Seeking Trust in AI

We want fairness in our lives. Driven largely by model behaviour that has touched an exposed nerve in our collective cultures and the growing awareness of the consequences of historical and existing bias on decision inference, global pursuit of explainable and fair AI has surged. These continued efforts have instigated new research areas, consultancy services, software and policy deliberations.

Given the nature of AI, specifically that it is artificial - is fairness from AI what we truly seek? The expected, logical and immediate response is resoundingly positive. If this is indeed the appropriate response, what then are the true objectives?

AI, or computational intelligence, is defined in Computational Intelligence: A Logical Approach as any device that perceives its environment and takes actions that maximises its chances of successfully achieving its goals. The AI algorithm is a statistical engine which searches for patterns within provided data that yields the optimal statistical outcome. For example, predicting which wood samples would have highest durability or which individuals are likely to buy a particular type of coffee.

Such a definition brings us to the first conundrum. Using the above example, an AI algorithm, while dependent on the data provided, treats both the wood and the individuals in a similar manner. It does not discriminate living from non-living, it is incapable of value judgment. The AI algorithm will statistically identify the patterns from which inference can be made on the wood with the highest predicted durability or individual with the likelihood to purchase coffee as the case may be.

Fairness is surmised as the quality or state of being, marked by honesty and treating or affecting all equally.

Here, we have the second conundrum. Treating all equally may entail that all may be affected differently, and affecting all equally may entail that all may be treated differently. The majority of work in AI fairness and explainability predominantly seeks to marry the AI statistical* *engine with the objectives of treating all equally and affecting all equally. Perhaps in an attempt to embed value judgment in AI. Recent studies, such as that by Pleiss et. al ‘On Fairness and Calibration’, indicate that an algorithm based on such approach may be potentially meaningless.

In order to ascertain fairness across groups, the characteristics of such groups need to be predetermined. Predetermination inevitably results in the assumption that individuals who share the same group attributes also share similar behaviour. Ironically, in the pursuit of fairness, these processes not only involve bias but are also counter-intuitive to the aforementioned definition of AI as they predetermine the group characteristics and behaviour. Perhaps not knowing and uniformly allocating a decision may in fact be the better action when looking through the lens of fairness. Societal expectations of fairness in AI should not become a substitute for the equitable expectations we have of ourselves.

The concept of fairness is the intersect of multiple criteria and multiple objectives. History impacts present behaviour. Identified behaviour may differ entirely from our perception of what our behaviour has been or is. Data is likely to be eternally imperfect. Algorithms will always have errors.

Given the above, when seeking to create fairness in AI, it must be highlighted that what is truly fair is a complexity where both the process and solution are illusive at best. When pit one against the other, the ideal of trustworthiness is perhaps a far better and more pragmatic high watermark to aspire to than fairness. What we need from AI is trust and honesty - at the bare minimum, mechanisms and controls that enable us to identify when things go wrong so that we can intervene.

References

• https://www.straitstimes.com/opinion/singapores-ai-ethics-model-needs-more-bite
• https://algorithmwatch.org/en/project/ai-ethics-guidelines-global-inventory/
• Pleiss, G., Raghavan, M., Wu, F., Kleinberg, J., & Weinberger, K. Q. (2017). On Fairness and Calibration. In Advances in Neural Information Processing Systems, 2017.
• Kusner, M.J., Loftus, J., Russell, C., Silva, R. (2017). Counterfactual fairness. In Advances in Neural Information Processing Systems, 2017.
• Poole, David; Mackworth, Alan; Goebel, Randy (1998). Computational Intelligence: A Logical Approach. New York: Oxford University Press. ISBN 978-0-19-510270-3.