A Relic. Several years from now, I wonder if these two words might describe the very paragraph you are reading right now, a collection of sentences drafted painstakingly by a human author. It underwent multiple iterations—editorial notes included—before the final version confronting you. I struggled through word and sentence permutations until they felt acceptable, capturing the muscle and musicality I’ve come to believe characterizes my “voice.”
This next paragraph, on the other hand, was generated by ChatGPT: fluent, orderly, and almost indecently quick. It arrived with polish familiar to anyone who has spent time with these systems—the balanced clauses, the clean transitions, the faintly frictionless sheen of language assembled without visible exertion. Its construction was effortless: no false starts or private irritations at a stubborn phrase, or the slow negotiation between thought and expression. Only a prompt, a pause measured in seconds, and a paragraph that sounds plausibly composed, perhaps even refined, while having cost almost nothing to produce. (In keeping with the editorial spirit of The Walrus, this will be the only paragraph in this article generated with the chatbot.)
Perhaps, dear reader, you can feel the difference between prose sculpted through human sweat versus that generated synthetically. As recent research has demonstrated, artificial intelligence flattens language like a skilled bureaucrat, emitting polished text, but with turns of phrase that are distanced, manicured, and formulaic. Yet, to the extent that AI drains language of its soul, a second and more subtle inquiry is emerging: What, if anything, does it do to the inner life of the writer? What did I lose by not drafting the preceding paragraph myself? And, more broadly, what are the implications of this technology for the cognition of adults?
After three years of sustained use by the public, the data on generative AI is yielding clear answers.
T he refrain from AI proponents is straightforward: generative AI represents a change in how professionals work, not how they think. Tasks that once required deep focus—drafting this article, for example—can now be accelerated and partially automated, shifting the effort from composing to supervising, where individuals review content for accuracy, ethics, and taste. According to one study, ChatGPT led professional writers to spend less time on rough drafting and more on editing and idea generation.
The problem with this viewpoint is that modern AI systems do not merely automate tasks but participate directly in reasoning efforts that were once confined to individuals. Central to this idea is the practice of cognitive offloading, or any action that reduces the mental effort needed for a task. In our case, it means delegating portions of the thinking to external tools and technologies. While such behaviour is not new—evident in the use of maps, calendars, and calculators—the depth and scope of offloading to large language models (LLMs) is unprecedented, with significant implications for adult cognition.
At the centre of this dynamic lies a key tension: the very systems that accelerate surface-level cognition—platforms such as ChatGPT—reduce the frequency with which professionals exercise the cognitive skills that distinguish baseline competency from mastery. More specifically, these are the skills required to collaborate with machines effectively: strong writing abilities to craft precise prompts, and critical-thinking capacities to evaluate outputs, identify flaws, and iterate thoughtfully. Yet, paradoxically, these are the exact skills that erode when generative tools are leveraged without restraint.
The mental implications of generative AI extend beyond these two key areas, so much so that it is helpful to distinguish between first-order effects—those that are visible and easily observed—and second-order outcomes, which are subtler, cumulative, and often more consequential over time.
A t the most fundamental level, generative AI represents a shift from generation to evaluation, from writing an essay to reviewing one generated synthetically. Such a change reflects a profound shift in the way cognitive systems are exercised. Whereas blank-page creation (writing) necessitates the building of mental models—constructing causal chains and resolving ambiguities—evaluation is episodic and reactive, a lighter exercise that samples outputs instead of reasoning through them. The effects of this switch are significant.
An MIT study on student essay writing, for example, reported that 83 percent of participants who used generative AI could not recall a single quote from their own essay, whereas 89 percent of participants who did not use AI assistance were able to do so. At a minimum, the study highlights the diminished recall of factual information when AI is used as a writing tool.
The implications of the technology, however, extend beyond memory. The same study found that individuals who relied on AI were weaker at reconstructing chains of reasoning and transferring their knowledge to novel contexts.
Your author can attest to this fact with embarrassing clarity. Because I wrote it, the first paragraph in this article still hums in my mind: the use of the word “relic” at the beginning, the decision to go with “collection of sentences” instead of “consortium” in the same phrase, which was fun but a tad too pretentious. The second paragraph authored by ChatGPT, however, is a blur. There is a vague recollection of a polished passage with words such as “fluidity” and “exertions,” but I have no idea what these words are doing.
Simply put, generative AI removes the struggle of information processing from intellectual work—a prerequisite for long-term memory formation and, more fundamentally, learning itself.
