The government has reiterated that artificial intelligence developers should receive a blanket licence to use lawfully accessed content for training models without individual negotiation, despite opposition to this recommendation from multiple industries.
A white paper published by the Office of the Principal Scientific Advisory (PSO) on Sunday threw its weight behind the Department for Promotion of Industry and Internal Trade's (DPIIT) recommendation that India adopt a hybrid model of intellectual property and copyright.
A DPIIT paper released in December last year recommended that royalties become payable only once the AI tool is commercialised, with rates set by a government-appointed committee. The latest PSO paper supported this and a centralised mechanism handling royalty collection and distribution, which it said will reduce transaction costs, provide legal certainty and support equitable access for both large and small AI developers.
A blanket approval for AI models scraping published content has found opposition from many.
According to industry body Nasscom, entities holding rights over content should have clear protection against commercial and non-commercial text and data mining (TDM). "For content that is publicly accessible online (freely accessible without paywalls, logins, or other access restrictions), rightsholders should be able to reserve their works from TDM through a machine-readable opt out, at the point of availability. For content that is not publicly accessible, rightsholders should be able to reserve their works from TDM through contract or licence terms," it said in December.
Firms have also argued that tracking which specific data "contributed" to a probabilistic AI output to distribute royalties is technically impossible.
The paper called for creating India-centric benchmarks for AI systems ensuring they are evaluated against Indian cultural norms and tested across multiple scripts and dialects. Creating benchmarks specifically designed to identify bias and stereotyping relative to Indian social identities should be a focus, it said. "Benchmarks designed around specific governance priorities, such as fairness across languages and demographics, enable regulators to systematically test these aspects and set expectations," it said.
Currently, India's benchmark ecosystem is developing through a mix of public digital platforms that institutionalise evaluation, academic benchmarks for Indic-language capability and risk, and domain and multimodal benchmarks aligned to India's public-service and market needs, it said.
Digital India Bhashini and Universal Language Contribution APIs are developing a language technology evaluation stack. It also pointed out tools such as Indic-bias developed by Bhashini that can evaluate the fairness of large language models across 85 Indian identity groups, focusing on bias and stereotypes, and IndicXTREME, a human-supervised benchmark of nine diverse natural language understanding tasks across 20 languages, featuring 105 evaluation sets.
A white paper published by the Office of the Principal Scientific Advisory (PSO) on Sunday threw its weight behind the Department for Promotion of Industry and Internal Trade's (DPIIT) recommendation that India adopt a hybrid model of intellectual property and copyright.
A DPIIT paper released in December last year recommended that royalties become payable only once the AI tool is commercialised, with rates set by a government-appointed committee. The latest PSO paper supported this and a centralised mechanism handling royalty collection and distribution, which it said will reduce transaction costs, provide legal certainty and support equitable access for both large and small AI developers.
A blanket approval for AI models scraping published content has found opposition from many.
According to industry body Nasscom, entities holding rights over content should have clear protection against commercial and non-commercial text and data mining (TDM). "For content that is publicly accessible online (freely accessible without paywalls, logins, or other access restrictions), rightsholders should be able to reserve their works from TDM through a machine-readable opt out, at the point of availability. For content that is not publicly accessible, rightsholders should be able to reserve their works from TDM through contract or licence terms," it said in December.
Firms have also argued that tracking which specific data "contributed" to a probabilistic AI output to distribute royalties is technically impossible.
The paper called for creating India-centric benchmarks for AI systems ensuring they are evaluated against Indian cultural norms and tested across multiple scripts and dialects. Creating benchmarks specifically designed to identify bias and stereotyping relative to Indian social identities should be a focus, it said. "Benchmarks designed around specific governance priorities, such as fairness across languages and demographics, enable regulators to systematically test these aspects and set expectations," it said.
Currently, India's benchmark ecosystem is developing through a mix of public digital platforms that institutionalise evaluation, academic benchmarks for Indic-language capability and risk, and domain and multimodal benchmarks aligned to India's public-service and market needs, it said.
Digital India Bhashini and Universal Language Contribution APIs are developing a language technology evaluation stack. It also pointed out tools such as Indic-bias developed by Bhashini that can evaluate the fairness of large language models across 85 Indian identity groups, focusing on bias and stereotypes, and IndicXTREME, a human-supervised benchmark of nine diverse natural language understanding tasks across 20 languages, featuring 105 evaluation sets.
