So why is AI a bad idea for allocating a stimulus bill? Here is why and what AI can currently do. AI (using a very broad definition here and related inputs or capabilities so don't @ me) would be very good to make certain types of decisions in allocating stimulus funding 1/n https://twitter.com/CraigtheStoic/status/1360321908215054341

Let's use a simple example. Should I build Bridge A or Bridge B seeking to promote higher economic efficiency? AI could easily come to a better outcome than any Capitol Hill intern and much faster. No questions asked. However, what is the (word of the week alert) 2/n

Epistemological assumption about the funding priority? The phrasing of the question has an implied value priority being (in this case) economic efficiency. So what is wrong with prioritizing economic efficiency? Nothing per se but let's extend the implications briefly 3/n

This could easily result (to take a simple example again) in more public investment in Silicon Valley because there it would realize a higher return there than other places. It is also distinctly possible other places have preformed poorly because of low public investment 4/n

As the machine learning joke goes, a machine learning coder walks into a bar and the bartender asks what'll it be? The coder says: whatever he's having. The epistemological problem is that AI can make empirical judgements but not normative value judgements. 5/n

Rephrased, AI can answer which road project has a higher expected economic value but not answer which segment of society needs public investment assistance? Let me rephrase it again. AI can easily answer which is greater $100 or $50 but it cannot answer 6/n

Do you like red or blue better. It can answer, I see you wear red shirts more in pictures on your phone. However, this has additional implications. To take another simple example, there is very often a significant difference between humans stated and revealed preferences 7/n

Dating websites are a great example of this. Men for instance, will frequently say they are interested in tall blondes but then choose short brunettes. How do you know what people will value when they actually have to choose not just talk. 8/n

The short answer is AI could be used as a tool to say project A is expected to have greater economic efficiency/payback etc but given that public budgets are basically a declaration of public priorities and not necessarily economic priorities, AI cannot understand normative 9/n

Values that effectively dominate public spending priorities. There is likely place here for AI something but no way yet to avoid the budgetary, normative, and value priorities within public spending. Done.