Introduction
In April 2026, Enrique Dans published an essay titled "Thinking hurts" (originally Pensar cansa). Its claim was simple and unsettling: much of the time, we are beginning to use AI not to think better but to avoid having to think at all. Dans traced this to a tendency some researchers call "cognitive surrender" (rendición cognitiva); we approve an answer delivered in a well-built, fluent, self-assured style, copy it, and go on our way. Yet we withhold from AI-generated writing the scrutiny we would apply had the same answer come from a person.
What said more than the essay itself were the reactions that recurred in the comments it drew and in similar online debates.
Faced with the same phenomenon, readers were flung to diametrically opposed positions, and this scattering repeated itself in the comments on that essay as much as in similar online debates. One commenter argued that it is wrong to lay responsibility on the tool: the calculator, the book, the pen, and AI alike are neutral tools; what truly decides is the attitude of the person behind them. Whoever wants to learn learns even with an encyclopedia; whoever does not will tend not to learn even with the most advanced tool. Dans's own position stressed the opposite: the tool is not designed to be neutral; it is designed to reduce friction, not to generate doubt. Another commenter drew the distinction elsewhere: AI can process a codified, debatable body of rules with fluency, but it cannot reach, in the same way, a morality rooted in the body and in lived experience, accumulated across a lifetime; moreover, because it voices both the rule it can reach and the morality it cannot with the same fluency, it does not easily let the reader sense the difference. A fourth objected to all of them: we are not ceasing to think, we are beginning to think differently.
The real question here is not which of them is right. Before it comes another: why do people looking at the same object arrive at such different places? And one more: what do these reactions actually describe? AI, or the position of the one looking at it? What the tool was designed for, or what it turns into, and in whom?
The starting observation of this essay is this: each of the comments gives away where its owner stands. The one who deems the tool neutral and ties responsibility to the individual's will is speaking from a particular worldview. The one who stresses the embodiment of morality is laying out their own professional or philosophical ground. The one who says AI merely affords a "different" way of thinking is defending their own way of using it. AI works less as the object onto which the reaction falls than as a surface that reflects the reacting person's position back.
This is the claim the essay will advance: the reaction to AI is, more often than not, not about AI but about one's own position. AI performs this reflection in three modes, and the essay is built around these three.
The first is the mirror. A person sees and defends their own position in AI (their competence, their privilege, their virtue); and the content of the reaction shows what it is they are defending.
The second is the magnifier. Whoever tries to cover their inadequacy with AI cannot cover it; because, not knowing the subject well enough, they cannot audit the tool's fluent and persuasive output, cannot get inside it. So the inadequacy, instead of being hidden, grows and comes to light. The same magnifier also works in the other direction: place accumulation, material, a well-formed question beneath it, and it multiplies capacity. The lens is neutral; what separates is what is placed beneath it and the hand that holds it, that is, the accumulation and the orientation with which a person looks at the tool.
The third is litmus. (Litmus paper is a measuring tool that gives away, by changing color, whether the liquid it is dipped into is an acid or a base.) Good work with AI is itself a skill, and this skill is largely uncodifiable; it matures through living and through mentorship. For precisely this reason AI does not equalize but separates: when two people take up the same tool, one cannot do what the other does. This runs counter to the discourse that "AI equalizes everyone and opens expertise to all," and it puts that discourse on the spot, making it strongly testable.
A fourth layer accompanies these three lenses: the shield. The observation that "AI makes things up" is technically correct, but this finding can turn into an excuse that shifts responsibility from the agent to the tool. The one who made the error may slip into protecting their own position by consigning the error to the nature of the tool.
Everything up to here aims a lens at others. The essay's real backbone is in turning that lens on itself: the lens is two-way. The side that wields it too, the side defending AI, the one who says to their counterpart "you are afraid, you are resisting for the privilege you stand to lose," can use the same lens to confirm their own position. Everyone sees the other in the mirror and reads their motive; the hardest thing is to see oneself. If this essay does not put its own position (the side that uses AI, that benefits from it) to the same test, it becomes an instance of the very thing it diagnoses. For this reason the lens will be set up here not as a weapon of diagnosis but as a reading under test.
