acquired and used lawfully.

The newer meter is emerging on the content side. On July 1, 2026, Cloudflare described its Pay Per Use experiments with Ceramic.ai and You.com: participating publishers may be compensated when their content appears in a search result or when an agent requests premium material. Cloudflare explicitly calls the initiative an experiment, so it should not be mistaken for a universal royalty system or a broadly available creator income stream.

The distinction matters. Paying for every crawl rewards repeated retrieval, even if a page never contributes to an answer. Paying for use attempts to connect compensation to a visible result, but it introduces harder questions: what counts as contribution, how is it attributed, and how should revenue be divided when an answer draws on many works?

“They scraped the internet” hides several different acts

The strongest criticism of commercial AI is not that knowledge itself was privatized. Facts, ideas, public-domain works, licensed archives and protected expression do not share one legal status. Nor does crawling a publicly reachable page automatically establish permission to reproduce copyrighted expression for every later purpose.

At least three stages need separate scrutiny: acquiring a copy, using copies during training, and producing or retrieving material for a user. A company may have different legal arguments, licenses and technical controls at each stage. Treating the entire pipeline as one act of “harvesting” is rhetorically efficient but too imprecise to show where consent was absent or which market was harmed.

The official position in the United States remains fact-specific rather than categorical. The Copyright Office’s Part 3 report, released in pre-publication form on May 9, 2025, concludes that some training uses will qualify as fair use and others will not. It places noncommercial research without reproducing protected passages toward one end of the spectrum, while treating the use of expressive works from pirate sources to create competing commercial content, where licensing is reasonably available, as unlikely to be fair.

That is narrower than saying all training is theft, but it also rejects the idea that calling a system transformative settles the issue. The source of the material, the purpose of the model, output controls and effects on existing or potential markets can all change the analysis. Courts, not model vendors or critics, ultimately decide particular disputes.

Licensing changes the argument without solving it

A growing licensing market demonstrates that permission is feasible in some categories. It does not prove that every work needed for a general-purpose model can be identified and licensed at practical cost. Large publishers, image libraries and music catalogues can negotiate because their rights are comparatively concentrated; an independent creator may own valuable work without possessing comparable bargaining power.

Collective or automated licensing could reduce that imbalance, but only if attribution survives the full workflow. A payment system needs reliable information about who controls a work, what uses were authorized, whether the content actually influenced a result and what accounting rule produced the creator’s share. A crawl log alone cannot answer all of those questions.

For creators, the practical asset is therefore not “knowledge” in the abstract. It is a documented bundle of rights, provenance and distribution control. Clear ownership records, machine-readable access preferences, direct audience relationships and contracts that distinguish indexing, retrieval, training and answer generation make a work easier to license and harder to treat as ownerless material.

The real arbitrage is an accounting gap

Commercial AI combines many inputs, but its most legible transaction currently occurs at the customer end. A provider can count tokens or subscriptions because it controls the interface. The contribution of any individual article, illustration or recording is less visible, especially after material has been combined with millions of other works.

That asymmetry creates the arbitrage: downstream usage has a mature meter, while upstream creative contribution often lacks a shared unit of account. It does not follow that every token payment corresponds to a royalty owed for training. It does mean that the industry can measure consumption far more readily than it can attribute the value supplied by particular creators.

Pay-per-use experiments begin to close that gap at retrieval time, where a system can identify a page used in an answer. They do much less for historical training, where causal attribution to one work may be technically and economically contested. Training licenses, retrieval payments and revenue sharing are consequently different mechanisms, not interchangeable labels for creator compensation.

What remains unresolved

The AI utility model is real as a pricing structure, but the claim that it simply encloses all human knowledge goes too far. Models provide a computational service built from infrastructure, software and data; the unresolved issue is whether protected contributions inside that data were obtained under a valid license, a legally defensible exception or neither.

The 2026 shift is that compensation is starting to be tested closer to the moment an answer creates value. Whether that becomes meaningful for independent creators will depend on participation, attribution quality, bargaining terms and transparent reporting. Until those pieces are proven at scale, creators face two meters with very different maturity: a precise bill for consuming AI and an experimental ledger for supplying the work that can make its answers useful.

Also read:

• How Chatbots are Redefining Mobile Apps and their Usage
• Patreon vs YouTube Memberships: Where Should Creators Put Their Paid Community?
• What are The Digital Wallets and Their Types?