A nother surface-level effect of these tools is what a Microsoft Research study termed “mechanised convergence”—the phenomenon that users relying on generative systems tend to produce narrower, less diverse outputs than those working independently. A similar pattern was reported in Nature, where researchers found that while AI tools expanded the scale of scientific work, they often limited focus and originality. This result may seem surprising given AI’s capabilities, but as technologists point out, an LLM is simply providing the statistically most plausible response to a question, narrowing the range of answers one is likely to consider.
Again, your author can attest to such behaviour in writing this piece. Given thematic direction and precise instructions, generative AI will most certainly produce content that is useful, functional, and occasionally illuminating. Ask it for an original or novel insight—connecting mechanized convergence to the deeper observation that LLMs are designed to recombine existing patterns rather than invent entirely new ones—and the response will likely underwhelm. While GenAI can produce astounding results, as evidenced by the Mayo Clinic’s recent announcement about being able to detect pancreatic cancer up to three years before clinical diagnosis, guidance from an expert and human hand appears paramount.
A final implication of the move from execution to oversight is the mental fatigue that results when deep, deliberate activities are replaced with managing multiple streams of synthetic work—what the Microsoft study terms “task stewardship.” As the research illustrates, the responsibilities of prompting, reviewing, and correcting outputs for content one hasn’t authored can be both exhausting and of limited value in internalizing the material. A lawyer using ChatGPT to research case law, for example, may save time retrieving information only to spend it verifying citations that have been embellished or fabricated. Likewise, a junior engineer reviewing AI-generated code may understand the surface logic of the program while never fully grasping the underlying principles that produced it.
What’s more, when the contemplative cadence of authorship is replaced with the rapid-fire mode of AI, we lose something precious: the pauses, the breaks, the mental white space where ideas simmer before they boil.
W hile the first-order consequences of GenAI are noteworthy, a deeper set of effects emerges when cognition is outsourced to artificial intelligence—ones that begin to reshape the habits around thinking itself.
The introductory paragraph in this article took two hours to draft and entailed a process not unfamiliar to writers across the literary spectrum: false starts, followed by cringeworthy phrases and poorly expressed sentiments, before the gradual, sometimes painful, arc to polish and clarity. Such struggle is what researchers term productive friction—the cognitive resistance one encounters with sufficiently complex tasks. And it is through such friction that individuals perform the labour that blossoms into competence, expertise, and occasionally mastery.
Generative AI collapses this friction with impunity. Drafts appear instantly, and explanations arrive fully formed. Efficiency improves, it is true, but the technology also eliminates those moments when individuals traditionally discovered what they did not yet understand. For students and early-career professionals, generative tools present a precarious trade: the acceleration of surface-level competence at the expense of incremental struggle—the mental equivalent of building a house on sand.
If generative AI represents a profound shift in how adults author content, it is also changing the way it is consumed—specifically, the move from deep attention to sampling. The study of AI-assisted workflows reveals that individuals process synthetic outputs in short, iterative bursts—prompting, scanning, adjusting, and moving on—such that mental models that are built when one is immersed in material are supplanted with more fragile forms of pattern recognition.
An executive examining an efficiency briefing note prepared by ChatGPT, for example, may move quickly between summaries and bullet points, without remaining with the material long enough to develop a coherent mental model of the matters at hand. Such a shift is profoundly reshaping how professionals develop expertise. What happens when technology erases the friction of both grappling with material (deep attention) and constructing it internally in one’s own mind (authentic authorship)? The history of manufacturing automation in the 1980s provides an instructive parallel.
In her seminal paper “Ironies of Automation,” Lisanne Bainbridge observed a striking paradox: as systems became more automated, as operators were removed from routine and repetitive tasks, they became worse at handling scenarios in which those systems failed. In other words, widespread mechanization removed individuals from the very experiences that were necessary to sustain competence in rare, high-stakes situations where human judgment was essential.
It is not difficult to apply this finding to the cognitive realm—the Bainbridge paradox of generative AI: by automating the routine, the technology leaves professionals unprepared to handle the non-routine, whether for writing, law, radiology, or coding. Across professions, AI is reducing the opportunities individuals have to exercise the very skills they require when the technology fails. The result is a loss of judgment in unfamiliar situations and a reduced capacity to intervene when AI systems produce flawed or misleading outputs.
Take the example of computer programming, where industry titans have heralded AI’s ability to generate sophisticated code and elegant algorithms. Left unsaid is the engineering prowess needed to maintain such systems and the comprehension debt that accrues when generative AI is used as a development tool. As critics have noted, the dangers of AI-generated code include not only its fragility under unforeseen circumstances but the difficulty engineers encounter in understanding systems they did not meaningfully create.
