Since 2024 Indian courts have issued a string of interim injunctions treating a performer’s voice as a protected attribute — Arijit Singh, Asha Bhosle, Shruti Haasan. Meanwhile the labels that own the Bollywood catalogue are suing OpenAI and have signed no AI training licences. The result is a market with a very high implied price and no transactions.
On 26 July 2024, Justice R.I. Chagla of the Bombay High Court signed an ex parte ad-interim injunction in Arijit Singh v. Codible Ventures LLP. Singh is one of the most-streamed voices in Hindi cinema — a playback singer, the Indian profession of recording the songs that actors lip-sync on screen. Voice conversion tools had appeared online letting anyone make any lyric come out in his timbre. The judge held that name, voice, manner of singing, likeness and signature are protected attributes of the person, and that making artificial intelligence tools capable of replicating them publicly available without consent violates personality rights.
The technical object at the centre of that case is specific: machine learning models trained on existing vocal recordings, which learn a representation of one singer’s acoustic signature — timbre, note attack, portamento, the micro-mannerisms that make a voice identifiable in three seconds — and apply it to a different source audio. They do not compose a new song the way a composer would; they take a performance and re-skin it. Doing that requires material, and the material is decades of published recordings. At that point the question stops being about personality rights and becomes a very concrete economic one: who owns those recordings, and on what terms can they license them to whoever is training the models.
A list of orders, not a law
The Singh order is not an outlier. Since November 2022, when Amitabh Bachchan obtained relief from the Delhi High Court, a chain has formed: Anil Kapoor in September 2023, Jackie Shroff in May 2024, journalist Rajat Sharma and producer Karan Johar in June 2024, spiritual figure Sadhguru in June 2025, actor Nagarjuna in September 2025. In October 2025 the Bombay High Court intervened for Asha Bhosle — one of the foundational voices of Indian film music — specifically over AI replication of her voice. Then Akira Nandan, Baba Ramdev, Shilpa Shetty, Allu Arjun, Shashi Tharoor.
On 4 September 2026, Justice Madhav Jamdar in Bombay ordered the takedown of content concerning actress Shruti Haasan, citing a “very strong prima facie case”: eighteen defendants, damages sought of 15 crore rupees, and — alongside personality rights — sections 38, 38A and 38B of the Copyright Act 1957, the performer’s rights provisions.
That last detail is the most interesting and the most easily misread. Until now Indian AI-voice litigation rested on a judge-made construct, the personality right, which protects the individual rather than the work. Invoking performer’s rights shifts the fight onto copyright terrain, where there are owners, assignments and contracts. But the outcome of that argument is not yet written: it has been pleaded, not decided.
And that is the point to hold on to against the temptation of describing India as the country that “recognised the right in one’s voice”. It has not. It has issued a series of interim orders, many of them ex parte — granted without the other side being heard. As things stand there is no final judgment defining how wide that right runs, and no statute codifying it. What exists is a fast judicial practice favourable to the famous claimant, which is a different thing and, for anyone doing arithmetic, a far less solid one.
The catalogue nobody has sold yet
On 13 and 14 February 2025 the Indian Music Industry body, together with T-Series and Saregama, filed at the Delhi High Court to intervene in ANI v. OpenAI, arguing their recordings had been used without authorisation for training. OpenAI has invoked fair use over publicly available data. India is its second-largest market by users.
Those two companies matter. T-Series releases roughly two thousand songs a year. Saregama holds over 150,000 tracks, including the repertoires of Lata Mangeshkar and Mohammed Rafi — two voices that between them carried much of twentieth-century Indian cinema. For scale: FY26 revenue of 9.85 billion rupees, about 102 million dollars, of which 8.14 billion came from music, up 17% year on year, with a record quarterly adjusted EBITDA of 1.33 billion.
Now the part that actually counts. On the 14 May 2026 earnings call, Saregama managing director Vikram Mehra said two things in the same hour. First: streaming platforms should assign no value in the royalty pool to AI-generated music. Second: “at the right time” the company will deal commercially with generative AI firms, citing the deals already signed by the global majors. Saregama has set up an “AI efficiency team” and, per statements in February 2026, is already using generative tools to produce new video around its historic catalogue.
Trade press flagged the tension itself: keeping AI out of the royalty pool while selling the catalogue to the people training AI are two positions that sit badly together. This is not an announced plan — it is a contradiction stated out loud.
As matters stand, no Indian licence for generative model training is on record. The labels’ public posture is still litigious rather than transactional. The two are not mutually exclusive — it is exactly the path Universal took when it settled with Udio in October 2025, and Warner with Udio in November and Suno in December — but until a signature exists, it remains a hypothesis.
Why the price rises while the market stands still
Here is the argument: India has built, judicially and very fast, an extremely high implied price for the voices of its best-known performers, without building the market in which that price could be paid. Every order raises the legal risk for anyone training a model on Indian material without consent, and therefore in theory raises what a licence is worth. But an interim injunction does not say what a voice is worth. It only says someone can stop you.
Waiting is not free. Deezer’s figures show the supply-side pressure: in April the platform was receiving around 75,000 fully AI-generated tracks a day, more than 44% of new uploads — up from 10,000 a day in January 2025 and 60,000 in January 2026 — with up to 85% of streams on those tracks reportedly fraudulent. The 2024 CISAC/PMP Strategy study estimated up to 25% of creators’ revenue at risk by 2028, roughly 4 billion euros. A legacy catalogue does not lose value because someone copies it. It loses value because the denominator keeps growing.
One thing almost nobody notes here. The Indian orders protect the person: Singh, Bhosle and Haasan sue in their own name and win in their own name. The recordings, however, sit on the labels’ balance sheets. These are two separate estates attached to the same sound, and so far each has moved independently — artists in court against cloners, labels in court against trainers. The day a training licence is actually signed, those two tracks will meet at a single table for the first time. Nobody, to date, has written the rules for that meeting.