Imagine I told you in 1970 that I was going to invent a wondrous tool. This new tool would make it possible for anyone with access — and most of humanity would have access — to quickly communicate and collaborate with anyone else. It would store nearly the sum of human knowledge and thought up to that point, and all of it would be searchable, sortable and portable. Text could be instantly translated from one language to another, news would be immediately available from all over the world, and it would take no longer for a scientist to download a journal paper from 15 years ago than to flip to an entry in the latest issue.
What would you have predicted this leap in information and communication and collaboration would do for humanity? How much faster would our economies grow?
Now imagine I told you that I was going to invent a sinister tool (perhaps, while telling you this, I would cackle). As people used it, their attention spans would degrade, as the tool would constantly shift their focus, weakening their powers of concentration and contemplation. This tool would show people whatever it is they found most difficult to look away from — which would often be what was most threatening about the world, from the worst ideas of their political opponents to the deep injustices of their society. It would fit in their pockets and glow on their night stands and never truly be quiet; there would never be a moment when people could be free of the sense that the pile of messages and warnings and tasks needed to be checked.
What would you have thought this engine of distraction, division and cognitive fracture would do to humanity?
Thinking of the internet in these terms helps solve an economic mystery. The embarrassing truth is that productivity growth — how much more we can make with the same number of people and factories and land — was far faster for much of the 20th century than it is now. We average about half the productivity growth rate today that we saw in the 1950s and ’60s. That means stagnating incomes, sluggish economies and a political culture that’s more about fighting over what we have than distributing the riches and wonders we’ve gained. So what went wrong?
You can think of two ways the internet could have sped up productivity growth. The first way was obvious: by allowing us to do what we were already doing and do it more easily and quickly. And that happened. You can see a bump in productivity growth from roughly 1995 to 2005 as companies digitized their operations. But it’s the second way that was always more important: By connecting humanity to itself and to nearly its entire storehouse of information, the internet could have made us smarter and more capable as a collective.
I don’t think that promise proved false, exactly. Even in working on this article, it was true for me: The speed with which I could find information, sort through research, contact experts — it’s marvelous. Even so, I doubt I wrote this faster than I would have in 1970. Much of my mind was preoccupied by the constant effort needed just to hold a train of thought in a digital environment designed to distract, agitate and entertain me. And I am not alone.
Gloria Mark, a professor of information science at the University of California, Irvine, and the author of “Attention Span,” started researching the way people used computers in 2004. The average time people spent on a single screen was 2.5 minutes. “I was astounded,” she told me. “That was so much worse than I’d thought it would be.” But that was just the beginning. By 2012, Mark and her colleagues found the average time on a single task was 75 seconds. Now it’s down to about 47.
This is an acid bath for human cognition. Multitasking is mostly a myth. We can focus on one thing at a time. “It’s like we have an internal whiteboard in our minds,” Mark said. “If I’m working on one task, I have all the info I need on that mental whiteboard. Then I switch to email. I have to mentally erase that whiteboard and write all the information I need to do email. And just like on a real whiteboard, there can be a residue in our minds. We may still be thinking of something from three tasks ago.”
The cost is in more than just performance. Mark and others in her field have hooked people to blood pressure machines and heart rate monitors and measured chemicals in the blood. The constant switching makes us stressed and irritable. I didn’t exactly need experiments to prove that — I live that, and you probably do, too — but it was depressing to hear it confirmed.
Which brings me to artificial intelligence. Here I’m talking about the systems we are seeing now: large language models like OpenAI’s GPT-4 and Google’s Bard. What these systems do, for the most part, is summarize information they have been shown and create content that resembles it. I recognize that sentence can sound a bit dismissive, but it shouldn’t: That’s a huge amount of what human beings do, too.
Already, we are being told that A.I. is making coders and customer service representatives and writers more productive. At least one chief executive plans to add ChatGPT use in employee performance evaluations. But I’m skeptical of this early hype. It is measuring A.I.’s potential benefits without considering its likely costs — the same mistake we made with the internet.
I worry we’re headed in the wrong direction in at least three ways.
One is that these systems will do more to distract and entertain than to focus. Right now, the large language models tend to hallucinate information: Ask them to answer a complex question, and you will receive a convincing, erudite response in which key facts and citations are often made up. I suspect this will slow their widespread use in important industries much more than is being admitted, akin to the way driverless cars have been tough to roll out because they need to be perfectly reliable rather than just pretty good.
A question to ask about large language models, then, is where does trustworthiness not matter? Those are the areas where adoption will be fastest. An example from media is telling, I think. CNET, the technology website, quietly started using these models to write articles, with humans editing the pieces. But the process failed. Forty-one of the 77 A.I.-generated articles proved to have errors the editors missed, and CNET, embarrassed, paused the program. BuzzFeed, which recently shuttered its news division, is racing ahead with using A.I. to generate quizzes and travel guides. Many of the results have been shoddy, but it doesn’t really matter. A BuzzFeed quiz doesn’t have to be reliable.
