Artificial intelligence (AI) is high on the Indian government's agenda. Some days ago, Prime Minister Narendra Modi inaugurated the Wadhwani Institute for Artificial Intelligence, reportedly India’s first research institute focused on AI solutions for social good. In the same week, Niti Aayog CEO Amitabh Kant argued that AI could potentially add $957 billion to the economy and outlined ways in which AI could be a ‘game changer’.

During his budget speech, Finance Minister Arun Jaitley announced that Niti Aayog would spearhead a national programme on AI; with the near doubling of the Digital India budget, the IT ministry also announced the setting up of four committees for AI-related research. An industrial policy for AI is also in the pipeline, expected to provide incentives to businesses for creating a globally competitive Indian AI industry.

Narratives on the emerging digital economy often suffer from technological determinism — assuming that the march of technological transformation has an inner logic, independent of social choice and capable of automatically delivering positive social change. However, technological trajectories can and must be steered by social choice and aligned with societal objectives. Modi’s address hit all the right notes, as he argued that the ‘road ahead for AI depends on and will be driven by human intentions’. Emphasising the need to direct AI technologies towards solutions for the poor, he called upon students and teachers to identify ‘the grand challenges facing India’ – to ‘Make AI in India and for India’.

To do so, will undoubtedly require substantial investments in R&D, digital infrastructure and education and re-skilling. But, two other critical issues must be simultaneously addressed: data bias and access to technology gains.

While computers have been mimicking human intelligence for some decades now, a massive increase in computational power and the quantity of available data are enabling a process of ‘machine learning.’ Instead of coding software with specific instructions to accomplish a set task, machine learning involves training an algorithm on large quantities of data to enable it to self-learn; refining and improving its results through multiple iterations of the same task.  The quality of data sets used to train machines is thus a critical concern in building AI applications.

Much recent research shows that applications based on machine learning reflect existing social biases and prejudice. Such bias can occur if the data set the algorithm is trained on is unrepresentative of the reality it seeks to represent. If for example, a system is trained on photos of people that are predominantly white, it will have a harder time recognizing non-white people. This is what led a recent Google application to tag black people as gorillas.

Alternatively, bias can also occur if the data set itself reflects existing discriminatory or exclusionary practices. A recent study by ProPublica found for example that software that was being used to assess the risk of recidivism in criminals in the United States was twice as likely to mistakenly flag black defendants as being at higher risk of committing future crimes.

The impact of such data bias can be seriously damaging in India, particularly at a time of growing social fragmentation. It can contribute to the entrenchment of social bias and discriminatory practices, while rendering both invisible and pervasive the processes through which discrimination occurs. Women are 34 per cent less likely to own a mobile phone than men – manifested in only 14 per cent of women in rural India owning a mobile phone, while only 30 per cent of India’s internet users are women.

Women’s participation in the labour force, currently at around 27 per cent, is also declining, and is one of the lowest in South Asia. Data sets used for machine learning are thus likely to have a marked gender bias. The same observations are likely to hold true for other marginalized groups as well.

Accorded to a 2014 report, Muslims, Dalits and tribals make up 53 per cent of all prisoners in India; National Crime Records Bureau data from 2016 shows in some states, the percentage of Muslims in the incarcerated population was almost three times the percentage of Muslims in the overall population. If AI applications for law and order are built on this data, it is not unlikely that it will be prejudiced against these groups.

(It is worth pointing out that the recently set-up national AI task force is comprised of mostly Hindu men – only two women are on the task force, and no Muslims or Christians. A recent article in the New York Times talked about AI’s ‘white guy problem’; will India suffer from a ‘Hindu male bias’?)

Yet, improving the quality, or diversity, of data sets may not be able to solve the problem. The processes of machine learning and reasoning involve a quagmire of mathematical functions, variables and permutations, the logic of which are not readily traceable or predictable. The dazzle of AI-enabled efficiency gains must not blind us to the fact that while AI systems are being integrated into key socio-economic systems, their accuracy and logic of reasoning have not been fully understood or studied.

The other big challenge stems from the distribution of AI-led technology gains. Even if estimates of AI contribution to GDP are correct, the adoption of these technologies is likely to be in niches within the organized sector. These industries are likely to be capital- rather than labour-intensive, and thus unlikely to contribute to large-scale job creation.

At the same time, AI applications can most readily replace low- to medium-skilled jobs within the organized sector. This is already being witnessed in the outsourcing sector – where basic call and chat tasks are now automated. Re-skilling will be important, but it is unlikely that those who lose their jobs will also be those who are being re-skilled – the long arch of technological change and societal adaptation is longer than that of people’s lives.  The contractualization of work, already on the rise, is likely to further increase as large industries prefer to have a flexible workforce to adapt to technological change. A shift from formal employment to contractual work can imply a loss of access to formal social protection mechanisms, increasing the precariousness of work for workers.

The adoption of AI technologies is also unlikely in the short- to medium-term in the unorganized sector, which engages more than 80 per cent of India’s labor force. The cost of developing and deploying AI applications, particularly in relation to the cost of labour, will inhibit adoption. Moreover, most enterprises within the unorganized sector still have limited access to basic, older technologies – two-thirds of the workforce are employed in enterprises without electricity. Eco-system upgrades will be important but incremental. Given the high costs of developing AI-based applications, most start-ups are unlikely to be working towards creating bottom-of-the-pyramid solutions.

Access to AI-led technology gains is thus likely to be heavily differentiated – a few high-growth industries can be expected, but these will not necessarily result in the welfare of labour. Studies show that labour share of national income, especially routine labour, has been declining steadily across developing countries.

We should be clear that new technological applications themselves are not going to transform or disrupt this trend – rather, without adequate policy steering, these trends will be exacerbated.

Policy debates about AI applications in India need to take these two issues seriously. AI applications will not be a panacea for addressing ‘India’s grand challenges’. Data bias and unequal access to technology gains will entrench existing socio-economic fissures, even making them technologically binding.

In addition to developing AI applications and creating a skilled workforce, the government needs to prioritize research that examines the complex social, ethical and governance challenges associated with the spread of AI-driven technologies. Blind technological optimism might entrench rather than alleviate the grand Indian challenge of inequity and growth.

source: https://www.chathamhouse.org/expert/comment/artificial-intelligence-apps-risk-entrenching-india-s-socio-economic-inequities