Licence To Train: Decoding India’s Plan To Solve The AI-Copyright Debate

India has proposed what could be the world’s most ambitious move to make AI companies pay for the copyrighted content they use to train large language models (LLMs) and frontier models. 

A recently published working paper by the Department for Promotion of Industry and Internal Trade (DPIIT) has sparked debate after it recommended a new licensing regime that would let companies and developers train AI models on virtually any copyrighted material, if they procure the licence and agree to royalty terms. 

Thus far, no such regulations exist and this is the first formal proposal in this regard. Coming from the DPIIT, which manages a lot of the Startup India engagements, the recommendations have resulted in some concerns. But what exactly is the DPIIT recommending? Here’s a quick glance: 

• Blanket Access: AI developers and companies gain universal access to all copyrighted content for training purposes upon obtaining a single, blanket licence.
• Post-Revenue Royalty: Once the AI model begins generating revenue, the developer must pay a royalty to copyright holders based on pre-determined, fixed rates.
• Centralised Body: A proposed body would handle the complex task of collecting and distributing these royalties

In suggesting this major overhaul to India’s copyright laws, the government is seeking to placate major media houses and studios, many of whom have accused AI giants of unauthorised usage of copyrighted content for training.

But while the government is betting that this hybrid model can balance creator rights with tech innovation, legal experts have raised concerns that the proposal in its current form could strip creators of rights over how their work is used, reward content farms churning out junk and spawn a bureaucratic royalty-collection nightmare. What else is at stake? Let’s dive deeper.

The Cost Of Universal Access

While the intent to create a licence regime is commendable, some experts warn that the practical execution is fraught with flaws. The most immediate concern is the lack of control. 

Unlike regions like the UK and EU, which allow creators to opt out, the proposed blanket licence rule offers no mechanism to copyright holders to protect their work or restrict its modification.

“The lack of a mechanism for gate access is a concern… Anybody who is developing AI will effectively be able to use your work as long as they are willing to pay the applicable tariff,” said  Hemant Krishna, a partner at law firm Shardul Amarchand Mangaldas & Co.

Echoing the sentiment, Abhinav Shrivastava, partner at LawNK Partners, argues that the recommendations deprive copyright holders of a critical moral right to protect the integrity of their work. 

“In copyright licensing for music, there are restrictions and separate licences for editing. You just get a licence for usage, not modification or adaptation. But the proposed blanket licence system wouldn’t enable any such control over how a creator’s work is used, who is using it and for what,” he added.

The one-size-fits-all approach also means AI developers may end up paying for irrelevant data. For instance, a healthtech startup, training a diagnostic model, has little use for influencer lifestyle content. Yet under a blanket system, it would still pay for a licence to access it.

“In a blanket licence, 90% of the dataset would be irrelevant to me. I’d rather pay a premium for a specific dataset that would be more useful,” LawNK’s Shrivastava noted. 

The Compensation Puzzle

The flaw in the payment mechanism is even more critical.

The recommended policy’s core assumption that compensation can be fairly distributed via royalties appears highly questionable, given the complex nature of AI. 

Unlike music streaming, where revenue can be tracked by play count, Nikhil Narendran, partner at Trilegal, noted that the connection between a piece of copyrighted data and the final AI output is difficult to track. As such, Narendran hinted that copyright holders may not be equitably compensated for their high-quality datasets.

Elaborating on this, Krishna cited the example of two literature publishers – one that curates high-quality work by eminent writers and one that publishes amateur work. 

“Whereas the former may provide more value to an AI developer, the royalty payouts may not reflect that proportionately as a blanket licensing system doesn’t enable that differentiation,” he added. 

This lack of clarity on how to measure contribution not only makes royalty distribution arbitrary but sets up an incentive for volume plays. Without differentiating based on quality, the proposal risks rewarding entities that flood the system with low-quality content just to maximise royalty payouts, irrespective of the actual value provided to the AI model.

“The incentive to put out slop just to get royalties is high,” added Krishna. 

Need For A Simpler Solution?

Despite the flaws, experts acknowledged that the recommendations deserves credit for recognising the strategic urgency around AI and attempting a hybrid solution rather than importing a foreign model wholesale. 

“This creates a sandbox in which you can play while covering for legal exposure,” said Shrivastava, noting that the lack of a legal framework could create risks for investments and intellectual properties (IPs) of AI companies.

However, many still rue the lack of a solid foundation. 

Narendran argues for a simpler approach – monitor only AI-generated output for copyright infringement, rather than administer a royalty collection and distribution apparatus.

“The proposed policy needs extensive consultation with industry and the creator economy to find appropriate balances and manage implementation,” said Shrivastava.

Ultimately, the debate exposes a deeper tension — IP frameworks designed for static works do not fit into continuously learning systems like AI. The blanket licence proposal tries to split the difference – granting AI developers access while promising creators compensation.

But without clear mechanisms for tracking usage, attributing value or preventing low-quality flooding, the proposed mechanism risks creating a bureaucratic apparatus that satisfies neither camp. With the ball now in the Indian government’s court, it will be interesting to see how lawmakers juggle these concerns and whether the recommendations will be formalised. 

[Edited by: Shishir Parasher]