As AI enters public service delivery, examine the risks of disability bias and suggest a governance framework for inclusive AI in India.

AI now screens job applicants, triages patients and verifies welfare beneficiaries. In Rajive Raturi v. Union of India (2024), the Supreme Court read accessibility into Article 21 [1]. Disability-blind AI in public services is therefore a constitutional risk as well as a design flaw.

Risks of Disability Bias

• Biased historical data: 7 in 10 persons with disabilities are outside the labour force, against 4 in 10 others [3]. AI trained on past hiring data learns to repeat this exclusion.
• Proxy discrimination: AI résumé screeners ranked CVs that mentioned disability-related awards lower than otherwise identical CVs. Video-interview tools misread the non-verbal cues of disabled candidates as disengagement [3]. The bias enters through features that look neutral.
• Inaccessible interfaces: e-KYC "liveness" checks that require blinking locked out acid-attack survivors and blind users. The Supreme Court held that digital access is part of Article 21 [2].
• Scale without exit: One biased state system repeats the same error for every citizen, and they have no other provider. AI that monitors work performance also penalises employees who need reasonable accommodation [3].
• Regulatory vacuum: In Raturi, the Court found that guideline-based self-regulation was inadequate. It ordered mandatory rules within three months [1]. India still has no binding accessibility test for government AI.
• Silo approach: DEPwD's Mission AI Accessibility and its AI chatbot build AI for disabled people [4]. Mainstream AI run by other ministries escapes the same scrutiny.

Governance Framework for Inclusive AI

• Binding standards: Notify enforceable rules under the RPwD Act, 2016 that cover digital and AI interfaces, as Raturi directed [1].
• Risk-tiered bias audits: Make pre-deployment disability testing compulsory for high-stakes uses such as recruitment, health and welfare, and a condition of government AI purchases. Low-risk tools can get lighter checks. The EU AI Act already flags recruitment and worker monitoring as discrimination risks [3].
• Representative, consented data: Build training datasets that include disabled people, collected with consent under the DPDP Act, 2023.
• Human-in-the-loop and redress: Give every citizen a right to human review of AI decisions [3]. Extend the nodal officers and accessibility audits that the Court ordered for e-KYC [2].
• Co-design: Put disabled people's organisations on design and testing teams, so that AI is built "with lived experience at its heart" [5].
• Mainstreaming: Have DEPwD and MeitY coordinate so that disability checks apply to every AI project, not just disability schemes.

Inclusive AI therefore needs binding standards, risk-based audits and genuine co-design. Building inclusion in from the start is cheaper than retrofitting it later. With these safeguards, India's AI push can deliver the RPwD Act's promise of dignity and SDG 10's goal of reduced inequality.