How an AI Learns, The 5 Minute Version for Writers

You wake up one day to the news that all content is now generated by artificial intelligences. Even that news was written by an AI. Can this happen? Before we go there, us flesh-and-blood writers should understand how that artificial intelligence does it.

I asked ImageFX, Google’s main AI offering, for “a drawing of an artificial brain.” Pure horror.

The Artificial Brain

The “brain” of an AI is more than just a software application. Applications can do work, but they are limited when it comes to learning and adapting.

AIs that learn to compute mass amounts of written data—which us simple humans call language—are called large language models (LLMs). These AIs can learn to interpret, converse, summarize, generate unique content (generative AI), and more. For example, at my day job I use an AI to summarize the oodles of messages posted on our internal Slack channels, and to fish out definitions for names and abbreviations of things I hadn’t heard of before and that are unique to my company. Each time I do that, I’m asked for feedback on how good the AI’s answer was, and what if anything needs improvement.

But how exactly can an AI learn? How does an artificial brain process untold amounts of knowledge, retain that knowledge, and apply it in new ways?

A Simple Example of Machine Learning

An AI is a kind of extremely complicated calculator. It does an enormous quantity of math. But unlike a calculator, AIs can learn from the experience of processing data and evaluating results. This is called machine learning.

Here’s a simple example: An AI trying to predict the length of a commute. It starts with 3 types of factors, or inputs, that use numbers to represent the value of each input. The goal is getting a result that tells us how long the commute will be.

The node is where some computation happens. Drawing by the author. Who’s not any kind of artist.

Those 3 inputs in the graphic make sense, don’t they? Weather, the particular time and day, and any traffic incidents will determine the length of a commute. Each of those contributes some values to the node, which is where computation happens to produce a number representing how long it’ll take you to get to work today.

But which of these 3 inputs is the most important, and under what circumstances? It’s hard to know without doing a lot of testing and then assigning each a number (or weight) that determines its influence on the final result, all based on the accuracy of the commute predictions.

All this computation also has to take into account that the 3 inputs can influence each other. If, for example, it’s night time, then bad weather may have more influence on the commute than it does during the day.

There are a lot of combinations like that to work with. In fact, since numbers are infinite, there are infinite combinations. We can assign different weights to these inputs all day long and create new combinations until the end of time.

Obviously we can’t, though, so we’ll have to pick some finite number of combinations and go with it. Keep testing, keep looking at results, keep adjusting. We’ll end up with many millions of nodes, each one providing a result. And then, based on how our testing goes, we do some more adjusting of the answers (called biasing) to combine the results in a way that makes the final answer (called output) accurate.

The artificial brain must train.

In reality complex AIs have a great number of inputs, not just 3. And that diagram above? There are layers upon layers of it: inputs and nodes and results and weights and biases and combinations and recombinations and so on. That’s the AI brain, called a neural network. And what’s amazing is that the whole design (or model) that determines what that brain is and does uses training feedback to learn to adjust the numbers itself to improve the outputs.

I’m representing that AI brain with the mess I drew in this final diagram.

The AI brain. By the author, who really tried his best.

There are so many layers, and so much convoluted computation, that it’s possible no one understands exactly how the output is finally arrived at. And even with the same inputs, the final output could be different each time.

And what if this AI also learns to write an essay? Or a poem? Or a short story? Or a whole novel? A play? A script? Makes a movie based on that novel or script? To different degrees, it’s capable of all of those, and it’s getting better at all of those.

Writers, it’s time to pay attention.

I’m not a computer scientist, so this is a much simplified sketch of how an AI brain works. However, I work with people who are designing the future of AI, and I want to convince you not to underestimate artificial intelligence in its early days. It’ll only get better. At everything. Does that mean human creativity is over? Remember how the game of chess was “over” when computers gained the ability to consistently beat the greatest of players? Chess is more popular today than it has ever been. If you’re a human, I’d like to read your thoughts about this topic. If you’re not, keep those thoughts to yourself.