When you're thinking about writing a patent, you want to consider two key things (aside from obvious stuff like it being novel and nonobvious):
First, it needs to be something that's not easy to work around. If you patent some particular form of learning, say, it needs to be one where there aren't tweaks that others could make that would route around the patent.
Second, it needs to cover things that are detectable in order to matter. If you patent some form of learning, but then open AI uses it but doesn't disclose what they're doing, it becomes essentially useless.
That second factor makes this pretty hard to use, I think. Most of the stuff you would want to protect from a safety standpoint is the kind of thing that is hard to identify in use, as opposed to applications which are more observable.
Source: work on AI, talk to patent lawyers sometimes, have a few software patents.
I see... so trolling by patenting something akin to convolutional neural networks wouldn't work because you can't tell what's powering a service unless the company building it tells you.
Maybe something on the lines of "service that does automatic text translation" or "car that drives itself" (obviously not these, since a patent with so much prior art would never get granted) would be a thing that you could fight over?
The US patent system is broadly considered to be broken, and patent-trolls make a healthy profit abusing it. Agressively pursuing and enforcing software patents does not strike me as a sufficient or desirable means to delaying AI timelines, but I can imagine it as tool in a sliding scale of deterrence options.
I haven't seen much discussion on this other than this question by CristianKI and there is some evidence that adversarial strategies are generally disfavored at MIRI (esp. the comment by AnnaSalamon), so I thought I'd ask: