AI Book Club: Machines Behaving Badly: The Morality of AI by Toby Walsh
I have some questionable views of what makes a good poolside read, and recently pondered the morality of AI by a resort pool with Professor Toby Walsh.
Machines Behaving Badly: The Morality of AI covers very similar subject matter to my last AI read (Moral AI And How We Get There) but in a vastly different way. Both discuss the ethics of AI in terms of privacy, transparency, fairness and justice, but refreshingly with very different views on the all.
While Schaich Borg and friends seem to err on the side of pushing for transparency with algorithms (to be able to explain how an AI has come to a decision), Walsh disagrees. His stance is that we should be building trust between users and the AI at large, instead of a blanket rule of transparency. Transparency across the board has issues, like IP and to prevent bad actors for manipulating AI for nefarious purposes.
Another stark contrast was whether we approach what is right through the lens of fairness or equality, which at face value may seem like synonyms but in reality can have very different outcomes. A very reductive definition (sorry) for equality is everything being roughly the same for everyone to level the playing field, but fairness seems to be more nuanced, and takes other factors into consideration. As Walsh so beautifully illustrates with this example: in 2012 the Eu mandated that insurance companies charge all genders the same for equality. Sounds fair, yeah? What happened in reality was that women ended up paying higher premiums to subsidise men’s poor driving, because men are far more likely to get into accidents than women. The same goes for AI, sometimes a one-size-fits-all approach might not be right.
I enjoyed how Walsh laid out his lessons for each chapter, and there were far too many to list, but here were some of the more salient for me:
Be wary of machine-learning systems where we lack the ground truth and make predictions based on proxy for this.
Do not confuse correlation with causation, because AI systems with this may further perpetuate society’s biases.
Fairness means many different things to do people/cultures/communities/industries, and often they can be at odds with each other. This may require trade-offs.
Building trust in AI is key. Do this through explainability, auditability (think a little black box for robots), robustness (not breaking with little changes), correctness, fairness, respect for privacy and transparency (which Walsh again reminds us is overrated).
We judge AI so much harder than humans, but Walsh notes that this is not a double standard. He writes: “We should hold humans to a higher standard than humans. And this is for two important reasons. First, we should hold them to higher standards because machines, unlike humans are not and likely can never be accountable. We can put up with a lack of transparency in human decision-making because, when things go wrong, we can call people to account, even punish them. And second, we should hold machines to higher standards because we can. We should aspire to improve the quality and reliability of human decision-making too.
Find a copy via Black Inc Books.
