Long before Artificial Intelligence, in Apollonius Rhodius’ Argonautica, Talos was gifted to Europa by Zeus to guard the island city of Crete. His body was entirely made of bronze, which made him indestructible and invulnerable. He circled the island three times a day, preventing approaching ships from reaching the shore. Except he had one vulnerability. Beneath his ankle was a single blood-red vein connected to his life and death. Only Medea, the sorceress who recognized the vulnerability hidden within the impenetrable body of Talos, could subdue him. She cast a bewitching spell, causing him to lose control and injure his ankle against a sharp rock while attempting to stop the Argonauts. Like melted lead, the hidden essence of his life flowed out, and he fell to the ground.
The subdual of Talos offers a striking metaphor of the power the mind wields over matter. Thousands of years later, machines are no longer imagined as bronze guardians, but are being trained to perform tasks once considered uniquely cognitive and human. This article argues that as generative AI makes the production of information abundant, the value of education may have to shift from simply acquiring information to developing the ability to question and independently reason about it.
When answers become abundant
Before the mainstream use of generative AI like ChatGPT in our academic and professional lives, students and professionals used simple web search to seek information. Such a person would type relevant keywords into the search bar, scroll through the results, decide which source is relevant, and evaluate its sufficiency. If its sufficiency was deemed inadequate, they would look for additional sources, which would lead them to connect different pieces together and eventually construct an answer. A study published by Fisher, Goddu, and Keil in 2015 examined how access to the internet affects people’s perception of what they know. Across nine experiments, participants who searched for explanations subsequently rated their own explanatory knowledge more highly than those who had been given the same information without searching for it. Almost a decade later, in 2024, Stadler et al. examined the effects of using an LLM compared with a conventional search engine during information gathering. Their experiment with 91 students compared research using Google and ChatGPT. It was observed that students who used ChatGPT reported lower cognitive load, while those who used Google demonstrated better performance at producing arguments and logical reasoning.
It was observed that students who used ChatGPT reported lower cognitive load, while those who used Google demonstrated better performance at producing arguments and logical reasoning.
The previous study noted that digital technology had already made information easier to access, but the student or the user still had to search, select, evaluate, interpret, and assemble information. One can argue that this process still required some intellectual work. Large language models (LLMs) are trained using vast quantities of text to generate new content. They can be used to generate, summarize, and translate text and essentially respond to user prompts. With the surge in usage of LLMs like ChatGPT, students have unprecedentedly easy access to generated information and explanations. This automation can also increasingly lead to the delegation of the intellectual process through which an answer is constructed. The result is the disjunction created between getting an answer and learning how to arrive at an answer.
The student who no longer has to struggle
The usage of GenAI can make task completion more efficient, but it is an entirely different question whether it improves learning. This distinction is important because the result is not uniformly negative. The OECD’s Digital Education Outlook 2026 has explicitly reported that GenAI can improve learning when it is designed or prompted to function as a scaffold. On the other hand, if it is prompted to behave as a direct-answer system, it will perform tasks efficiently but without providing equivalent learning gains. The study by Stadler et al. serves as a concrete illustration of this distinction. The ChatGPT group experienced lower cognitive load, but their final recommendations contained lower-quality reasoning and argumentation than the other group. It is important to clarify that the study did not find that ChatGPT simply made students incapable of thinking or that LLMs harm learning. It instead revealed the trade-off within this information-gathering task, where lower cognitive load was reported to accompany weaker argumentation. This raises the question of when delegating a task begins to compromise the intellectual capacity that task is meant to exercise.
Thus, the issue is not necessarily how much AI a student uses, but which part of the intellectual process the student retains.
The answer is that cognitive offloading is not a new or necessarily harmful behavior. Humans have always used external aids such as writing, gestures, other people, and technological tools to reduce the cognitive demands of a task. The important question is which cognitive activity is exactly being offloaded. The 2026 Poquet et al. study makes this distinction rather than classifying “AI use” as a single behavior. Their lab study involved college students, who were asked to synthesize two opposing texts on predictive modelling in education into an essay using ChatGPT. The researchers analyzed 564 prompts and classified them according to whether students remained active or passive and whether they were dealing with local details or the broader structure of the text. The study found that “Full Offloaders” produced essays with only 61% student authorship, while “Strategic Offloaders” had 24% and “Clarifiers” 91%. This demonstrates that students might actually tell the LLM what they think and ask it to evaluate or refine their reasoning. Alternatively, a student might just ask it to summarize or write the answer without specifying their own reasoning. As a result, the researchers distinguished the broad usage of AI from division of labor, where only parts of the cognitive process are delegated to AI. Thus, the issue is not necessarily how much AI a student uses, but which part of the intellectual process the student retains.
