India’s AI copyright proposal makes the royalty formula the central bargain

A single licence for training AI could solve a real coordination problem. A model developer may draw on works owned by thousands of publishers, writers and other creators. Clearing each work individually favours firms that can afford lawyers, existing data contracts and years of delay. A central licence lowers that barrier. India’s proposal then faces a harder question. Which developer owes a payment, on what revenue, and which creators get it? Those decisions will determine whether the scheme opens the market or builds a new advantage for incumbents. DPIIT, working paper on generative AI and copyright · PIB, publication of the working paper

A proposal to separate training from payment

DPIIT’s December working paper examines several ways to treat copyrighted works used for generative AI training. Its committee proposes a blanket licence for training on lawfully accessed content, a royalty due when tools are commercialised and a central mechanism to collect and distribute payments. It proposes that rates be determined through an appointed committee, with judicial review available. These are consultation recommendations. Product budgets can model the proposal as a possible future liability while current rights and contracts continue to govern their choices. DPIIT, working paper on generative AI and copyright · PIB, publication of the working paper

The timing feature matters. Upfront per-work fees consume cash before a developer knows whether a model will work or find customers. Payment on commercialisation shifts some financing risk away from the entrant. Creators then bear the risk that the use earns little. In a market with a high rate of failed experiments, that is a meaningful transfer. Its legitimacy depends on whether successful products deliver a sufficiently predictable return to the pool.

Why the base matters more than the headline rate

Suppose an illustrative company earns ₹2 crore a year directly from a model API. At an assumed 1% rate on that revenue, it pays ₹2 lakh. Another company offers an equivalent model inside an office subscription earning ₹20 crore. A 1% royalty on the whole subscription is ₹20 lakh. A formula attributing just 10% of the bundle revenue to the model yields the same ₹2 lakh as the first firm. The rate is identical throughout. The accounting definition determines whether the rule treats economically similar uses alike.

A rate based on a narrow, observable sales line is easy to administer where the model is a standalone product. It creates an advantage for a company with a bundle, advertising business or related-party transfer if the model has no separate price. A rate covering all corporate revenue would charge unrelated activity. Defined allocation rules, consolidated reporting where appropriate and small-firm safe harbours could address those problems. Any exception should have a test for businesses that grow past it, since a sharp payment cliff can discourage honest reporting or expansion.

Paying creators when attribution is noisy

Allocation among rights holders has a different measurement problem. Counting files or tokens could reward duplication and favour the most easily scraped categories. Using current popularity as the sole weight could leave work in a regional language, small research press or specialist archive with a tiny share despite its scarcity. A developer can supply corpus-level information without producing a reliable estimate of any one work’s contribution to model performance. Sampling and category-level distributions may be practicable. They are still policy choices about whose claims the pool can recognise.

The best argument for voluntary agreements is that licences can be tailored. A publisher with a specialised archive can set terms for access, attribution and updates. A blanket arrangement should make routine clearance cheaper while preserving a workable market for higher-value services such as curated data and timely access. Evidence to watch includes the fee and time required to launch a small commercial model, the share of proceeds reaching smaller rightsholders and whether commercial bundles bear a comparable burden to standalone products. A transparent distribution methodology will matter as much as the number of works the mechanism says it covers.

The proposal is most promising as a way to reduce bargaining costs. Its wider distributional claim needs a royalty base that survives bundling and a payout formula that creators can understand and challenge.