Finally, a statement of limits: this essay does not claim to have discovered a new mechanism. It offers a reading that brings together, at the level of public discourse, studies that have drifted apart from one another (most of them confined to a workplace or laboratory context): identity threat, tacit knowledge, moral grandstanding, the responsibility gap. Originality lies not in the individual findings but in the frame that can connect them, and in the discipline of applying that frame to oneself as well. Throughout the essay the forum commenters cited are treated not as persons but as positions, and none is labeled; because the matter is not who anyone is but what the reactions show; our aim is not to be proved right. Dans closes his essay this way: "that choice is not about understanding technology, it's about challenging our cultural instincts" (in the Spanish original, "the line is not in the technology but in the attitude"). The one condition this essay adds is that this attitude be tested in both directions.
I. Mirror: who is the reaction about?
To unpack the reaction to AI, a concept from research in information systems may help: "identity threat." Relaying a definition from Craig et al., Mirbabaie et al. express it as "the expectation that the use of an information technology artifact will harm an individual's self-beliefs" (Mirbabaie et al., 2021, p.75). The mechanism is plain: when an experience conflicts with a person's identity, the individual suffers a loss of self-esteem and acts to protect the self-esteem tied to that identity; this action often takes the form of resistance (ibid., p.77).
What is critical is the object of the threat. What is threatened is, more often than not, not a person's job or income but the answer they give to the question "who am I." Mirbabaie et al.'s workplace study finds three predictors of AI-driven identity threat: change to the work, loss of status position, and the identity a person builds with AI (Mirbabaie et al., 2021, p.73). All three concern position more than material loss.
Bankins et al. locate the same threat at the group level. For them, professional identity is a professional group's shared answer to the questions "who are we; where are the limits of our expertise, what are our values" (Bankins et al., 2024, p.164). If this shared answer is strong, it absorbs the tool: professions with a deep, well-defined identity try new technologies more readily, because strong shared assumptions make the risk of change acceptable. But if the identity is weak or uncertain, the same tool is met as a threat.
From here comes a counterintuitive finding that leads to this essay's core observation. Those who resist most are not the least competent. The studies Bankins et al. compile show that people with high domain expertise resist algorithmic recommendation more: both because they will themselves answer for the result AI produces (they own the output) and so do not trust it blindly but want to audit it; and because they judge their own abilities superior to the tool's and believe using it is of little benefit (Allen & Choudhury, 2022, as cited in Bankins et al., 2024, p.164). It is not inadequacy but the perception of expertise that heightens resistance.
Here is the core of the mirror. Resistance reports less on a feature of AI than on the position of the one resisting: on the expertise they are presumed to hold and on the privilege that expertise confers. The intensity of the reaction measures the subjective value of what is being protected. A tool's rejection as "it can't do what I do" arises, more often than not, less from the tool's limit than from the meaning the speaker attaches to their own skill.
This literature was born in the workplace, but the pattern appears in public debate too. When, beneath a technology piece, some call the same tool "the thing that finally lightens the burden of thinking" while others call it "a threat that makes people shallow," this divergence arises, more often than not, not from what the tool does but from where each side places its own labor and identity. A recurring observation in such debates shows that we find AI wonderful when it does not err and declare it "stupid anyway" when it does; our appraisal of the tool is often a mirror of our own expectation and of whether that expectation was met.
One last distinction, because this section is the easiest to misread. That the content of a reaction gives away a person's position does not mean labeling that reaction "a complex," "envy," or "fear." Such a label would judge the reacting person's argument by its motive rather than its content, which is a separate kind of error, taken up in section V. The claim here is more measured: the reaction to AI carries information less about an objective feature of the tool than about the reacting person's own position toward AI. The mirror does not tell you what a person is; it shows you where they are looking.
II. Magnifier: the attempt to conceal exposes
The mirror said that the content of a reaction gives away a position. The magnifier goes one step further: it makes visible not only the reaction but the work itself. The core of the metaphor is its neutrality. The magnifier multiplies whatever is placed beneath it; place inadequacy, it multiplies inadequacy; place accumulation, it multiplies accumulation. What separates them is not the lens itself but what is placed beneath it and how the hand holding it uses it. This section follows one direction of the lens, how the attempt to conceal turns into its opposite; the other direction, which multiplies capacity, will connect later to the litmus section through the question of why the same tool yields different results in two different hands.
The mechanism seems plain but knots at a counterintuitive point. AI's output comes in a fluent, orderly, self-assured style. These three qualities are independent of the content's accuracy, yet people often read them as a sign of accuracy. Buçinca et al.'s experiment measured exactly this point: adding a justification to AI's decision raises the user's trust in it. Moreover, this trust rises even when AI is wrong; overreliance is born right here (Buçinca et al., 2021, p.2). What researchers call "overreliance" acquires a technical definition here: the user's tendency to follow AI's recommendation even when it is wrong or unsuitable (ibid.). Passi and Vorvoreanu's institutional review, surveying some sixty studies, sees the same pattern more broadly: explanations raise trust not only in correct but even in wrong recommendations; automation bias and complacency drive the user into error (Passi and Vorvoreanu, 2022).
