The Singularity Institute Blog

Despite chronic flaws, science fiction has long amused us with stories about AI.

A favorite AI stereotype is an exaggerated nerd: instant in learning, limitless in memory, yet strangely inept in social situations. It just wants to be human, or to be loved. It wants to do some touching thing that runs completely opposite to any practical AI goal.

Then there’s the other archetype: AI that is just plain evil. Whether by design or defect, the AI goes on a killing spree with weapons of mass destruction, wicked coolness, subtle creepiness, or whatever other else will rouse our red-blooded heroes into action. When the nerd goes postal, you wrastle him to the ground.

In either case, we fleshies feel smugly validated. It’s jock vs. nerd writ large, and since the AI is the nerd, we know who we must be. We may suck at math and smell like socks, but we’re the one who gets the girl.

Right.

Writers have plenty of reasons for depicting AI the way they do. Most importantly, writers of mere human intelligence simply can’t know what it’s like to be truly smarter. Moreover, any AI character interesting to humans must have something in common with them. These and plot requirements lead to a common result: recognizably human personalities, with positronic brains or acronymic names. Their special powers of intelligence are only slightly exaggerated versions of talents seen in human savants. They get along well with computers and use excessively precise language, but they still think mostly like us.

But we shouldn’t blame the producers of fictional AI when we use their products for off-label uses. It is we who are to blame when we assume that exposure to their popular conceptions qualifies us to make assertions about what real AI will be like.

Even if writers and directors had the capacity to portray greater intelligence, there could be no definitive portrayal of AI, for there are many possible designs. When you speculate today about the “first” AI, for example, you’re not extrapolating from apples to oranges. You’re extrapolating from apples to a mystery point in a range that might include tuna sandwiches, galactic clusters, and everything in between.

Even if there was only one possible kind of AI, the human brain shows us that similar minds can behave very differently given different conditions. You ought to know yourself better than anyone, but imagine that you somehow consumed all the knowledge available on the internet. Can you really predict what you might do afterwards?

Why, then, should anyone think they can predict what some hypothetical AI would do when they don’t know anything about how it works, or anything about what it might learn? It’s like trying to guess what the millionth customer at your neighborhood supermarket will have in her shopping cart, only the customer isn’t human and the store carries a million products you’ve never heard of.

The only people in a position to make even the vaguest guesses are those actually working on AI concepts that may reach maturity. But don’t expect them to know much, either. At the time of this writing, nobody has been able to prove they can guarantee conservation of core values in an AI with the potential to modify itself. (SIAI is working on this problem.)

Could you make an AI like the Terminator? Or Commander Data? Or HAL? Maybe. But why would you? Those designs are shackled by our own imaginations, and by the needs of their plots. The only mission a Hollywood AI is ideally suited for is entertainment.

In my previous post to this blog, I tried to drive home the idea that thinking is a process, and one that will soon be producible outside the human brain. I said that AI researchers wanted to do for the thought process what mechanical engineers have done for the locomotion process.

Could you imagine early industrialists purposely limiting all locomotives to the speed of a running man? How about Henry Ford forcing an 8/24 sleep cycle into the Model T?

Those are the kinds of design choices we are assuming when we think that real AI would plateau at the Hollywood level; that it would linger at a level of intellectual sophistication roughly equal to our own (plus a few savant-like tricks and silly mannerisms), and stay there long enough to star in a few films.

Perhaps we should hope that designers would choose these restrictions, for safety reasons. But history does not inspire much confidence.

Besides, we already know that even Hollywood AI can be very dangerous. We’ve seen the movies.