Is Relying on AI Cognitive “Offloading” or “Outsourcing”?

The terminology we use makes a difference.

Image generated by AI.

---

Cognitive psychologist Paul Kirschner has written a Substack post noting that he recently read two pieces by “leading scientists” who described AI as a tool for “cognitive offloading.” In fact, Kirschner argued, what people do with AI is different: it’s “cognitive outsourcing.”

What’s the distinction? Offloading, as Kirschner explains, involves writing down things we have in our heads so we can think about other things. I often use the term “offloading” to describe creating an outline before writing an essay.

Writing imposes a heavy cognitive load, especially for those of us who aren’t experienced writers and are trying to write at length. We might have to keep pausing to ask ourselves what we were going to say next, or if we’ve already made that point—and hold all of that in working memory, which has an extremely limited capacity. But if we offload what we plan to write onto a clear linear outline that we can use as a guide, we open up capacity in working memory for the many other tasks involved in writing.

Cognitive outsourcing, on the other hand, is when we ask ChatGPT or one of its ilk to write the essay for us—or even just to create the outline. Sure, it’s easier, and the final product might be better, but we’ve basically handed over our brain to a bot.

In other words, as Kirschner writes, tools that provide offloading support cognition. Tools that provide outsourcing replace it.

Just hours after I read Kirschner’s post—as if on cue—I came across a new report from the Brookings Institution warning of the dangers of overreliance on AI. It used the term “cognitive offloading” 57 times, largely to describe what I (and Kirschner) would call cognitive outsourcing. It did also use “cognitive outsourcing” 15 times, but it used the terms interchangeably and never suggested there was a difference between them.

A quick review of the cognitive science literature (aided by AI) indicates that the Brookings report and the pieces Kirschner came across are not outliers. There appear to be many studies and articles that treat the terms “offloading” and “outsourcing” as synonymous. And the AI-generated definitions of the terms aren’t all that different. Both talk about using “external tools” to reduce mental effort, and both cite some of the same examples—using a calculator, for instance.

Drawing a Clear Distinction

You might argue that it doesn’t make much difference which term you use and that “outsourcing” is just an extreme version of “offloading.” After all, both Kirschner and the authors of the Brookings report make the same basic points about the risks of AI, although the latter make them at much greater length. (I’ve discussed those risks in a previous post, here.)

But like Kirschner, I believe there’s a value to drawing a distinct line between offloading and outsourcing because the risks of each are fundamentally different. Let’s take a concrete hypothetical that, while not a perfect analogy to cognitive processes, may be useful.

Imagine a company that makes costume jewelry entirely in-house and has produced more than it can store in its own building. It then “offloads” some of the jewelry it has created by renting space in someone else’s warehouse.

Now imagine another costume jewelry company that finds it’s efficient to subcontract out some of the work. Rather than manufacturing every bead or crystal or wire, it buys those things ready-made from other companies that specialize in making them. In other words, it outsources those parts of the manufacturing process.

Nothing unusual or alarming about that. But now imagine that the company starts to outsource other components of the process: putting together the jewelry, then marketing it, then designing it. At a certain point, it’s outsourced so much that it no longer makes sense to call the company a company. It’s not actually making anything anymore. It’s just an empty shell.

Translating this to the cognitive context, offloading involves storing information that you’ve generated yourself—making a grocery list, for example, rather than trying to remember everything you need to buy. Some might argue that too much of this can weaken our ability to remember things, but it seems to me the risks are minimal and the benefits clear.¹

Cognitive outsourcing, on the other hand, doesn’t just mean finding storage space for thoughts we’ve produced through our own mental effort; it means we don’t make the effort ourselves. As with the second costume jewelry company, outsourcing can make sense as long as we don’t overdo it.

For example, most of us, if not all, would opt to have our phone take on the task of alphabetizing a contact list rather than doing it ourselves. There’s a value in having kids learn to alphabetize, but whatever learning might be involved in alphabetizing hundreds of names is clearly outweighed by the amount of time it takes a human being to do that.

Using the GPS on our phone involves more of a trade-off. If we rely on it all the time, we may lose our ability to navigate or use a map on our own. But if I’m trying to get to a place I’ve never been before or I’m not sure what traffic is like on my route, I’m happy to let my GPS guide me.

The Risks of AI Outweigh the Benefits

With generative AI tools like ChatGPT, the stakes become much higher, as the Brookings report makes clear. The authors acknowledge that AI has its uses,² and also that there isn’t yet a lot of rigorous or longitudinal research on its effects. But they argue that right now, “the risks of utilizing AI in education overshadow its benefits.”

The authors note that as students increasingly use AI, “a positive feedback loop emerges.” They’re getting better grades without putting in much time or effort, seducing them into becoming dependent.

Surveying the available research and drawing on their own interviews in 50 countries, the authors conclude that this dependence leads to a decrease in critical thinking skills, creativity, and comprehension. Teachers, they say, report “a digitally induced ‘amnesia’ … where students cannot recall information they submitted or commit that information to memory.”

There are obvious risks in allowing students to use AI to do their reading for them, or to provide superficial or simplified summaries of complex texts, but what worries me even more is the widespread use of AI to avoid writing. While we need some information about a topic before we can write about it, the process of writing itself strengthens our retention of information, deepens our understanding, and helps develop our analytical abilities.

Because writing is so difficult, however, it’s where AI may be most seductive. Why bother to sweat out a ten-page essay when a bot can do it for you in mere minutes, or even seconds? Students may not realize that by taking that shortcut, they’re cheating themselves out of a lot of learning—or even if they do realize it, they may not be able to resist the temptation.

We Need to Teach Kids to Write

The Brookings report includes various recommendations for action, including “Use AI tools that teach, not tell” and “Prepare teachers to teach with and through AI.” But there’s one possible solution that goes unmentioned: Equip teachers to teach writing in an explicit, logically sequenced manner.

Writing will always be difficult, but for many students it’s far more difficult than it needs to be. That’s because no one has explicitly taught them how to do things like construct sentences and clear linear outlines. That’s not the fault of teachers. It’s the fault of a system that has failed to train teachers in how to teach writing. In addition, the instructional materials teachers are given generally assume, erroneously, that kids will just pick up writing skills without much instruction.

Once students learn the basics of writing, it can enable them to experience the benefits of cognitive offloading. But if they outsource their writing to AI, they’ll fail to realize their full academic and cognitive potential.

We should be clear on the difference between offloading and outsourcing, but the more fundamental message is that we’re on an AI train that appears to be heading off a cliff if someone doesn’t apply the brakes—and soon. Neither Kirschner nor the Brookings authors advocate banning AI. But they do see serious risks in continuing to allow students to use it to supplant learning rather than supplement it.

The report includes some chilling quotes from teachers, parents, and students themselves about what they’re seeing and experiencing. “Students can’t reason,” one teacher said. “They can’t think. They can’t solve problems.”

“It doesn’t make sense for us to read a 50-page book,” a student commented, “when in five seconds we can research it and ask questions in AI.” (You have to wonder how familiar this student is with books, since few of them are only 50 pages.)

Another student provided a succinct explanation of the dangers of cognitive outsourcing: “It’s easy. You don’t need to use your brain.”

1. In the classroom context, for example, the authors of The Teach Like a Champion Guide to the Science of Reading have noted the value of having students jot down their thoughts for class discussion so they don’t have to hold them in working memory while waiting to speak. They describe this as a guard against the transient information effect—which it is. But it can also be seen as a form of cognitive offloading that frees up students’ working memory capacity so they can focus on what their classmates are saying.

2. In fact, they acknowledge they used AI to some extent in drafting their report.