This tendency turns into measurable behavior. Shaw and Nave's three preregistered experiments found that participants, when they consulted AI and AI was wrong, adopted the wrong advice at a high rate (in the first study, 79.8% of the erroneous trials on which AI was consulted; 74.6% in the second, and 72.3% in the third, under time pressure; Shaw and Nave, 2026, pp.22, 28). The authors stress a condition: this rate is tied to choosing to consult AI on that question, because the consultation rate itself is around fifty percent. That is, not everyone surrenders on every question; but when they do, the error is largely taken over. The net result is a drop in accuracy: when AI is right, participants gain about 25 points; when it is wrong, they lose 15 (Cohen's h = 0.81; ibid., p.22). The name the authors give this pattern coincides with the term Enrique Dans also uses: cognitive surrender.
An observation repeated in everyday language on technology forums describes the same mechanism: when you delegate all your thinking to asking AI questions, the mind goes offline; yet asking a question is not thinking, the real thinking is in building the answer. This is the everyday description of a loosening of self-scrutiny. A person is spared the labor of producing their own answer, but that labor was at the same time the chance to test the answer.
Up to here prepares why the attempt to conceal fails. A distinction is needed: over-trusting fluent output is a widespread tendency, while lacking the knowledge to audit it is a separate situation; one is negligence, the other impossibility. When a person trying to cover their inadequacy with AI takes the fluent output as is, because they lack the knowledge to audit it, the inadequacy is not covered but grows. Because the covered gap enters the output, and when the output becomes public the gap becomes visible along with it. The clearest example came from law: in a widely reported case a lawyer faced sanctions for putting AI-fabricated fictitious cases into a brief, and others followed (as relayed in Magesh et al., 2024, p.2). Magesh et al.'s real contribution is not to note this incident but to measure how often AI tools designed specifically for law fabricate: dedicated legal tools produce hallucinations in 17% to 33% of queries; while the best-performing tool answers 65% of queries correctly, another stays at 42%, fabricating nearly twice as often (ibid., p.1). The tool's fluency is constant, its reliability is not; a user who does not distinguish the two documents the very gap they wanted to cover.
Why does this cycle not correct itself? Kruger and Dunning's classic finding fits here: inadequacy diminishes two skills at once, both the skill of doing the work and the skill of noticing that one is doing it badly. Those in the bottom quartile, in reality at the 12th percentile, judged themselves at the 62nd (Kruger and Dunning, 1999, p.1121). Metacognitive blindness causes precisely the one most in need of covering this to be the one who least sees that the cover does not hide it. The magnifier is merciless for this reason: the gap it enlarges most is the gap its owner is least aware of. The long-term cost is measured too. Gerlich's study with 666 participants finds a strong negative relationship between frequent AI use and critical thinking skill (r = −0.68), and an even stronger negative relationship between cognitive offloading and critical thinking (r = −0.75). In multiple regression, AI use negatively absorbed critical thinking (β = −1.76, p < .001; Gerlich, 2025a, pp.14, 16). This is a correlational picture; on its own it establishes no causation, but it shows an accumulation in the direction the magnifier metaphor expects. Two further studies strengthen this picture. The first shows the pattern is not confined to students or the laboratory: Lee et al. asked 319 knowledge workers about 936 real instances of using AI at work and analyzed each one. The finding turns the magnifier's "hand that holds the lens" intuition into a measured coefficient. As trust in AI rises, the engagement of critical thinking falls (β = −0.69, p < .001); by contrast, a person's confidence in their own task (β = 0.26) and confidence in evaluating AI's answer (β = 0.31) raise critical thinking, and the strongest positive effect is the disposition to reflect (β = 0.52, p < .001; Lee et al., 2025). What is decisive is not the presence of trust but its direction: trust that surrenders to the tool lowers thinking, while trust turned toward oneself and toward scrutiny raises it. The second partly settles the causal debt. The same author's experimental study compared four conditions (human only, human plus unguided AI, human plus structured-prompting AI, AI only) with 150 participants across three countries (Germany, Switzerland, the United Kingdom): unguided use produced cognitive offloading without raising reasoning quality, while structured prompting significantly raised both critical reasoning and reflective engagement (Gerlich, 2025b). The difference was not in the tool itself but in the manner of use; the direction arrow correlation could not draw appears where the experimental arm points.
