The IRS currently analyzes data from filed returns and looks for patterns from firms that are already under suspicion. It usually takes years to unravel the Gordian knot of accounting that surround dubious partnerships and good old fashioned tax evasion. The researcher's propose a system that targets partnerships and looks not just at individual returns (which on their own seem legit) but the whole network surrounding those returns that add up to fraud.
Because the system is always looking for signals of nefarious company practices, it wouldn't need to be focus on an individual source. It could just waits until a series of tax regulations are being used concurrently that usually means someone is exploiting the system for their own gain. And, the system could evolve to detect new ways tax evaders are cheating the government.
In their paper the researchers state: "Our approach is to model the co-evolutionary arms race between transaction sequences in ownership networks with their corresponding audit observables." Which is a fancy way of saying, it can help the IRS win. Of course it has to convince the same entity that makes you fill out a mountain of forms if you buy a house. So it might be a while before AI starts making sure we're all paying our fair share.