The fear that AI will reshape how we think repeats a worry as old as writing itself — and the long record shows the effects are real but rarely the ones the alarmists predict.
Worry about artificial intelligence has settled on a specific fear: that handing reading, writing, and remembering to machines will hollow out the mental skills those tasks used to build. Students will let a chatbot draft their essays and never learn to think in paragraphs; professionals will stop holding facts in their heads because a model will always supply them. The same worry appears in one of the oldest surviving discussions of writing, a text roughly 2,400 years old whose target is writing itself.
In Plato’s Phaedrus, Socrates tells a story about the Egyptian god Theuth, who invents many arts and brings them to King Thamus in the hope they will spread. When Theuth presents writing as a cure for poor memory, Thamus refuses the praise. Writing, he says, “will create forgetfulness in the learners’ souls, because they will not use their memories; they will trust to the external written characters and not remember of themselves.” It offers “an aid not to memory, but to reminiscence,” and gives students “only the semblance of truth,” so that they “appear to be omniscient” while knowing nothing. The complaint maps closely onto current worries about search engines and language models: a new external medium lets people retrieve knowledge without internalizing it, and leaves them looking wise while knowing less.
The distance is useful because the outcome is already on record. Thamus identified a real mechanism. Cultures that transmitted law, genealogy, and epic by word of mouth had built formidable oral memory, supported by rhythm and repetition, which is what made long poems memorizable without a text. As writing spread, that trained oral memory became less necessary and less cultivated. Walter Ong, who studied the shift from oral to literate cultures, argued that writing restructures thought rather than simply recording it. Writing did move some memory practices outside the body and reduced the need for trained recall. The stronger forecast, that literacy would leave people forgetful and only apparently wise, did not match its long-run role. Writing let knowledge accumulate, allowed ideas to travel across centuries and languages, and made possible the philosophy and science that followed, including the work of Plato, who preserved the argument by writing it down.
That gap between an accurate mechanism and a mistaken forecast is the pattern worth tracking, and it rests on a straightforward point about cognition. There is no tool-free baseline mind that a new technology corrupts. People think with artifacts around them, and even ordinary thinking leans on arrangements outside the skull, such as a written list or a calendar. Psychologists call part of this cognitive offloading, the use of action or external structure to reduce the mental demands of a task. Philosophers have made a stronger version of the claim, arguing that a reliable external tool can function as part of an extended cognitive system. The human mind is usually a mind-in-an-environment, and a new technology changes the environment in which thought operates. Since every tool shapes the mind, the useful question is which parts of thinking a given technology reshapes, and at what cost.
The printing press supplies the second instance, and the worry it raised was volume rather than memory. Once books could be produced in quantity, scholars faced more text than any person could read, and they said so in terms that sound modern. The complaint predated print: Seneca, in the first century, had already warned that “in reading of many books is distraction” and advised sticking to a few master authors rather than drifting among many. In early modern Europe, as printed material expanded, the unease spread. Ann Blair, a historian of how early-modern scholars coped with the flood of print, documents a period convinced it was drowning in books, struggling to manage a sudden volume of text that felt repetitive and pulled attention in every direction. The feared outcome was a mind unable to hold a coherent body of knowledge under the weight.
The response to overload was reorganization. Readers and scholars built tools for managing texts: indexes, tables of contents, and the commonplace book, a personal notebook in which readers copied passages under thematic headings for later retrieval. These tools helped readers select and retrieve material, letting them navigate more than they could memorize, which is exactly the skill a book-rich culture requires. Reading shifted from the intensive study of a few core texts toward strategic extraction from many. The abundance that had felt like an affliction became the ordinary condition of educated life, and the coping tools faded into invisible infrastructure. The fear had identified something real, since attention is finite and more text does compete for it, but the result was a new set of mental habits rather than a ruined generation.