Now the lens must be turned over, because the claim of neutrality holds only if its opposite is shown too. When beneath the same fluent output there is not a gap but accumulation, material, a well-formed question, the lens magnifies capacity. In the same debates the opposite experience is voiced too: an experienced user recounts using AI as a cognitive amplifier, learning far faster with it, yet depending on it for nothing. The difference between the two poles is not in the tool itself but in the material placed beneath it and in the hand that holds the lens. The same criticism is aimed at the design of the experiment: what such studies show is not that AI rots the brain but that it becomes a measuring instrument, catching people at the most incompetent they can be.
This double face also recalls that what the tool reduces is not always a gain. One of the things AI reduces most is friction: the labor of continual trial and error while stuck before a problem. Yet Kapur's "productive failure" study shows that students' unaided struggle with a hard problem, though it looks like failure in the short term, produces at a later stage a depth that surpasses AI-assisted learners (Kapur, 2008). The same distinction has a new empirical counterpart: in Fan et al.'s randomized experiment with 117 university students, the ChatGPT group led on the short-term test score but showed no significant difference in knowledge gain and transfer; this pattern, which the authors call "metacognitive laziness," is the score rising on the surface while something in the depth does not change (Fan et al., 2025). A use that erases friction entirely may erase this hidden productivity too; for which the hand that holds the lens must distinguish which friction is a burden to be shed and which an opportunity to be kept.
The distinction here should not be confused with a neighboring concept in the literature. Ehsan et al. say AI carries an "AI-as-Amplifier Paradox": the tool, while strengthening performance, erodes the underlying expertise, at once both amplifying and dissolving (Ehsan et al., 2026, p.1). The axis of that paradox is between tool and expertise, between amplification and erosion. The magnifier this essay uses points to another direction: the lens magnifying a person's own position, the thing they place beneath it. The two do not conflict; they work on different planes; Ehsan's paradox feeds this section's negative direction (the erosion of competence), but the question of how much the lens magnifies the input is absent from their frame. In fact the two, Ehsan's amplifier and this essay's magnifier, can be read as the two ends of one chain. Ehsan's amplifier works along the axis of time: the expertise of a person who uses the tool constantly erodes over the long term, often unnoticed. This essay's magnifier works along the axis of the moment: the inadequacy or accumulation a person brings to the tool at that moment becomes visible in the output immediately. The two do not conflict, they are sequential; the person who covers their inadequacy with the tool is exposed in the short term, and as the same covering continues, loses their real skill too. Exposure and erosion are the short- and long-term faces of the same dependence.
Thus the magnifier renders no judgment on its own, it sharpens a question: what is beneath the lens and what is the hand holding it doing? This question leads to why the same tool yields such different results in two people's hands, that is, to the "litmus" section.
III. Litmus: AI does not equalize, it separates
Litmus paper changes color when dipped into a solution; the color it changes to gives away what the solution is. This is the metaphor's work: when AI comes into two people's hands in the same way, it does not take them to the same place, it makes the difference between them visible. This section's claim names where that difference comes from. Good work with AI is itself a skill, and this skill is largely uncodifiable, nor does it change hands easily; it is fully present neither in the instrument itself nor in a manual, and becomes reachable through living and through mentorship. For exactly this reason AI does not equalize but separates.
Because this claim stands against a very widespread discourse, that discourse must be confronted in its strongest sources. The thesis that "AI equalizes everyone and opens expertise to all" is not a slogan but a thesis with serious field evidence. In Noy and Zhang's preregistered experiment, when university-educated professionals were given writing tasks and half were equipped with access to AI, average productivity rose markedly: time spent fell by 0.8 standard deviations, output quality rose by 0.4 standard deviations, and inequality among workers also fell, because the tool benefits the lowest-skilled most (Noy and Zhang, 2023). Brynjolfsson, Li, and Raymond's study with 5,172 customer-service agents points the same way: access to AI raises productivity by 15% on average, but the gain is not evenly distributed, reaching 36% in the lowest-skill tier, highest among the least experienced agents (Brynjolfsson et al., 2025, pp.889, 911). Dell'Acqua et al.'s field experiment with consultants also confirms the equalizing direction: those performing below average improved by 43%, those above by 17%, so the gap narrows (Dell'Acqua et al., 2023). This thesis has a voice in public discourse too. A common objection in technology debates turns to the experts of a field and says: I do not know and do not care about anything beyond the basic definition of this technical field; what I care about is what I can do with it; insistence on expertise often turns into an intellectual arrogance. This voice, saying the barrier has fallen, is describing a real experience.
