INEFFABLE

Generative AI is the fastest growing technology in human history. More than 800 million people use it at least weekly, and data centers account for about half the growth in the United States gross domestic product.

There is no question that large language models can sort through more information than a human brain and do that orders of magnitude faster than we can. There is, however, an ongoing argument about whether we are on the verge of artificial super intelligence—a machine that can do everything a brain can do and do it better—or whether that is even possible.

To deal with that question, we need to start by defining intelligence. Think about what a large language model is. One feeds it a huge amount of information composed of words, parts of words, or combinations of words broken into tokens. The input is all language. The machine then does its statistical magic to decide which words go together to reply to a prompt which is also in language and produces an answer in—you guessed it—language.[i]

If we measure AI against the brain solely on how many facts one or the other contains and how fast each can get to those facts (semantic memory), the machines have already won. Of course, they make mistakes and make things up, but those are solvable problems. But more to the point, semantic memory is not the only kind of memory.

What about stored experience? We have a rich library of memories of the people, places, and situations we have encountered during our lives, and the library is almost all in images rather than words. Those memories help us make predictions and know what to do when we encounter similar situations.

Large language models only have the words that people have used to describe experiences. They have no experiences of their own. A number of AI experts who think superintelligence is inevitable—most prominently Dario Amodei of Anthropic—recognize that problem and offer a solution. Their “world view” comprises AI-equipped robots that can wander around seeing, hearing, and touching. The robots have all the senses of a human and remember what they have encountered. Does that mean they have experiential memory equivalent to that of humans?

Perhaps language does not fully represent experience. A situation described in the same words may have entirely different impact from one person to another. The same view from the edge of a cliff may be exhilarating to one person and terrifying to another. Metaphors are one way we account for experience (“Sky high” or “Scared to death”), and metaphor is one of the things large language models are especially bad at.

Another way to look at this is to ask whether there is something in human intelligence that takes place without words. It seems to me that the answer is yes. Consider for instance, the experience of pain. Not only do we experience pain without putting words to it, we actually have a hard time finding useful words. In reacting to pain, we resort to sounds that pre-date language—“Ouch” or “Arghh”—or expletives that are still sounds without meaning. As a surgeon one of my most difficult tasks was getting a reliable sense of a patient’s pain. We had a short list of descriptors (stabbing, throbbing, burning) and an inadequate set of measures (one to ten or various frowny and smiley faces) but the risk of over or underestimating the amount of pain or not figuring out where it came from was a constant source of worry and a frequent source of error.

Much of what makes experience powerful is the emotion associated with it, and that emotion is entirely personal to the person having the experience. In addition, it is intricately tied to an endocrine system that humans have and machines lack. We release oxytocin in the presence of human attraction and norepinephrine when we are scared for instance, and the endocrine response is involuntary and entirely personal.

Beyond that, we can all generate images in our brains that are independent of language. There have been recent experiments using electrodes implanted on the brain’s surface in people who have lost the ability to generate speech. Those brains generate specific, reproducible patterns when the people are asked to imagine doing a specific activity. Just the activity—not the words that describe it. Think about that. You have no trouble generating an internal picture of a thing, a place, or a situation that is quite clear and uses no words. A large language model can describe that internal world, but it cannot reproduce it.

Brains can create metaphors, they can generate emotion, and they can paint ineffable pictures. Generative AI can do none of those things.

And that brings us to the last point. If we are to make the best use of generative artificial intelligence, we need to maximize what we can do that the machine cannot. Much of that may be in the realm of ideas outside the bounds of language. That is the sort of right brain creativity we have undervalued at least since we learned to write, more since the Enlightenment, and even more since we digitized information.^[ii] Perhaps we will have a different measure of intelligence as we learn to work with AI.

[i] I am only dealing with large language models here, although the same argument would apply to AI that generates images, music, and so forth.

[ii] My thanks to Sean Trott and his Substack “The Counterfactual” for helping put this into words.