Will the Rise of AI Destroy Our World? A Calm Optimistic Economic Sketch of How We Will Adapt

 On my Substack, I have posted a new post that I reproduce below.

The CEO of Anthropic must be a fascinating dinner guest. At the same time that he prepares for a multi-trillion $ IPO for his firm, he is claiming that his firm’s product may end humanity. Other companies such as Starbucks and United Airlines and Tesla don’t offer such drama. Can Econ 101 offer any valuable insights here?

First, I suggest that you hear directly from Dario Amodei. Here is recent TV interview of his. He looks like an innocent dude.

He actually looks like most Assistant Professors in Economics.

When I hear Chat GPT’s leaders and Anthropic’s leaders discuss the need for an FDA regulatory body to regulate AI models, I agree with David Sacks from the All In podcast who knows his Stigler Economics. This is an earnest effort to achieve regulatory capture so that the frontier closed lab models face less competition from China and Open Source models.

The real reason I have written this column is to talk about my own work. At this link, you download my new paper.

In my paper; “Adapting to AI Risk: Lessons from Climate Change Economics”, I explore how risk averse people and firms respond to a known unknown risk. We have many, many options to protect ourselves from risk.

In my paper, I return to the lulling hypothesis. When government claims that due to its benevolence that it is protecting you from harm, people reduce their own costly efforts to protect themselves. Safety regulation evaluations have demonstrated that regulatory efforts displace effort by people to protect themselves from risks. For example, when people feel safer driving (because of mandated safety equipment in the car), they drive faster and injure people walking around and crash more.

I claim that this lulling hypothesis is relevant for AI safety. My paper explores many strategies for how we will protect ourselves in our emerging AI economy.

Now, the main way that government can enhance our safety is by enforcing liability law. If these companies are liable for damages caused by unauthorized uses of their models then this would be a strong incentive for the leading Labs to invest in safety.

One privacy question arises. Should there be randomized audits of the prompts posed to the leading labs each six months? Similar to Google knowledge about each of us, the leading labs are collecting a huge amount of information about each person and there is a question of whether this information is a private good and/or a public good that our National Security officers have a right to examine.

My article emphasizes that if ideas are now common knowledge because of AI, then rules limiting access to key physical inputs become even more important. In the past, to build a dangerous weapon —- you needed the know how and the inputs. AI addresses the former but not the latter.