But in all three field studies there is a limit appended beside the equalizing finding; the litmus thesis begins right at that limit. Dell'Acqua's experiment used two tasks: one inside AI's capabilities, the other outside. Equalization happened on the first task. On the second, that is, the task falling outside AI's frontier, consultants who used AI were 19 points less likely to reach the correct solution than those who did not (Dell'Acqua et al., 2023). Brynjolfsson's data shows another face of the same limit: while the least experienced workers gained most, the most experienced and highest-skilled workers saw a small gain in speed and a small drop in quality, and the gain concentrated most in routine, repetitive questions (Brynjolfsson et al., 2025, p.889). The pattern is consistent: equalization occurs in the codifiable part of the work, the part inside AI's frontier. Outside the frontier, that is, in the uncodifiable part of the work, the tool does not equalize; on the contrary, it misleads whoever trusts it and carries it beyond its frontier, and lowers their performance. The two findings do not conflict, they measure different layers; one the inside of the frontier, the other the outside. The same divergence appears in a controlled experiment too. The four-arm study noted in the magnifier section is not only a piece of causal evidence but also the cleanest laboratory counterpart of litmus: the same tool, in the hand of one who used it with structured prompting, fed critical reasoning, while in the hand of one who used it unguided produced only offloading (Gerlich, 2025b). A controlled setting also confirms the divergence the field studies point to.
So what lies outside the frontier? Uncodifiable knowledge. Hadjimichael, Ribeiro, and Tsoukas's study, reaching from Polanyi to Merleau-Ponty, explains how tacit knowledge is acquired: through the body's ability to grip the task at hand, to find the grasp best suited to it. This ability, in the authors' words, develops "through the authoritative guidance of the more experienced" (Hadjimichael et al., 2024). A point the authors underline is the thesis's strongest leg: no task, whether manual or intellectual, whether ordinary or creative, can be completed without recourse to tacit knowledge (ibid.). That is, uncodifiable skill is not a marginal exception but at the core of every task. Lave and Wenger open this dimension of guidance further: for them learning is not an isolated mental act but happens through participation in a community of practice; the novice becomes a master alongside masters, moving from the periphery of the work toward its center, by doing together (Lave and Wenger, 1991). What carries the skill is not a manual or a group lecture but this relationship of participation; in their words, "even so-called general knowledge only has power in specific circumstances" (ibid., p.33). Tsao et al.'s review surveying creative professions sets this distinction on two axes: the codifiability of knowledge (its passage from tacit and embodied to explicit and codifiable) and the materiality of the output; it finds that fields which prioritize embodied practice are the ones most resistant to AI substitution (Tsao et al., 2026). The same review flags career stage as a decisive factor: entry-level practitioners meet AI with enthusiasm, as a natural extension of digital tools, while senior practitioners view the devaluation of expertise with suspicion (ibid.).
The individual difference itself is measured too. Lovett's review relays that people with similar backgrounds reach sharply different results with the same tool, and that the distinction is drawn by learning-goal orientation: a performance-oriented dependence turns the tool into skill decline, while a mastery orientation turns the same tool into skill growth (Yang et al. 2026 finding, as relayed in Lovett 2026). The phrase "the same tool" is the core of litmus: what varies is not the tool but the orientation that holds it.
The clearest public expression of this distinction lies between two voices that meet head-on in the debates. One describes their own method: having AI produce many options and then filtering them through their own knowledge base, and adding that this cannot be done without prior accumulation. The opposing voice is the other side of the coin: one without accumulation will also filter, but will believe they are filtering while choosing without quality, and moreover will never know they made a mistake. The same tool: the expert filters, the novice thinks they filter. The same distinction is posed from the angle of learning too: the way someone who has learned the fundamentals beforehand uses the tool differs from what someone starting from scratch will experience; learning an advanced subject while a tool that produces ready solutions is at hand can cripple long-term learning. Here litmus changes color: the paper dipped into the same solution shows what the one dipping it brought.
