Artificial Intelligence (AI) has penetrated all areas of our lives and law is no exception. Even the Supreme Court of India has decided to use AI to transcribe live proceedings.

The SC had called for sealed bids in a three-cover system (EMD, technical bid and financial bid) from experienced and reputable firms for the design, development and implementation of AI solutions for transcribing arguments and court proceedings.

Following which, a paper titled “Analysing the endeavours of the Supreme Court of India to transcribe and translate court arguments in light of the proposed EU AI Act,” published on September 20, 2023, categorises this application as high-risk AI under the rules of the proposed European Act. This Act seeks to guarantee safety and adherence to the fundamental rights and values of AI systems within the European Union.

“I believe that the European Union’s AI Act provides one of the most comprehensive guidelines for using AI responsibly. While it’s not perfect and there may still be room for improvement, it’s a very impressive piece of legislation,” says Kshitiz Verma, Assistant Professor, JK Lakshmipat University, and the author of the paper.

Verma has taken the example of the 2019 Anil Ambani case to highlight the point: “As long as AI is committing fewer mistakes than what human beings would commit, it should be fine to use; otherwise, it is not.”

“Caution is essential, as seen in the case of Anil Ambani, where a crucial word was mistakenly omitted by court employees. This led to the Supreme Court having terminating their employment. The incident underscores the potential impact of even minor errors in legal proceedings,” adds Verma.

AI in transcriptions

Speech recognition technology has lagged in development compared with vision and natural language processing. While significant advancements have been made, such as training models with millions of hours of data, like OpenAI’s Whisper, their accuracy remains imperfect, with error rates around 12 per cent.

Additionally, diverse accents in India and variations from the Western English datasets make speech recognition unique and challenging. “In a noisy and chaotic courtroom environment, even humans may struggle to accurately recognise words, especially with India’s diverse accents. Expecting AI to flawlessly identify every word and speaker in such conditions, considering the multitude of Indian accents, is a formidable challenge,” says Verma.

To improve the accuracy of transcriptions, it’s essential for the Supreme Court to provide a wide range of videos for training datasets. AI models are crucial, but the quality and diversity of data are equally important.

AI in Translations

The success of natural language processing and machine translation varies significantly among languages. Research and resources are more abundant for languages in higher resource classes, while languages like Hindi and other Indian languages fall into lower classes.

This means the datasets and labels for these languages are much smaller compared to English. Translating legal documents from English to Indian languages requires caution, as it may introduce confusion due to varying interpretations of words. The eighth schedule of the Constitution recognise 22 languages; however, AI isn’t efficient enough to translate all of them.

Needs human intervention

Increasing AI system risks necessitate a corresponding increase in human oversight. Those overseeing AI must have sufficient AI literacy and authority for thorough investigations when needed.

This oversight aims to mitigate risks to health, safety, rights and the rule of law while considering specific risks and the level of automation. This prevents automation bias and unquestioning acceptance of AI outputs.

In terms of the safety, Verma says, “If an AI violates the fundamental rights of the citizens and their safety, then it should be considered a high-risk AI. And it should be used after a lot of conformity and precaution,” adds Verma.