The first time I used generative artificial intelligence, I felt like a kid at an amateur magic show. Is the card really floating in midair? The parents at this kind of show, of course, are less dumbstruck than the kids: The card is not floating but instead swinging on some string. It’s not magic. Not even a particularly good illusion. You simply have to know where to look.
The same goes for generative artificial intelligence. Once you know where to look, even the most powerful AI stops looking like magic. No string here — instead, look at the AI’s training data.
Training data is the information used to construct an AI. After programmers feed an AI a massive diet of training data, the AI learns to identify patterns in it and then generates output. The more data you feed an AI, the more subtle the patterns it can recognize and generate. That’s why ChatGPT can churn out travel itineraries, B-level college essays and social media marketing campaigns.
In all the hubbub around AI, it can be tempting to think that AI will eclipse us. That it will expand infinitely, until it can do all that a college-educated human can do — and more. That it will take over not only the jobs of data crunchers and coders and copy editors, but also poets and artists and high-level managers.
We are probably right to worry about some of our jobs. But many predictions about AI are overblown. The technology faces crucial limitations.