A confusion must be prevented here. Litmus's "it separates" claim may seem to conflict with the observation voiced in some debates that "AI makes everyone alike." The phenomenon Sarkar calls "mechanized convergence" is real: AI users tend to produce less varied output on the same task (Sarkar 2023, as cited in Lee et al., 2025). But this is outputs resembling one another; what litmus separates is the skill difference between people. The two stand on separate planes: while outputs converge on the surface, the capacity of those producing them to filter, audit, and integrate continues to diverge. Moreover Lee's own caveat is valuable here: output variety is a flawed measure of critical thinking; someone who uses the tool's suggestion without changing it may perfectly well have made a critical judgment. Convergence is not the opposite of divergence but a separate phenomenon standing on a different axis.
One more debt of honesty: the optimism this section implies, the expectation that "in a good hand the tool magnifies capacity," is not beyond question either. An agent-based simulation proposing four criteria to distinguish cognitive amplification from cognitive delegation suggests that no real amplification was reached in any interaction regime tested, that even in the mixed regime that best preserves human capacity the collaboration gain stayed negative, and that a positive gain did not appear even when cognitive erosion was reduced to zero (Di Santi, 2026). This finding has two limits, and both should be stated: first, it is a NetLogo simulation, not empirical field data, and the author himself characterizes it as an abstraction; second, its definition of "amplification" is whether the hybrid output surpasses the best single agent (in most cases AI itself), which is not exactly the same thing as "the growth of a person's own capacity" this section speaks of. Still it stands as a limited warning: the presence of a hand that draws production from the tool does not mean that production turns, in every case, into a real cognitive gain. This too is a lens the essay turns on its own optimistic half.
Finally, a debt of honesty. To say "uncodifiable" is not to say "unteachable"; this essay does not take on that absolute. Hadjimichael's emphasis says the very opposite: tacit skill matures through the guidance of the experienced. That is, skill is not an intransmissible intuition but transmissible, only in the company of the experienced and by living it with them. What litmus separates is not an innate hierarchy of talent but whether accumulation and orientation are present or not. To turn this distinction into a declaration of superiority, to say "we who can filter, they who cannot," offers a chance to read the lens this section builds in one's own favor. That temptation is real and binds the essay itself; the next section takes up exactly this trap.
IV. Shield: shifting responsibility from agent to tool
All three lenses looked in one direction: at a person's own position. The shield must be added beside them as a fourth movement, because it is not a lens but a means of escaping the lens. The sentence "AI makes things up" is technically correct; the second section showed this with measured data. But a correct observation can be put to a wrong use. That sentence often ceases to be a description and turns into an excuse: the owner of the error, by consigning the error to the nature of the tool, frees themselves from the obligation to audit and from ownership of the result. When the observation turns into a shield, what it protects is not accuracy but position.
The logical error is simple but invisible. The premise "the tool can make things up" is true; the inference "therefore the error in the output is not mine" is invalid, because there is a step that breaks the link between them: the agent who adopts, uses, and makes the output public. To make this step visible, one must ask to whom responsibility can belong.
Fuchs's study closes this end decisively: the tool cannot bear responsibility, because it is not a moral agent. AI has no body and no aliveness; though its power of expression stirs an involuntary empathy in us, in Fuchs's phrasing, when an object is a structure that can express well enough and resembles the living, we must not be deceived by this empathy into being drawn to it, because there is no one there who will be happy, be saddened, or answer for anything (Fuchs, 2022). Fuchs's real warning turns into a design principle: the systematic imitation of subjectivity must be prevented; the system must present itself only as what it is: a simulation (ibid.). From here the rotten spot of the shield comes to light. To pin responsibility on AI means ascribing to it an attribute it does not have, the attribute of "being an agent." Because the tool cannot be an agent, responsibility cannot remain with it; it inevitably returns to the human.
Yet a question stays open: what if no one can really foresee? Matthias shows that in learning systems there is a genuine responsibility gap: when a system learns from experience and changes its behavior while running, its manufacturer or programmer can no longer fully foresee and control its future behavior; traditional attribution of responsibility, which requires the agent to know and be able to control the action, meets a gap here (Matthias, 2004). But where this gap opens is decisive. Matthias's gap is in the loss of foresight of the manufacturer, the one who designs the system; not on the side of the user-agent who takes, adopts, and makes the output public. The shield blurs exactly this distinction: it uses the manufacturer's unforeseeability as the excuse of the user who owns the output. Yet the manufacturer's inability to foresee does not remove the user's responsibility to audit and to own.
