Context: Artificial Intelligence (AI) is increasingly being used by persons with disabilities to overcome barriers in everyday activities such as reading documents, interpreting images and accessing government forms.
- However, the benefits of AI will not automatically reach disabled citizens because many digital systems, datasets and AI models remain inaccessible or biased.
- This concern becomes more significant as AI expands into government services, recruitment and healthcare, alongside rapid growth in data-centre infrastructure.
- Persons with disabilities therefore need to be included in AI policymaking, system design and infrastructure planning from the outset.
Accessibility as a Fundamental Right
- In the Rajive Raturi judgment (2024), the Supreme Court recognised accessibility as an aspect of the fundamental right to life and dignity.
- The Court observed that existing accessibility rules lacked adequate enforceability, whereas the Rights of Persons with Disabilities (RPwD) Act, 2016 envisages binding obligations.
- The Court directed the Union Government to frame mandatory accessibility standards.
- Continued litigation indicates that implementation of accessibility requirements remains inadequate.
- The broader principle is that persons with disabilities should not be expected to adapt to systems designed without accessibility in mind.
Digital Accessibility Gap
- The Chief Commissioner for Persons with Disabilities has penalised 155 establishments, including government ministries, for websites and applications that persons with disabilities could not effectively use.
- Although the RPwD framework has required accessible digital services since 2019, compliance remains limited.
- This creates a fundamental concern as governments increasingly shift towards AI-enabled digital public services.
- An inaccessible AI interface can therefore reproduce or deepen existing barriers rather than eliminate them.
AI Bias Against Disability
- AI systems are not inherently neutral; their outputs reflect the data, design choices and assumptions embedded in their development.
- The AccessEval benchmark, covering nine types of disability, found that 21 language models became more error-prone, more negative in tone and more likely to stereotype when disability-related contexts were introduced.
- Research on the CLIP image model found substantially lower accuracy for photographs taken by blind and low-vision users compared with web images.
- Objects central to disabled people's lives, such as white canes and Braille displays, were significantly under-represented in training datasets.
- Such biases can also appear in ordinary interactions, where AI may respond to disability primarily through assumptions of limitation rather than addressing the user's actual request.
AI Does Not Always Translate into Independence
- A survey of 2,462 users of the NClude platform, which assists disabled people with job applications and government forms, illustrates the gap between AI's potential and actual accessibility.
- 1,313 users reported completing tasks that had previously been inaccessible to them.
- However, only 543 users were able to complete these tasks through AI alone; others required assistance from staff.
- This demonstrates the importance of maintaining human fallback mechanisms when AI systems fail or produce inaccessible outputs.
- Accessibility should therefore be assessed not merely by whether an AI tool exists, but by whether it enables independent and reliable completion of tasks.
AI Infrastructure and Environmental Concerns
- India's data-centre capacity is projected to increase more than fourfold by 2030, reaching 6.5 GW or more.
- States are competing for data-centre investments through measures such as power subsidies and duty waivers.
- Rapid expansion raises concerns regarding electricity demand, grid stress, renewable-energy sourcing and heat generation.
- Data centres consume large amounts of electricity and generate substantial waste heat, creating additional challenges for cities already experiencing high temperatures.
- Therefore, AI policy needs to consider not only the software and algorithms but also the physical infrastructure required to operate them.
Disability and Energy Security
- Energy reliability has a direct accessibility dimension for persons dependent on powered wheelchairs, oxygen equipment and other assistive devices.
- Power outages can therefore have more serious consequences for disabled citizens than for the general population.
- Inaccessible emergency alerts and communication systems can further increase vulnerability during disasters or infrastructure failures.
- AI and digital infrastructure policies should consequently incorporate disability-inclusive energy and emergency-response planning.
What Should Be Done?
- Government-deployed AI systems should undergo mandatory disability-bias and accessibility testing before deployment.
- AI training datasets should contain meaningful and representative disability-related data, collected with informed consent.
- Data-centre incentives should incorporate conditions relating to renewable-energy sourcing, electricity demand, grid impact and heat management.
- AI systems providing essential public services should retain human assistance or alternative access channels when automated systems fail.
- Persons with disabilities should participate in designing AI systems—from data collection and model development to deployment and infrastructure planning.
RPwD Act, 2016
- The Rights of Persons with Disabilities Act, 2016 replaced the Persons with Disabilities (Equal Opportunities, Protection of Rights and Full Participation) Act, 1995.
- It gives statutory recognition to the rights of persons with disabilities and expands the recognised categories of disabilities.
- It emphasises equality, non-discrimination, accessibility, reasonable accommodation and participation.
- Accessibility is not limited to physical infrastructure; it extends to information, communication technology and digital services.
- The Act is therefore increasingly relevant to AI-based governance and inclusive digital public infrastructure.