If some forms of learning depend upon the effort involved in struggling and making mistakes, what happens when technology becomes capable of removing much of that friction?
This issue can further be complicated when applied to a discipline like mathematics because it considers how GenAI can aid in mathematical reasoning as well as practical performance. In 2025, Bastani et al. studied nearly 1000 high school mathematics students. Students who had access to a standard GPT-based system performed 48% better during practice. But when access to AI was removed, they performed 17% worse than students who had no access to it. The study describes it as the “crutch effect” because while the tool improved performance when the students had access, it did not produce equivalent independent capacity. The same experiment also compared the unrestricted “GPT base” system with a deliberately designed GPT tutor that was configured to avoid giving direct answers and support learning. It was observed that the tutor version mitigated the negative effect while also producing larger gains during assisted practice. This connects to the idea of “productive struggle”. Effective learning can often require effortful engagement with difficult material, which can contribute to better long-term learning. A correct mathematical answer may demonstrate that the student can obtain the correct answer. But can the learning process be equated to being objectively right? If some forms of learning depend upon the effort involved in struggling and making mistakes, what happens when technology becomes capable of removing much of that friction?
The value of being wrong
The value of education may not lie only in reaching the correct answer but in developing the capacity to recognize when an answer is wrong and determine how to improve the cognitive process behind it. In 1915, Roy Ivan Johnson argued that when teachers simply identify an error and provide its correction, the student’s initiative is weakened because the intellectual work of finding and correcting the mistake has been removed. In contrast, it develops critical judgement and habits of accuracy if students identify and work on errors themselves. In another work, Swartz argues that mistakes are an inevitable feature of human inquiry because human knowledge is potentially fallible. Identifying and eliminating mistakes helps move inquiry towards better solutions because knowing what not to do frees us to reconsider established ideas and explore alternatives. He therefore describes problem solving as a dynamic process, where contemporary knowledge can be changed and improved rather than treated as final. This reveals how humans do not merely make errors; they can instead learn from, expose, critique, and revise them. In the same study, Swartz also argues that openly revealing decisions and mistakes allows others to identify errors that an individual may miss. This essentially means one person’s mistake can prevent others from repeating it.
Along the same lines, DeBrincat termed productive student errors “springboards of inquiry” because an incorrect answer can take classroom discussion down an unforeseen but productive path. This gives students greater agency and opens up alternative perspectives. However, neither Swartz nor DeBrincat advocates careless error. Swartz further stresses that mistakes must be made within appropriate problem situations, because some errors can have serious or irreversible consequences. The educational value lies in creating an environment where mistakes can be examined, exposed, and revised.
Human error has the potential to become productive because humans can recognize an error or question the reasoning behind it. AI systems do not possess this same reflective relationship with their own errors.
It is pertinent, however, to remember that AI itself is also fallible, but its errors are produced differently from human errors. Barassi argues that AI failures emerge from the ways systems combine patterns and information through mathematical and statistical reasoning. Human error has the potential to become productive because humans can recognize an error or question the reasoning behind it. AI systems do not possess this same reflective relationship with their own errors. Its outputs still require human judgement and scrutiny. She argues that AI systems are shaped by the limitations and inequalities embedded in their training data and design. This means that an answer produced by AI can reproduce existing gaps or biases in knowledge rather than providing a neutral correction to human error. Barassi further discusses how 93% of GPT-3’s training data was in English by word count, alongside the problem that multilingual models can perform unevenly across languages and may rely on machine-translated material that does not properly capture local linguistic contexts. This makes the capacity to question an answer, even an AI-generated one, more important and not less. If education removes the process through which students encounter and correct errors, it risks producing people who can obtain answers efficiently but are less capable of judging when those answers should be challenged.
What should remain human?
Generative AI is not inherently harmful to education. Its ability to make information accessible, reduce cognitive burden, and provide assistance can genuinely improve learning when used as a scaffold rather than a substitute for thinking. The central problem is not that AI provides answers, but that it can make the process of arriving at an answer optional. This matters because some intellectual capacities develop precisely through struggle, uncertainty, error, and revision. If students consistently remove these stages through AI, they may become efficient at producing outputs without developing the capacity to interrogate those outputs. Therefore, the appropriate response is not rejection or prohibition of AI, but a better approach to using it, particularly among students. AI should assist in explanation, questioning, feedback, and exploration, while students retain the responsibility for reasoning and verification.
Medea did not defeat Talos through brute force. She identified the single vulnerability within an apparently invulnerable machine. This essay is not a debate of AI versus human. AI’s growing power similarly makes it important to identify what must remain human. Education’s purpose is not merely to produce people who can reach the right answer, but people who can recognize when an answer is wrong, ask why, and decide what should replace it. This is not witchcraft, but perhaps it is how humans should proceed.