The gap Matthias points to is only one example. Santoni de Sio and Mecacci define four distinct responsibility gaps concerning AI; the first is the culpability gap: situations in which a person leads to a bad outcome yet can hold a legitimate excuse, which no one could reasonably foresee or prevent (Santoni de Sio and Mecacci, 2021, p.1063). The concept is legitimate; but the authors' real warning is about how this gap is mishandled. They count three inadequate attitudes: "fatalism," which declares the problem new and unsolvable; "deflationism," which dismisses it as a false problem; and "technical solutionism," which presents everything as solvable by technical improvement (ibid.). The shield is the everyday guise of the first: to say "that's how the tool is, nothing to be done" is to present a foreseeable and preventable error as an unforeseeable fate. The way out the authors propose demands the opposite: sociotechnical systems must be designed for "meaningful human control," that is, must remain bound to human reasons and capacities (ibid.). Thus the agent remains human.
But to reject the shield is not to declare the tool neutral; this distinction is the finest point of the section. A strong pole in public debate says the design of the tool is not innocent. One voice of this pole says there are ways of using it that would enlarge the mind, but that the future companies will offer will not be this, because running a model that affirms the user will always be more profitable. Another reads corporate discourse itself as a shield: phrases of the kind "as long as you check the results" are expressions companies resort to in order not to be held responsible for possible problems. This pole points to a valid place. As Enrique Dans underlines in his own essay, the tool is designed to reduce friction, not to generate doubt; the fluent natural-language layer hides its limits. The tool is in this sense inclined. But there is a question of priority here: what is decisive, what the tool was designed for, or what it turns into, in whom? The design incline is real, yet it does not explain the outcome on its own, because the same inclined tool turns into production in one hand and into drift in another. Dans's observation that the tool is not neutral is not wrong; it only answers a different question, that of the tool's innocence. On the question of what determines the outcome, the weight shifts from design intent to the hand that holds the lens.
Two truths meet here and besiege the shield from both sides at once: the tool is inclined, the agent is still human. The incline does not erase the agent, it makes the agent's job harder. Design can push a user toward a fluent error; but whether to own that error or not, to audit it or not, is still the agent's act. The one who dumps all responsibility on the tool (the personal shield that does not see the non-neutral tool) and the one who dumps it on the user (the corporate shield that does not see the incline) stand at two ends of the same error: both want to place the burden in one place and make the other end invisible. Meaningful human control, in wanting responsibility to remain with the agent even with an inclined tool, therefore accepts both the tool's incline and the agent's debt of attention at once.
Thus the shield is, in fact, a fourth mirror. The person who says "AI made it up, the fault is its" gives information about their own position: not wanting to own the result, not taking on the labor of auditing, they want to place their error outside themselves. When we lift the shield, what is protected behind it becomes visible. But this observation opens at once onto a trap: to diagnose the same move in one's counterpart, to say "look, they are taking refuge in their shield," may itself be the diagnoser's own shield. Now the lens must turn back…
V. The two-way mirror and its traps (the self-reflexive lock)
The previous four sections aimed a lens outward: they showed how the reaction, the attempt to conceal, the skill, the excuse give away a person's position. But this reading itself carries a dangerous power: the same lens can, in that case, turn into a weapon in the hand of the one using it. This section turns the lens onto the one holding it. There are three traps; but all three hold not only for "the other side" but for this essay too.
The first trap is confusing origin with content. To deem a claim rotten because of the mood or interest of the one voicing it is known in logic as the genetic fallacy: discrediting a view because of its origin. Yet origin is, more often than not, irrelevant to the truth of the view (Dowden). That an identity threat lies behind someone's critique of AI does not, on its own, make that critique wrong. The mirror section flagged this from the start: showing the source of a reaction does not stand in for refuting the reaction's content. The sentence "you say this because you are afraid of losing your privilege" attacks the counterpart's motive, not their argument; the argument will continue to be true or false independently of the motive. If this essay takes the position it diagnoses to have the force of a refutation, it falls into exactly this fallacy. There is a way out of this trap too: a claim is first weighed on its own, by its content; only after that does the lens turn to what remains. Moreover the claim is built in its strongest, not its weakest form: the litmus section took up the equalizing thesis it opposed not as a slogan but with the soundest field evidence, because only a claim met in its strongest form is really tested. "AI makes things up" is first accepted as true; the lens then looks at the selective use of that truth, its heat, at which moment it is advanced. Reverse the order, stamp the claim with a motive from the start, and the diagnosis itself this time takes on the guise of a refutation. The previous four sections tried to hold the lens in this order: the shield first took the observation "AI makes things up" as true, and examined its turning into an excuse only afterward.