A.I. will be great for creating content where reliability isn’t a concern. The personalized video games and children’s shows and music mash-ups and bespoke images will be dazzling. And new domains of delight and distraction are coming: I believe we’re much closer to A.I. friends, lovers and companions becoming a widespread part of our social lives than society is prepared for. But where reliability matters — say, a large language model devoted to answering medical questions or summarizing doctor-patient interactions — deployment will be more troubled, as oversight costs will be immense. The problem is that those are the areas that matter most for economic growth.
Marcela Martin, BuzzFeed’s president, encapsulated my next worry nicely when she told investors, “Instead of generating 10 ideas in a minute, A.I. can generate hundreds of ideas in a second.” She meant that as a good thing, but is it? Imagine that multiplied across the economy. Someone somewhere will have to process all that information. What will this do to productivity?
One lesson of the digital age is that more is not always better. More emails and more reports and more Slacks and more tweets and more videos and more news articles and more slide decks and more Zoom calls have not led, it seems, to more great ideas. “We can produce more information,” Mark said. “But that means there’s more information for us to process. Our processing capability is the bottleneck.”
Email and chat systems like Slack offer useful analogies here. Both are widely used across the economy. Both were initially sold as productivity boosters, allowing more communication to take place faster. And as anyone who uses them knows, the productivity gains — though real — are more than matched by the cost of being buried under vastly more communication, much of it junk and nonsense.
The magic of a large language model is that it can produce a document of almost any length in almost any style, with a minimum of user effort. Few have thought through the costs that will impose on those who are supposed to respond to all this new text. One of my favorite examples of this comes from The Economist, which imagined NIMBYs — but really, pick your interest group — using GPT-4 to rapidly produce a 1,000-page complaint opposing a new development. Someone, of course, will then have to respond to that complaint. Will that really speed up our ability to build housing?
You might counter that A.I. will solve this problem by quickly summarizing complaints for overwhelmed policymakers, much as the increase in spam is (sometimes, somewhat) countered by more advanced spam filters. Jonathan Frankle, the chief scientist at MosaicML and a computer scientist at Harvard, described this to me as the “boring apocalypse” scenario for A.I., in which “we use ChatGPT to generate long emails and documents, and then the person who received it uses ChatGPT to summarize it back down to a few bullet points, and there is tons of information changing hands, but all of it is just fluff. We’re just inflating and compressing content generated by A.I.”
When we spoke, Frankle noted the magic of feeding a 100-page Supreme Court document into a large language model and getting a summary of the key points. But was that, he worried, a good summary? Many of us have had the experience of asking ChatGPT to draft a piece of writing and seeing a fully formed composition appear, as if by magic, in seconds.
My third concern is related to that use of A.I.: Even if those summaries and drafts are pretty good, something is lost in the outsourcing. Part of my job is reading 100-page Supreme Court documents and composing crummy first drafts of columns. It would certainly be faster for me to have A.I. do that work. But the increased efficiency would come at the cost of new ideas and deeper insights.
Our societywide obsession with speed and efficiency has given us a flawed model of human cognition that I’ve come to think of as the Matrix theory of knowledge. Many of us wish we could use the little jack from “The Matrix” to download the knowledge of a book (or, to use the movie’s example, a kung fu master) into our heads, and then we’d have it, instantly. But that misses much of what’s really happening when we spend nine hours reading a biography. It’s the time inside that book spent drawing connections to what we know and having thoughts we would not otherwise have had that matters.
“Nobody likes to write reports or do emails, but we want to stay in touch with information,” Mark said. “We learn when we deeply process information. If we’re removed from that and we’re delegating everything to GPT — having it summarize and write reports for us — we’re not connecting to that information.”
We understand this intuitively when it’s applied to students. No one thinks that reading the SparkNotes summary of a great piece of literature is akin to actually reading the book. And no one thinks that if students have ChatGPT write their essays, they have cleverly boosted their productivity rather than lost the opportunity to learn. The analogy to office work is not perfect — there are many dull tasks worth automating so people can spend their time on more creative pursuits — but the dangers of overautomating cognitive and creative processes are real.
These are old concerns, of course. Socrates questioned the use of writing (recorded, ironically, by Plato), worrying that “if men learn this, it will implant forgetfulness in their souls; they will cease to exercise memory because they rely on that which is written, calling things to remembrance no longer from within themselves but by means of external marks.” I think the trade-off here was worth it — I am, after all, a writer — but it was a trade-off. Human beings really did lose faculties of memory we once had.
To make good on its promise, artificial intelligence needs to deepen human intelligence. And that means human beings need to build A.I., and build the workflows and office environments around it, in ways that don’t overwhelm and distract and diminish us. We failed that test with the internet. Let’s not fail it with A.I.