Kerala HC Mandates Adalat AI for Witness Deposition Transcription

Dona recently spoke to Medianama for their report on the Kerala High Court's recent mandate to use Adalat AI for witness deposition transcription starting November 1, 2025.

This Court-mandated use of Adalat AI’s transcription tool in the state’s district courts makes Kerala the first Indian state to mandate AI transcription as a default.

Dona cautions about the risks of seemingly innocuous AI-enabled tools for transcription, drawing from her ongoing engagement with our work on AI in courts and voice technologies. Read some of the excerpts from the report below:

"100% accuracy is not possible [with such tools]. [Errors can] creep into the model during the training stage. For example, If the training dataset is too clean (like with no background noise), it might overfit in real-world settings and not perform well."

Malayalam as an “Indic, Low-Resource Language”

AI models require “vast quantities of text” to become accurate, but for Indic languages, there is a relative paucity of training data available. The digital ecosystem for these languages… is also quite limited. Malayalam is likely to be a “low-resource language” which means that it suffers from a significant scarcity of linguistic resources, such as corpora… and annotated datasets. Most AI models are also “oriented towards the alphabet” and struggle to process Indic “abugida” scripts (where letters are “mushed together”), which can lead to errors.

She further explained that speech is a “complex space for transcription tools,” outlining the following basic risks:

On Linguistic Complexity: Malayalam has a range of dialects… On top of that, the state also has a significant population from other states… [with] different accents. People code-switch and code-mix in speech, switching between languages across phrases and sentences, especially between English and Malayalam in court settings. Moreover, legal language is different from both colloquial and literal Malayalam.
On Courtroom Chaos: Court proceedings will also have multiple speakers – the prosecutor, defence lawyer, judicial officer, and witnesses, sometimes speaking over each other… Courtrooms are noisy settings, so the mic and tool should be able to pick up isolated voices and reject background noise.
On Digital Infrastructure: The use of transcription tools would also require good quality microphones… Depending on whether the tool is accessed from a server or is software on a digital device, considerations like consistent electricity and internet will play into enabling uninterrupted services.

Gap in Accountability

If an AI transcription does contain errors, who is responsible for correcting such errors, and what would be the guidelines for the same? The Kerala HC policy on AI use seems to be placing the full responsibility on the “meticulous verification” by the judge. However, could this not potentially be equally time-consuming since it has to be “meticulous”? Dona Mathew also highlights the risk for sensitive cases:

For in-camera proceedings or cases of a sensitive nature, personal information may be shared during depositions. Risks arise from the handling of such data – is the system storing the data… Where is it stored, who can access it, and will it be used for fine-tuning the model…? As keepers of data from cases, courts have to ensure that service providers are fulfilling data security and protection obligations and handling data in a responsible manner.