The change in habits is worth stating precisely. Before print was common, a scholar read a small number of texts slowly and often knew long passages by heart. After print became widespread, a scholar read more texts more quickly and relied on an index to return to a passage rather than on memory to hold it. What counted as knowing a subject moved from having it in memory toward knowing where in the literature it lived and how to get back to it. This is one ordinary way a technology changes what counts as knowing something, and it is easy to miss because the people who benefit rarely mourn the practice they no longer need.
Psychologists have measured the modern version of Thamus’s worry by studying how the internet affects memory. In 2011 Betsy Sparrow, Jenny Liu, and Daniel Wegner ran experiments on how the expectation of future access changes what people remember. Across four studies, people who believed a fact would remain available to look up later recalled the fact itself less well and recalled where to find it better. Participants who typed statements into a computer and were told the file would be saved tended to remember the folder rather than the content. The authors framed this as transactive memory, an idea Wegner had developed in the 1980s to describe how a couple or a team distributes storage across members, with one person keeping the finances while another keeps the calendar. The internet works as a nonhuman partner in that system, and the finding entered public conversation as the “Google effect,” often reported as proof that search engines are eroding memory.
The studies support a narrower claim. What they show is a change in strategy, the reallocation of effort from storing a fact to knowing how to reach it, which is the same move a scholar makes in remembering which reference book holds an answer rather than the answer itself. The evidence also asks for caution on its own terms. A large preregistered replication effort in 2018, the Social Sciences Replication Project, targeted one of the study’s experiments, the computer-priming result, and did not reproduce it. The direction of the offloading effect is better supported than that particular experiment, but the episode is a reminder that even the famous demonstrations here are more modest and less settled than the headlines suggest. Search can reduce recall for some content when people expect later access, while increasing reliance on retrieval cues. The cost, when it appears, is specific: less practice storing some content internally, and more need to judge whether what comes back is right.
The three cases share the same sequence. A tool changes the task, some established practice becomes less central, and new compensating skills develop. When writing moved memory outside the body, trained oral recitation faded, and literacy went on to expand what minds could do. Print overwhelmed readers with volume and fragmented their attention, and the culture answered with search tools and grew used to abundance. Digital search has shifted some remembering from the content of a fact toward its location, which is a kind of offloading people have practiced for a long time. In each case the alarm caught a genuine mechanism and then overshot the consequences, predicting collapse where the record shows reorganization and trade-off.
The historical pattern still leaves real risks for AI, and it makes no promise that the outcome will be benign. It does point to a more productive question than whether AI will rot the brain or leave thinking untouched. The better question is which capacity is being moved onto the tool, what is gained by moving it, and what is lost by no longer practicing the skill. Naming the capacity honestly comes first, because offloading a phone number differs from offloading the ability to build an argument, and some skills matter only for their result while others matter because the practice builds the person who can judge the result. A second requirement is that the offload preserve supervision: a calculator is safe when the user has enough number sense to catch an impossible answer, and a model is safe when the user can evaluate the structure and evidence of what it produces. Fluency makes this harder, because well-formed language can hide weak reasoning. The third requirement is institutional, because schools, workplaces, and platforms decide which forms of dependence become normal, and a society can adopt a tool in ways that keep skill alive or in ways that quietly remove the occasions for practicing it.
Large language models extend this offloading from storage and retrieval into drafting, summarizing, and stretches of reasoning, producing the finished product of thought and adjusting it to the user on request. The same feature that makes a model useful is what makes the offload deep. It is reasonable to expect, though still an open empirical question, that heavy reliance on generated summaries could leave a student reading less patiently, or that routine acceptance of generated code could leave a programmer with less grip on the structure of a system. These are questions about particular tasks and users rather than a single verdict on AI. The reasonable expectation is that AI will change how people think in real and measurable ways, as writing and print and search did before it. Each of those earlier tools reduced the need to practice some skill while making other work easier or more valuable, and the mind reorganized around them rather than declining. Thamus named the mechanism correctly at the start and misjudged the result. Externalizing a mental skill really does change it, and that change has repeatedly turned out to be a reorganization of how people think rather than the decline he predicted.