On 21 October 2026, US District Judge Eumi Lee heard Anthropic's request for summary judgment that training Claude on copyrighted song lyrics is fair use. She did not rule. Universal Music Group, Concord and ABKCO sued in 2023, alleging that lyrics from hundreds of songs by artists such as Beyoncé, the Rolling Stones and the Beach Boys were used in training, according to MediaNama's report of the hearing. MediaNama set the hearing beside a proposal from India's Department for Promotion of Industry and Internal Trade (DPIIT) that would skip the question altogether. The proposal is not law.
The case for DPIIT's approach
The strongest argument for the DPIIT model deserves a fair statement. Courts take years, and outcomes differ by judge. Rights holders, especially independent songwriters and regional-language publishers, cannot fund Californian litigation. A developer that scrapes first and litigates later gets free inputs while creators bear the uncertainty. A statutory licence offers a predictable rule, and its payments flow to creators whether or not they can afford a lawyer.
That is roughly what DPIIT proposes. Its working paper on generative AI and copyright, Part 1, was released in December 2025. It came from an eight-member committee formed in April 2025, and the government has said it rejected a zero-price model because that would weaken incentives to create (PIB press release). Instead it proposes a mandatory blanket licence: developers could train on lawfully accessed works, rights holders could not withhold them, and royalties would fall due once an AI system is commercialised. A government-appointed committee would set the rates, subject to judicial review. Creators would be paid equally whether or not they belong to a collective management organisation, and collection would be centralised (Storyboard18). Stakeholders were given 30 days to comment.
Why the US hearing matters to the Indian proposal
The lyrics case shows how contested the underlying facts still are. Anthropic's counsel argued that training is transformative and that AI does not reduce demand for lyrics. She noted that UMG itself has described AI as a net positive for the industry. The publishers argued the opposite: that users made nearly 120,000 requests for synthetic lyrics over six months and that Claude "could overwhelm the market with AI-generated content very easily" (MediaNama). Judge Lee pressed both sides on fair use and on whether Claude's lyric-writing ability threatens publishers.
The empirical question is whether training on a work substitutes for that work in the market. Different US judges have answered it differently, and MediaNama notes that Judge Alsup and Judge Chhabria reached different conclusions in 2025 cases involving books. The relevant facts also vary by medium. Lyrics are short, easy to reproduce and sit in a concentrated licensing market. A model trained on millions of web pages, code repositories or Indian-language news does not look like that.
A single statutory rate cannot reflect that variation. Compulsory licences work best when the licensed use is standardised and the market failure is clear, as with broadcast of recorded music. Here neither condition has been shown. Setting the price before the substitution evidence exists risks setting it wrongly in both directions. A rate that is too high pushes Indian start-ups and open-weight developers out. A rate that is too low shortchanges creators and, ironically, strengthens the case that the scheme is a subsidy.
Design problems that the paper has to answer
- Attribution. Royalties distributed equally across creators avoid the impossible task of tracing outputs to inputs. But an equal-share rule pays a prolific aggregator and a one-song lyricist alike, and the amounts per creator may be negligible.
- Who pays. Payment on commercialisation is gentler on research than upfront fees. It still taxes domestic developers, who must also compete with foreign models that may be trained under looser regimes elsewhere.
- No opt-out. A rights holder who cannot withhold a work loses the bargaining leverage that licensing markets rely on. This also sits awkwardly with the speech interest in controlling how one's expressive work is used.
- Nasscom's dissent. The industry body recommended a text-and-data-mining allowance with opt-out mechanisms, while MeitY backed the hybrid model (Storyboard18). The split suggests that the question is not settled even inside government.
The existing law is narrower, but it is not the problem
India has no general fair-use defence. Section 52(1)(a) of the Copyright Act, 1957 protects fair dealing with a work for listed purposes, including private use (research included) and criticism or review. Whether model training fits those purposes is untested. The ANI v. OpenAI case, with the Indian Music Industry, T-Series and Saregama intervening, is still pending, per MediaNama. A Delhi court ruling would give Indian policymakers evidence from their own market.
A more proportionate path
India does not have to choose between a free-for-all and a compulsory price. Three steps would keep options open:
- Wait for the pending litigation and for Part II of the working paper, which MediaNama could not find, before legislating.
- If a statutory scheme goes ahead, limit it to uses where substitution is demonstrated, and exempt research and small developers.
- Encourage voluntary licensing and collective deals, which can price lyrics differently from news or code, instead of fixing one national rate.
The US court may rule for either side, and either ruling will tell Indian regulators something. A government that legislates before reading the evidence will have made the same bet as a court that decides without a record, with less chance of correcting it.