While platforms such as Claude and Cursor can undoubtedly accelerate the development of software, they are most useful for senior engineers who cultivated their skills before such tools existed—learned professionals who anticipate failures before they occur and combine their contextual understanding of virtual systems with the strengths of the machine to produce robust applications.
By automating granular tasks once performed by junior workers, generative AI may be destroying the training ground for our next generation of experts.
A re there remedies? In a recent interview, Terence Tao, the renowned mathematician who won the Fields Medal in his thirties, offered an interesting analogy. A hundred years ago, when food was scarce, ideas around diet and exercise were not part of our vocabulary. It was only after the Green Revolution, when food became abundant (at least in privileged parts of the world), that human beings began thinking deliberately about nutrition and physical activity.
In the same way, says Tao, individuals and knowledge workers must treat the brain as a muscle that requires constant resistance to difficult problems to remain sharp in a world of effortless answers. In practice, the most effective strategies are those that reinject a degree of mental struggle into cognitive work.
One way is to draft first, prompt second. Resist the urge to prompt AI with impulsive, half-formed thoughts. Pause, open a separate note-taking app, and draft a prompt as if you were writing a letter. Deliberate, thoughtful prompting not only facilitates critical thinking but will yield more fruitful responses from the model.
Next, leverage AI as a tutor, not an answering engine. One might call this Socratic Tutoring, which favours prompts such as “walk me through this” instead of “give me the answer” and asks AI to challenge your assumptions and thinking.
But that only solves part of the problem. Another habit worth cultivating is treating generation as cheap and evaluation as expensive. The amount of time spent on a task prior to generative AI should now be used to verify its outputs. This includes identifying hallucinations and gaps in synthetic outputs to mitigate the effects of offloading.
It might also be a good idea to use AI selectively—that is, leverage the tool to heighten your understanding of a particular topic rather than as the default tool for all tasks. Exercising restraint helps prevent dependency while keeping core faculties engaged.
Lastly, fast. Designate specific days or projects as AI-free zones, particularly for skills that are key to your profession (writing, statistical reasoning, debugging, etc.). Deliberate practice zones are a powerful way to maintain and even enhance cognition.
As generative AI increases in power and scope, the discipline embodied in the points above—along with growing calls to use the technology deliberately and with caution—will become ever more pronounced. Prudence with AI, however, includes not only best practices with the tools themselves but restraint and wisdom beyond them. As the growing body of research indicates, seasoned professionals are more likely to resist using AI as a shortcut, engaging it, instead, to deepen their thinking in a way that augments cognition. In other words, they understand not only how to prompt GenAI but when not to use it.
Professional seniority aside, a final safeguard against offloading is AI literacy and an appreciation that, at their core, LLMs are providing statistical responses they deem most useful to you. Such bias was dramatically illustrated in a Harvard study which evaluated AI for medical purposes. In this case, the system gave markedly different answers, when presented with the same clinical facts, depending on the persona—patient, doctor, insurer—that queried it. As this example illustrates, using AI intelligently requires understanding bias, rigorously checking factual claims, and appreciating the technology’s inherent limitations.
I n January of this year, researchers who worked for OpenAI announced that ChatGPT had helped derive a new result in theoretical physics. Alex Lupsasca, one of the paper’s co-authors, stated that the model operated “at the level of a very talented contributor.” More telling were the words of Andrew Strominger, another co-author. “It is the first time I’ve seen AI solve a problem in my kind of theoretical physics that might not have been solvable by humans,” he said. “Two things changed: the model improved and we figured out how to talk to it.”
ChatGPT’s impressive contribution to this work—a deft and delicate calculation—required persistence and thoughtful dialogue from world-class physicists. The value of humanity, however, involved not only extracting the result produced by AI but interpreting it: situating the outputs of the machine into an underlying theory and mapping its significance to the physical world.
In scanning the previous paragraph, two points come to mind. First is the emphasis of human value in what was heralded as an achievement for AI—a reflection of the capabilities of the technology, I suspect, but, more pertinently, our need to matter amidst its shadows. Second is the thought experiment underlying this entire piece: given the first 2,000 words of this article, could ChatGPT have devised the previous three paragraphs if left to its own devices?
Perhaps one day it will, and future models will appreciate the allure of an emphatic finale, fuelled by the invocation of theoretical physics and references to the text itself. Until then, it would seem that what distinguishes our faculties from the machine is the grounded awareness of what it’s like to be human and the meaning we attach to that which intelligence produces. Technology may continue to dazzle, but our capacity to think will remain something to both cherish and preserve.
Sheldon Fernandez is the former CEO of DarwinAI and an AI strategist.
