At Dataiku, we like to talk about Everyday AI. This is a way of expressing the culture change that needs to occur before investments in artificial intelligence can pay off. Everyone from the boardroom to the frontlines needs to be thinking about data — how and when they collect it, how and when they process it, and how and where they secure it.
In the Middle East, if we went looking for Everyday AI, we should probably start our search in the United Arab Emirates (UAE), the home of innovation and the first country on Earth to appoint a minister of state for AI. The UAE is set to see a 14 per cent bump to 2030 GDP thanks to AI, according to PwC. This is, to say the least, not nothing. But as innovation continues in the country, we should make mention of “fairness” — a tricky term, but one that calls for clarity if we are to deliver Everyday AI at scale.
At Dataiku’s 2022 Everyday AI Conference in London, Luke Vilain, Data Ethics Senior Manager at Lloyds Banking Group called fairness a multi-faceted “socio-technological problem”. This implies that fairness requires a comprehensive Responsible AI strategy, where we ensure that the benefits of the technology are distributed equally across society and do not bring undue advantage to a privileged few. “No discrimination” would be another way to put it.
Fair vs ethical
Lloyds uses a special tool for fairness and transparency (often referred to in AI circles as “explainability”). The tool tracks various categories on which it is prohibited to discriminate, including age and race. If a decision is made based on criteria that are irrelevant to the matter at hand, then the algorithm or analyst that made the decision would objectively be considered as being discriminatory. An example could be two groups of people, one of which were all successful loan applicants and the other unsuccessful, and the only difference between them was race.
In fact, machine learning models must be biased in order to reach conclusions but being swayed by non-relevant data is considered discriminatory bias. We should also bear in mind that, as Vilain pointed out in his talk, fairness and ethics are not necessarily the same thing. Fairness denotes balance; ethics is more socially defined and brings in notions of morality and a subjective assessment of what is “good”. For example, if someone stole your orange and you stole their apple, that could easily be characterised as fair, but its ethical properties are a matter of debate.
Fairness can be quantified and that is why we focus on it when engaging in responsible AI. We must ensure that at each stage of construction in our machine-learning models, we do not introduce discriminatory bias. We must do this during data sampling and during the labeling stage when we are training the algorithm to recognize categories. Simple things like the treatment of null values, averages, and outliers can introduce counterproductive biases.
Diversity is crucial
And we should not forget to account for the team that designs the algorithms and pipelines behind the model. Diversity is key. A balanced team will ensure a balanced output. But even if you take all the right approaches, externalities such as legal restrictions on the gathering and use of data can prevent an ideal outcome. Data proxies can be the answer, such as when European insurance companies inferred gender from car color because they had data that suggested males should be charged higher motor insurance premiums (but were legally barred from asking applicants for this information). Proxy metrics can be a powerful tool but given the metadata surrounding them and the probabilistic nature of the data itself, they can have an adverse effect on explainability.
Assuming enterprises can define fairness in a straightforward way, the next step is introducing it into a responsible Everyday AI framework. Data and AI teams must be exposed to clear standards, simple processes, and streamlined oversight. Proper AI governance, implemented through platforms that track processes and give automated alerts based on policy, will ensure stakeholders and decision makers can keep track of developing biases and remediate any variances so that unfairness does not emerge.
As with any best practice, vigilance is an indispensable part of maintenance. Good governance can erode if, once a process is developed and deployed, it is left to roam wild. Everyday AI only works if mechanisms are in place to monitor outcomes continuously. Such mechanisms’ primary functions will be to ensure that no bad outcomes are fed back into the model from which they came and that errors are corrected quickly and efficiently.
Beware the purist
Vilain went as far as to urge companies to build centres of excellence dedicated to fairness and explainability. The centres would be tasked with making sure AI team members of different disciplines worked together toward fairness. While fairness may seem simple on paper, the need for persistence calls for constant assessment and the ability to make bold decisions. As time goes on, fairness requirements may change. Training data may prove inadequate after a time and may need to be updated manually to ensure fairness. Purists may object, but in the end, if fairness is not part of the outcome, then can the model really be considered to be in working order?
This is the case that must be made to dissenters who believe fairness is an accidental side-effect that we can live without as long as we follow the data where it leads. We already know that data itself can be the problem, so fairness advocates must be prepared to dig in and make the argument that unfair AI is not responsible AI. And irresponsible AI is no AI at all.
