Can we operationalize AI ethics?
The AI Ethics Lab creates an AI index listing companies/institutes that subscribe to some AI principles. But what do they really mean? Are they just general promises, with no way to measure them? And even if we could quantify and assess them, would there even be a penalty, since these ethical codes are essentially promises without real consequences?
Quotes from here:
It's not clear that all those declarations mean much at this point. A study by the AI Ethics Lab published in December, in the prestigious journal Communications of the ACM, concluded that all the deep thinking by those organizations couldn't really be put into practice.
As Cansu Canca, director of the Lab, wrote, the numerous declarations were "mostly vaguely formulated principles." More important, they confounded, wrote Canca, two kinds of ethical principles, what are known as core, and what are known as instrumental.
What I find interesting about this paragraph is the implicit idea: if you want to talk about operationalizing, it should at least be clear which principles are fundamental, i.e. they are not themselves operationalizations but are the things we aim to operationalize.
Drawing on longstanding work in bioethics, Canca proposes that ethics of AI should start with three core principles, namely, autonomy; the cost-benefit tradeoff; and justice. Those are "values that theories in moral and political philosophy argue to be intrinsically valuable, meaning their value is not derived from something else," wrote Canca.
This would for example mean that transparency and explainability are instrumental, because they are valuable insofar as they uphold the intrinsic ethical values of human autonomy and justice. This also has the implication that instrumental values are more interchangeable, whereas a conflict in intrinsic values always requires further in-depth ethical deliberation.
Notes on the source paper here: Canca - Operationalizing AI Ethics Principles .
Remark from Ryan Calo, same source:
"Principles are largely meaningless because in practice they are designed to make claims no one disputes," said Caro. "Does anyone think AI should be unsafe?"
Instead, "What we need to do is roll up our sleeves and assess how AI affects human affordances, and then adjust our system of laws to this change. "Just because AI can't be regulated as such, doesn't mean we can't change law in response to it."
TODO: https://cacm.acm.org/magazines/2020/12/248788-operationalizing-ai-ethics-principles/fulltext
Notice the remark fragmented tooling and framework landscape:
Abhishek claims the book will recover the signal from the noise in the "fragmented tooling and framework landscape in AI ethics." The book promises to help organizations "evoke a high degree of trust from their customers in the products and services that they build."
Suggests it's valuable to structure this fragmented landscape.