Understanding AI Risk

It Ain't Rocket Science and It Ain't Really About AI

I don't understand why the risk of AI is so hard to understand. Perhaps it is because we have all seen the Terminator movies and the dancing robots and we imagine an evil SkyNet taking over. Forget all that. Until robots get tiny fusion-powered battery packs, we need not worry about Terminator-style robots.

But AI does entail risks. So I am going to lay it out for anyone that makes policy decisions or loses sleep over SkyNet.

1. AI is like a really dumb person

We discuss the limitations of AI elsewhere in Intelligence Evolved so I won't repeat them here. AI is fantastic at retrieving what somebody else has programmed into it. It is not good at reasoning or solving original problems. It will make a mistake, acknowledge that it made a mistake, and then repeat the same mistake (see Funny Flubs).

2. Giving unlimited power to a stupid or evil actor is a bad idea

Churchill said that democracy is the worst form of government except for every other form. Dictatorships and autocracies are worse than democracies because they lack checks or balances. If you would not make an idiot or a psychopath a king, then you should not put an AI in absolute control. AI is not the problem. Magnifying stupid and evil is.

3. Networks scale the damage potential of computers and AI

Humans screw up. To really screw things up, you need a computer. To scale from a class A screwup to an apocalyptic disaster, you need to connect computers to the Internet. This is exactly what we are doing with AI. Networks give AI enormous power.

I worked at a software company 20 years ago that outsourced payroll to another company. All employees were required to give the contractor access to employee bank accounts so they could direct deposit payroll—the contractor refused to print payroll checks. During one pay period, instead of depositing funds into employee’s bank accounts, the payroll company withdrew funds instead. We got our money back and our pay a few days later, but it was after our checks bounced. We got charged on over-draft fees. Our credit ratings were dinged. I am sure that a human screwed up somewhere. AI will scale this kind of human error to tens of thousands of companies and execute it in less than a few seconds.

I love the way my AI-enabled authoring tool suggests active voice alternatives to my passive voice sentences. I love the way it makes customized graphic images. These are small things with virtually no risk. AI is great at making suggestions that get reviewed by a thinking human being. AI is not bad or dangerous. It is a useful tool…but a stupid one. Giving highly-scaled, autonomous control to a stupid actor (human or AI) is always a bad idea.

---

For more on this subject, see AI: Apocalypse Inevitable?