AI By The People, for The People
Why we need to democratise artificial intelligence
By Dharshi Harindra
Never before has technology played such a prominent role in the day to day lives of every individual. Our sick and elderly are being set up with Zoom accounts to keep in touch with loved ones; luddite country leaders are urging us to download contact tracing apps; professional and personal gatherings are being held online; social media networks appear to have long overtaken analog ones; and fake news travels faster over those digital networks than real news.
The stakes can’t get much higher.
How do we come up with ways to solve the COVID-19 mysteries that will lead to its eradication, as well as put us on the front-foot in the event of future viruses?
High stakes can call for innovative technology solutions. But their value can only be entrenched if developed transparently and inclusively.
Cue Artificial Intelligence
Or more specifically, machine learning. Around the world, technologists, corporates and researchers are coming up with new and clever ways to address the COVID-19 pandemic. DeepMind, a key player in the AI field, owned by Google’s parent company, Alphabet, is trying to establish which drugs could work best against COVID-19 by using data on genomes to predict organisms’ protein structure.
Similarly, London based company, BenevolentAI has re-tooled its existing systems powered by AI to investigate treatments for chronic diseases, to investigate potential treatments for Coronavirus.
Others are using data to be able to understand spreads by location, and forecast those that are next likely to be hit.
But without proper human oversight, AI isn’t all that intelligent. And if it’s to lead the charge in automating our lives and solving our biggest problems, more manpower, with diverse skills and backgrounds need to be involved in the creation of tools, and their every iteration.
By itself AI is not a silver bullet. For machine learning tools to be effective, they need vast amounts of data as fodder. In the context of COVID-19, it’s unlikely that we have enough data to feed the tools in such a way as to replace human oversight in epidemiological studies for example. Further, just as has often been the case with AI tools in the past, we need to be wary of what data is feeding the machine, to avoid data resulting in biased results, particularly given the imperative to develop products at pace.
Bias in, Bias out
There have been some high profile instances of AI tools showing direct discrimination in the past few years. Since 2014 Amazon sought to use its swathes of recruitment data to develop an AI recruitment tool to automate the hiring process – throw in 100 resumes, and the tool would spit out 5 or so to hire. But the tool only highlighted the reality based on those previous years’ data – that Amazon had a preference for hiring men. The tool was ditched in 2018 after it became clear that the product was positively discriminating against women.
Similarly, and with arguably higher stakes, purported “predictive justice” risk assessment tools have been used in the US justice system with alarming results. Among other things, these tools use opaque algorithms to suggest sentencing terms and likelihood of re-offending. However reports have found that these algorithms support racial bias in law enforcement – black defendants would be tagged as future criminals to the tune of twice of that of white defendants. Also, judges applying these tools typically have limited, if any, understanding as to how these scores are computed.
Even in the COVID-19 context, biased outcomes could occur. In a recent interview, AI specialist, Alex Engler of the Brookings Institution explained that “men, for instance, are more likely to be smokers, and they also show higher mortality risk. But if you didn’t account for the fact that they were smoking, or that there was smoking in their medical history, your algorithm might show that all men are more at risk, and thus all men are going to get prioritised for care – hypothetically.”
Data Quality as well as Quantity
So not only do we need high volumes of data to develop effective machine learning tools, but the quality of that data is key to producing products that benefit society without entrenching, and amplifying negative biases. This means that we can’t necessarily rely on existing data without testing it in front of a diverse group of people. Data going in, and the resulting output need to be put under rigorous scrutiny.
Further, the machinations of AI tools need to be transparent to all. It is often a black box, decipherable only by software engineers.
Just as bias in AI is implemented by humans, so too can humans mitigate its discriminatory impact. We need a much more diverse group of people in the room at the outset when deciding what data is inputted into AI, what the issues are, and what we are trying to avoid.
AI has the potential to entrench non-discrimination, but only if it’s designed and developed transparently and inclusively.
An AI product is only as good as the data that goes in it. Biased data in means biased data out, amplified. In order to eradicate bias in AI we need a balanced spread of views and backgrounds going in to the initial decision making process. We also need to question whether, just because data exists, is it the type of data that we want to feed a decision making process. Just as our courts in criminal cases, we need a jury of decision makers to review inputs, outputs, and iterations that can be developed and updated as times, needs, and social norms change.
