4 ms·

The question becomes, how much understanding is enough? Do we need AI engineers passing laws or training courses for politicians? Can the AI be treated like a

by ismail 7y ago

The question becomes, how much understanding is enough? Do we need AI engineers passing laws or training courses for politicians? Can the AI be treated like a black box [0] and the regulation still be effective? I think in some cases the answer would be yes. A simple thought experiment demonstrates this. Example: Consider regulation for AI in relation to job losses. Does one need to know the details of AI tech to develop regulation effectively? I would argue not. The economic impacts can be analysed. At some tipping point, job losses Lead to a cycle of: reduced demand -> reduced Revenue in companies -> reductions in jobs -> reduced demand and so on and on. To regulate this, one would have to model the economic impacts as best as one could. Understand what jobs are most likely to be lost, in which industries. Understand the key things to monitor (I.e replacement job rate. Assuming jobs are created in other industries). And pass regulation to mitigate some of the risks or make sure the number of Job losses never hit the tipping point. None of these items leads me to believe the person passing the regulation requires knowledge of AI at an engineers level. There is the Conant-Ashby Theorem or good regulator theory [1]. This states “Every good regulator of the system must be a model of the system” [2] In this instance the regulation is to deal with ECONOMIC impacts. Therefore an economic model is needed, while impacted by the tech the tech does not need to be modelled. Thus my conclusion. This poses an interesting question, could all eventualities be catered for in the regulation? There is other theories that could help us answer this question. Happy to hear different or contradictory views. [0] https://en.m.wikipedia.org/wiki/Black_box#Other_theories https://en.m.wikipedia.org/wiki/Black_box#Other_theories [1] https://en.m.wikipedia.org/wiki/Good_regulator https://en.m.wikipedia.org/wiki/Good_regulator [2] http://pespmc1.vub.ac.be/books/Conant_Ashby.pdf http://pespmc1.vub.ac.be/books/Conant_Ashby.pdf

sailfast 7y ago
Enough to reason about what the thing does in a basic manner and potential pros and cons? That seems like a straightforward heuristic but maybe I'm an optimist haha. You don't need to know the ins and outs of the mathematics to understand what encryption does, but you should be able to reason about how data is encrypted and what is necessary to decrypt, and what law provides you to request decryption. You sure as hell should also be able to grasp that, like the atomic bomb, once out in the world it's there to stay and you can't really prevent it from being used. As for economic impact - I find that measure does not really get ahead of the problem. In the case of machine learning if you're regulating economic impact you're decades late. The first-movers already have the data sets, the training has occurred and now you're living in their world. I think it's important to understand a bit more upstream so that you can get ahead of these kinds of issues. Furthermore, a baseline understanding of these algorithms and their potential biases / impacts of misclassification is important to understand to ensure we don't compromise society by their mis-use, or shift scales in a terrible way unintentionally. The earlier the better, and economic impacts are probably a lagging indicator, or even worse considered a huge boon for those areas that have companies doing this work but using labor from elsewhere to train their models.

by ismail 7y ago

sailfast 7y ago

Enough to reason about what the thing does in a basic manner and potential pros and cons? That seems like a straightforward heuristic but maybe I'm an optimist haha. You don't need to know the ins and outs of the mathematics to understand what encryption does, but you should be able to reason about how data is encrypted and what is necessary to decrypt, and what law provides you to request decryption. You sure as hell should also be able to grasp that, like the atomic bomb, once out in the world it's there to stay and you can't really prevent it from being used. As for economic impact - I find that measure does not really get ahead of the problem. In the case of machine learning if you're regulating economic impact you're decades late. The first-movers already have the data sets, the training has occurred and now you're living in their world. I think it's important to understand a bit more upstream so that you can get ahead of these kinds of issues. Furthermore, a baseline understanding of these algorithms and their potential biases / impacts of misclassification is important to understand to ensure we don't compromise society by their mis-use, or shift scales in a terrible way unintentionally. The earlier the better, and economic impacts are probably a lagging indicator, or even worse considered a huge boon for those areas that have companies doing this work but using labor from elsewhere to train their models.