Is AI Always Error Prone?

Ben Affleck says artificial intelligence is always going to be error prone. A calculator, the kind of computing people already know, gives the same precise result every time someone says four times four. Artificial intelligence is probabilistic. It works like a bookmaker setting a line and finding a mean, then landing in a zone. When a result is wrong, or when the system states something with tremendous authority and then admits the fault, the miss is how the technology does what it does successfully.

In 1982, traders had computers sell a stock as soon as it hit a certain number. Everybody did the same thing. The computer sales dumped the stocks, caused a panic, and wiped the market out. The lesson was to stop handing the wheel over entirely to the technology.

Ben Affleck calls the wish to remove the human from the loop the most pernicious notion. Treating efficiency as a count of how many people can be eliminated will fail. Nobody wants to live in a world like that. The technology needs management and expert guidance. Coding harnesses still have to be sat with, fixed, and iterated. Left unchecked, they will likely sail off and rewrite half a code base. They are bounded by context, they often forget things, and the permissions they receive have to be chosen carefully.