After criticism, govt clarifies: AI startups don’t need IT Ministry approval

On Monday, Chandrasekhar said in a post on X: “(The) advisory is aimed at the significant platforms and permission seeking from MeitY is only for large platforms and will not apply to startups… it is aimed at untested AI platforms from deploying on Indian Internet.” He said that the process of seeking permission, labelling of platforms that are under testing and consent-based disclosure to users is an “insurance policy to platforms who can otherwise be sued by consumers”.

At the time of issuing the advisory, the government had not explicitly mentioned that the conditions will only apply to bigger companies and not smaller businesses – it was sent to all online intermediaries.

However, since the government’s advisory was in many ways to limit the impact content from such platforms could have on the upcoming Lok Sabha elections, it is unclear why an exemption to start-ups has been offered, given that their systems are also prone to generating inaccurate information. The Indian Express had recently reported about Ola’s beta generative AI offering Krutrim’s hallucinations problems.

With Lok Sabha elections set to be announced soon, the IT Ministry had on Friday sent an advisory to generative AI companies like Google and OpenAI and to those running such platforms — including foundational models and wrappers — that their services should not generate responses that are illegal under Indian laws or “threaten the integrity of the electoral process”.

Platforms that currently offer “under-testing/unreliable” AI systems or large language models to Indian users must explicitly seek permission from the Centre before doing so and appropriately label the possible and inherent “fallibility or unreliability of the output generated”.

As per the advisory, the government also wants these platforms to add in a traceability requirement by adding an identifiable marker to content they generate in a way that can be traced back to the person who has instructed the service to create pieces of misinformation or deepfakes.

The advisory had drawn sharp criticism from founders and investors in the generative AI space. Pratik Desai, founder of KissanAI, which built the agriculture large language model Dhenu, had tweeted, “I was such a fool thinking I will work bringing GenAI to Indian Agriculture from SF (San Francisco). We were training multimodal low cost pest and disease model, and so excited about it. This is terrible and demotivating after working 4 years full time bringing AI to this domain in India.”

After Chandrasekhar’s clarification, Desai said, “Not for startups! A good start… The next step would be defining revenue and user base numbers to classify a startup, which would clear the uncertainty cloud for many of us.”

At the heart of the disagreement is a tussle between lawmakers and tech companies over the future of safe harbour protections to generative AI platforms like Gemini and ChatGPT. It is also as much about the government’s view of outputs generated by some of these platforms and if it disagrees with them, even if they may or not be entirely unlawful.

Google’s AI platform Gemini had recently come under fire from the Ministry of Electronics and Information Technology (MeitY) for answers generated by the platform on a question about Prime Minister Narendra Modi.

Although, it should be noted that outputs generated by such platforms depend on a number of factors, including the underlying training data scraped from vast swathes of the Internet and algorithmic filters added on top of that, a number of instances of errors — called hallucinations — generated by these platforms have been reported across the world can typically be attributed to shortcomings in these factors.

With wrapper models, that is where a separate entity uses open source foundational models prepared by another company, there is an additional level of complexity added to the final responses that such platforms may generate depending on their algorithmic filters.

Though the advisory is not legally binding, questions have also been raised on the legal basis – under which law the government can issue guidelines to generative AI companies since India’s current technology laws do not directly cover large language models.

The advisory was sent as “due diligence” measures that online intermediaries need to follow under the current Information Technology Rules, 2021. On Saturday, Chandrasekhar had said the reason the advisory specifically mentions the integrity of the electoral process is in the backdrop of the upcoming Lok Sabha elections. “We know that misinformation and deepfakes will be used in the run-up to the election to try and impact or shape the outcome of the elections,” he had said while responding to a question by The Indian Express on whether the advisory went beyond the remit of existing IT Rules.