Fairness / tackling the biases AI

Based on the premise that a large set of well-diversified data may be an accurate description of the world, most of the developer community takes a technocratic attitude that data-driven decision making is good and algorithms are neutral. However, this argument does not recognize the fact that the existing data may have biases, which may have got reinforced over time. The issue of fairness is at the forefront of discussion in academic, research and policy fora, and definitely merits a combined dialogue and sustained research to come to an acceptable resolution. One possible way to approach this would be to identify the in-built biases and assess their impact, and in turn find ways to reduce the bias. This reactive approach, use-case based, may help till the time we find techniques to bring neutrality to data feeding AI solutions, or build AI solutions that ensure neutrality despite inherent biases.

Transparency / opening the “Black Box”

Presently, most AI solutions suffer from what is commonly known as the “Black Box Phenomenon”, with very little or no understanding of what happens in between and only the input data and results being the known factors. This is due to the reliance in most current AI systems to incrementally improve the performance as defined by a narrow set of parameters, with developer’s emphasis being less on how the algorithms are achieving the requisite success. However, calls for explaining the decision-making process will gain momentum as AI systems are increasingly relied upon for decision making that has significant consequences for a large section of population. Opening the Black Box, assuming it is possible and useful at this stage (there is considerable debate on that as well), should not aim towards opening of code or technical disclosure – few clients of AI solutions would be sophisticated AI experts – but should rather aim at “explainability”. With extended disclosure though, what needs to be balanced is whether the algorithm’s parameter may induce the individuals and companies to change their behavior and in turn game the system. Clearly, more collaborative research is required in this area.

Source: Page 85-86; NITI AAYOG National Strategy for Artificial Intelligence, June 2018.