Bhashini-Powered Govt Chatbots Fail Basic Conversations

The Union government, in a recent Lok Sabha session, highlighted its flagship AI-powered chatbots as proof of how artificial intelligence is transforming citizen services in India. 

However, a fact-check by AIM revealed that several of these tools, including Kisan e-Mitra, Skill India Assistant and Bhashini, struggle with basic usability and fail to address the needs they were built for. Contrary to the LS response, which painted a picture of accessible, multilingual and citizen-friendly tools that can help people navigate schemes and services with ease.

The National Informatics Centre (NIC) defines a chatbot as “a piece of web-based software that conducts a conversation via auditory and/or textual methods and is often designed to simulate a human conversation partner”. 

Promises Vs Performance

Kisan e-Mitra, introduced under the digital development drive of the agriculture and farmers’ welfare ministry, is positioned as a farmer’s AI companion. The ministry claims that the chatbot supports multiple languages and is evolving to assist with other government programs. At present, it handles over 20,000 farmer queries daily, and so far, more than 92 lakh queries have been answered, it asserted. 

The app opens in Hindi by default, leaving users to manually change language settings, a step that could be a barrier for those with limited digital knowledge.

Once operational, it did not respond to most of the queries, even the ones specifically pertaining to agricultural schemes.

Even after several attempts, the voice bot was unable to capture the command.

For every conversation initiated, the chatbot replied, “Sorry, there seems to be some problem. I request you to check your internet connection, then refresh the session by clicking on the refresh button in the bottom left corner, and then try again.”

To seek clarity, AIM reached out to Amitabh Nag, Bhashini CEO and director of IndiaAI, who said the platform is “only open to registered farmers”. However, the chatbot itself presents a different picture.

There is no login required to access the service, and it explicitly notes that no personal details such as Aadhaar number, phone number or name are required unless specifically asked.

Most of India’s AI systems or chatbots, including Kissan e-Mitra, are powered by the government’s flagship language translation platform Bhashini. However, Bhashini’s capabilities remain limited, as it primarily focuses on basic translations.  

Meanwhile, Sunbird UCI, which co-powers the Kisan e-Mitra chatbot alongside Bhashini, declined to comment on the matter. 

Nag earlier told AIM that the Bhashini platform is “useful when I want to talk to a taxi driver in Kannada but I don’t speak the language. It works, and it works affordably.” 

Yet, when tested in real-life scenarios, the results tell a different story. For instance, when given a simple Malayalam prompt: “I want to go till Kadavanthra, how much will it cost?”, Bhashini translated it as “How much do I have to go to the mortgage?” (image attached)

In practice, these translations are often inconsistent or inaccurate, which affects the quality of responses citizens receive when interacting with AI-powered public services.

The Skill India Assistant (SIA), the government’s generative AI chatbot for job seekers, faces similar problems. Intended to help citizens discover training opportunities, the chatbot frequently fails to interpret straightforward queries such as “Where can I learn computer skills near me?” Instead, it returns irrelevant links or incomplete responses, leaving users without actionable information.

When AIM tested SIA, it often produced irrelevant results. Manu (name changed on request), a communication designer searching for opportunities on SIA, was shown listings for interior designer, Figma designer, and even appliance repair technician. “It’s a mix of relevant and irrelevant job openings,” he said.

Instead of providing direct guidance or solutions, they redirect users to generic help pages.

Inside Perspective from IndiaAI

A person working closely with the IndiaAI initiative, speaking on condition of anonymity, acknowledged that these problems are not isolated. “Most of the chatbots in India by the government are not working,” they said, adding that this is “an open secret” among those in the sector.

The source also shed light on how such tools are portrayed in official communication. “Whatever PIB releases come to us, we are supposed to cover it without even verifying the background or fact-checking it.” 

Further, when the Supreme Court of India launched its AI-driven tool, Supreme Court Portal for Assistance in Court’s Efficiency (SUPACE), it was hailed as a step toward modernising the judiciary. But concerns around its limited accessibility and lack of transparency have surfaced now.

“The problem with SUPACE is that it is used only by SC judges. There is no transparency with regard to the use, accuracy, recall, and the training data set,” said advocate Joseph Pookkatt. Unlike global best practices in AI deployment, the system’s working remains largely opaque.

Currently, only a handful of Supreme Court and select high court judges have access to SUPACE. “If such a tool is not accessible to advocates, litigants or the public, it is in effect restricting broad-based judicial innovation and scrutiny,” Pookkatt pointed out.

Questions also remain about accountability. No independent audits, accuracy reports, or details about the algorithms have been released. “Given SUPACE’s closed, judges-only operation and the absence of an external audit, it risks becoming a black box system,” Pookkatt added.

He added that citizens often recognise the gap between claims and reality. “As a voter, you know what is going around,” he remarked, pointing to the disconnect between promotional narratives and everyday experience.

India is making massive bets on AI at the infrastructure level. Under the IndiaAI Mission, approved in 2024, the government has committed ₹10,300 crore over five years to build one of the world’s largest public AI compute facilities, equipped with 18,693 GPUs.

This makes it nearly nine times more powerful than the open-source model DeepSeek and about two-thirds the size of ChatGPT’s infrastructure, the government claims.

While AI-driven chatbots could make public service delivery efficient and accessible, the current implementation leaves much to be desired. Without robust multilingual support, better natural language comprehension, and continuous user testing, these tools risk becoming symbolic gestures rather than functional solutions.

The post Bhashini-Powered Govt Chatbots Fail Basic Conversations appeared first on Analytics India Magazine.

Generated by RSStT. The copyright belongs to the original author.

Source