The Chief Minister of Maharashtra, Devendra Fadnavis, has announced the development of an Artificial Intelligence tool in collaboration with the Indian Institute of Technology, Bombay to identify illegal immigrants from Bangladesh and Rohingya Communities. This tool will deploy linguistic analysis to scrutinise speech patterns, tone and specific word usage to determine an individual’s origin.  This piece argues that reliance on opaque algorithmic decision-making undermines procedural fairness under Article 21 of Indian Constitution and risks violating International Human Rights Law (IHRL).

Dialect, Language and Nationality

The foundational flaw in using speech analysis – relying on accents and linguistic patterns – lies in the shared linguistic heritage of the Indian subcontinent.  The nature of the data used to train such a system remains unclear but it may include publicly available speech datasets, government-held audio records, and field-level collections. State authorities have previously misidentified Indian citizens as ‘foreigners’, leading to wrongful deportations.. An AI tool relying on Bengali dialects is inherently error-prone, as linguistic boundaries between West Bengal, Assam, and Bangladesh significantly overlap. Experts have shown concern noting that the Bengali language has regional variations and no unique markers to distinguish the language spoken in states like West Bengal, Assam and other parts of India from Bangladesh.

Historical context further underscores the problem. Following the Partition of India and the creation of Bangladesh in 1971, a consequence of the Bangladesh Liberation War, millions of people migrated across borders. Hence, language and its particular dialects remain rooted in geography, and not nationality or religion. For instance, the Bengali-speaking population in Assam’s Barak Valley speak Sylheti, a dialect spoken in Bangladesh’s Sylhet district. Any AI system attempting to infer citizenship from such speech patterns is likely to be structurally incapable of reliable accuracy without an extraordinarily large dataset.

AI Opacity and Risks of Violation of Fair Procedure

In Selvi v. State of Karnataka, the Supreme Court of India recognised that Article 21 included substantive due process rights.. In  Maneka Gandhi v. UOI, the Court held that any procedure affecting life or liberty must be fair, just, and reasonable. Thus, individuals faced by state decisions must be able to understand, challenge and have access to evidence and reasoning relied upon against them.

Most AI models predict through deep neural networks, operating as ‘black box models’ where only the input and the output is visible to users while the reasoning process remains opaque. Currently, most AI platforms run on the black box model and the probability of implementing such non-transparent systems is high in this case. This lack of explainability undermines procedural fairness, as affected individuals cannot understand, contest, or rebut the basis of adverse decisions.

India’s recent experience with algorithmic tools illustrates these dangers. In the investigations relating to the North-East Delhi riots, the Delhi Police used facial recognition tools like ‘Amped FIVE’ and ‘AI Vision’ without disclosing methodology or accuracy rates.  Matches above 80% were treated as positive identifications. However, such thresholds can be problematic. For instance, as the American Civil Liberties Union pointed out, a similar threshold applied in Amazon’s Rekognition system resulted in significant false identifications.

These concerns are magnified in sensitive contexts such as immigration enforcement when ‘error rate’ is measured in human lives, including refoulement, family separation, and statelessness. Such outcomes undermine India’s commitment towards IHRL and United Nation’s AI Resolution to promote transparent and explainable AI systems.

Privacy

While K.S. Puttaswamy v. Union of India recognised privacy as a fundamental right, India’s legislative framework offers limited protection against AI-driven profiling in immigration enforcement. The Digital Personal Data Protection Act, 2023 lacks specific safeguards for sensitive data such as linguistic biometrics. Broad exemptions under Sections 7(c) and 17(2)(a) permit state processing in the name of sovereignty. Further, Section 44(3) weakens transparency by permitting the government to withhold information involving personal data, effectively insulating opaque AI profiling systems from public and legal scrutiny.

Conclusion

The use of AI to determine citizenship raises profound concerns for due process and human dignity. While the European Union has recognised a right to explanation for AI systems, India has no comparable statutory recourse. In a context where dialects cut across borders and communities, algorithmic classification is structurally unreliable, rendering individuals without recourse.

As the Bombay High Court observed in TPL HGIEPL Joint Venture v. Union of India, ‘the principles of natural justice and fairness are too valuable to be sacrificed at the altar of AI expediency.’ Deploying opaque and error-prone systems in citizenship determinations directly infringes Article 21, which guarantees life, personal liberty, and procedural fairness.