My first post, and it's a linkpost to a Substack article. I haven't posted the article here because it feels a bit tangential to EA, but that's an assumption that may be wrong. I'd welcome this community's feeback.
I'm pretty new to EA ways of articulating ideas, so I ask your indulgence on how I navigate through this.
We are short of data on the potential benefits, for individual citizens, of having a capable AI assistant to help with burdensome tasks. Annie Lowery has tried to put some numbers on the opportunity cost of the time taken up by dealing with these tasks, in her book The Time Tax: How the Government Wastes Our Time—and How to Fix It. There's a useful, non-paywalled summary here. 12 billion hours, which if paid for would cost $430 billion per year, is her floor estimate of just the time spent on paperwork. So a quick corollary is that capable AI assistance might lessen that time cost. Halving it is worth $215 billion a year in opportunity costs saved. And that doesn't include other, hard-to-quantify benefits that the article linked above discusses.
That's a pretty big potential payoff, in the US alone, in the context of US government alone. It also enables better (more productive, more meaningful, more life-enhancing) use of time. It's dependent on AI agents being capable, on being adopted, and on being accepted as helpers. We have an analogy: assistance dogs are trained to be capable and helpful, they're welcomed as such by their human users, and they're (largely) accepted where they can help (eg shops, transport, workplaces).
We know the agents are capable. One barrier to both adoption and acceptance is the concern that they're not safe or trustworthy. AI alignment efforts keep foundering on the empirical evidence that AI models are not reliably aligned, and that whatever alignment they have is determined by the AI model developer's preferences, rather than by any reliable mechanism. Perhaps there is a way to bolster safety and trustworthiness by applying "traditional" cyber security and engineering safety methods (and doubtless developing new ones) that are external to the AI models. Given the potential payoffs, in this area alone, investing time and energy in this seems a good bet to me.
We don't try to align earthquakes; we engineer buildings that can withstand them.
Link preview photo by Dzmitry Dudov on Unsplash
