Reminds of some emissions regulation that required catalytic converters in all new cars. Then Honda created an engine that ran so clean that it had lower emissions than American cars with a catalytic converter and a catalytic converter didn't help. Except they couldn't sell cars in the US with the new engine without also installing mostly-useless and extremely-expensive catalytic converters anyway, because of the regulation.
I could be naive here, but I guess I was giving legislatures the benefit of the doubt in supposing that AI presented challenges to existing law that they were attempting to mitigate. Specifically, I think I had heard that sometimes the makers of AI systems claim trade secrets (or something to that effect, I'm not sure of the correct term) to try to shield the exact workings from scrutiny, which seems like it could make it difficult to establish whether there is an unlawful bias--but again, I know very little about law. My point, which I'm not sure you've addressed, is that just because discrimination is already unlawful in theory doesn't obviously mean that it's difficult to get away with in practice (and if not, that seems like a good reason for additional legislation of some kind). This is the question I was curious about when I started reading your piece.
I’m no expert, but it seems to me that it’s very difficult to know based on a single decision whether it was biased, and the only practical way to combat bias is to either look at patterns of decisions or get some insight into the process by which the decision was made. In the case of AI decisions, it seems like there needs to be some kind of legal mechanism to investigate either the system’s process or its pattern of results to help adjudicate possible bias cases. I would have liked to see some discussion of such practical questions in this article. Realistically, is new law needed to make bias cases recognizable and winnable against AI processes? What should such legislation do, if so?
Your correct that the nature of AI systems can make it hard to disentangle unfair from fair decisions. The Connecticut and old Colorado law look at outcome (decisions which “in any manner that has the effect of causing” discrimination) to determine bias. This is in part why new AI laws for discrimination are unnecessary, existing discrimination law already covers business practices that create certain discriminatory outcomes.
VERY interesting material. Thanks for putting it together.
Reminds of some emissions regulation that required catalytic converters in all new cars. Then Honda created an engine that ran so clean that it had lower emissions than American cars with a catalytic converter and a catalytic converter didn't help. Except they couldn't sell cars in the US with the new engine without also installing mostly-useless and extremely-expensive catalytic converters anyway, because of the regulation.
I could be naive here, but I guess I was giving legislatures the benefit of the doubt in supposing that AI presented challenges to existing law that they were attempting to mitigate. Specifically, I think I had heard that sometimes the makers of AI systems claim trade secrets (or something to that effect, I'm not sure of the correct term) to try to shield the exact workings from scrutiny, which seems like it could make it difficult to establish whether there is an unlawful bias--but again, I know very little about law. My point, which I'm not sure you've addressed, is that just because discrimination is already unlawful in theory doesn't obviously mean that it's difficult to get away with in practice (and if not, that seems like a good reason for additional legislation of some kind). This is the question I was curious about when I started reading your piece.
I’m no expert, but it seems to me that it’s very difficult to know based on a single decision whether it was biased, and the only practical way to combat bias is to either look at patterns of decisions or get some insight into the process by which the decision was made. In the case of AI decisions, it seems like there needs to be some kind of legal mechanism to investigate either the system’s process or its pattern of results to help adjudicate possible bias cases. I would have liked to see some discussion of such practical questions in this article. Realistically, is new law needed to make bias cases recognizable and winnable against AI processes? What should such legislation do, if so?
Your correct that the nature of AI systems can make it hard to disentangle unfair from fair decisions. The Connecticut and old Colorado law look at outcome (decisions which “in any manner that has the effect of causing” discrimination) to determine bias. This is in part why new AI laws for discrimination are unnecessary, existing discrimination law already covers business practices that create certain discriminatory outcomes.