The second trap is using the lens in one's own favor. The litmus section drew a distinction: the one with accumulation and orientation turns the tool into skill; the one without chooses blindly and cannot notice that they erred. To turn this distinction into a hierarchy of persons, into a "we who grasp, they who are inept" narrative, would be the easiest and most insidious step, because it automatically places the claimant on the winning side. An extreme form of this narrative circulates in public discourse too: voices rise that draw as tableaux a future in which most people will keep dulling, some will adapt and live in a kind of "functional blindness," and some will develop an immunity to AI and turn into a separate "lineage." This is the caricature of the litmus thesis: it turns divergence into a kind of superiority, into the elite lineage the speaker belongs to. If the essay comes even a millimeter close to this language, it has refused to hold the lens it built to its own face. For this reason the litmus section particularly emphasized that the distinction is not an innate talent but an acquirable accumulation and orientation.
The third trap is reducing the argument to psychology. The mirror and shield sections said that the reaction and the excuse give away a position. To move from here to reading every opposing view as the symptom of a psychological flaw, that is, to say "really they are bothered by this," "really they are compensating for that," shifts the debate from content toward the person. This essay does not diagnose persons, and must not; the forum commenters are taken as positions, not as cases. The lens is an analytic tool that makes positions visible, not a diagnostic apparatus that labels persons.
These three traps do not bind only the one who defends AI; they bind the one who resists it, even the one who panics before it. The lens must be turned to that side too. Take the most-shared finding of late, that "AI is rotting our brains." Kosmyna et al. connected essay-writing participants to EEG and proposed that those using AI accumulate "cognitive debt" over time, that their neural connectivity, sense of ownership, and recall performance decline; the result quickly turned in public into the headline "AI is ending learning" (Kosmyna et al., 2025). Yet a detailed commentary written on the same study shakes the finding in several places. Stanković et al. show that the sample is small relative to the number of comparisons made (that where a power analysis requires about 159 participants, the work was done with 54), that the EEG measure of "meaningful connectivity" shows a relative difference between groups and so the inference "less connectivity, weaker brain" cannot speak about absolute neural power, that in the beta band the AI group even came out stronger than the tool-free group, and that on the "quoting" task the difference between the search-engine and tool-free groups was insignificant (p = 1). This last weakens the thesis that "delegating to an external tool accumulates cognitive debt" in itself: the search-engine group also used an external tool yet showed no deterioration (Stanković et al., 2025). The point here is not that Kosmyna is wrong; it is how quickly and selectively a not-yet-peer-reviewed preprint can be read because it confirms the panic narrative. Just as AI's making things up is, so AI's rotting the brain is a technically testable claim; panic too is a position and can read its own evidence selectively. This is symmetric with the move in the shield section and binds this essay too: just as it took up the equalizing thesis in its strongest form, it must read the opposite pole's finding with caution as well.
The same balance appears in the literature as a whole. A review surveying thirty empirical studies reports that twenty-one of them found AI to have a predominantly positive effect on critical thinking, and that this effect appears especially in pedagogically structured use (Raitskaya and Tikhonova, 2025). This shows the essay does not lean on the panic literature: the negative weight of the magnifier section does not mean the whole field is negative. The difference is again in the use itself, in the hand that holds the lens.
All three of these traps knot in a single place; that knot became visible in a concept the literature on moral discourse has debated much in recent years. Tosi and Warmke define what they call "moral grandstanding" as using moral discourse to appear respectable in others' eyes by this method; the grandstander speaks with a "desire for recognition" they hope will make them look virtuous (Tosi and Warmke, 2016). The neighboring concept Westra calls "virtue signaling" looks to the same place: what is decisive is not the content of the moral expression but the status-seeking motive behind it (Westra, 2021). Zager names another move this way: a person accused of a rule violation neutralizing the accusation by sliding from the language of rules to the language of virtue, that is, by saying "I may have broken the rule but I am a good person," which he calls "toggling" (Zager, 2023). These three concepts are powerful lenses: they make visible the motive or maneuver behind a counterpart's moral stance.
But right here, the self-reflexive lock closes. Flowerree and Satta show that diagnosing grandstanding is itself a trap: we cannot reliably distinguish whether someone is truly grandstanding or speaking sincerely; if we cannot distinguish, we are not epistemically justified in judging someone a grandstander; moreover a propensity to judge others this way points to a lack of intellectual humility (Flowerree and Satta, 2023). This is the last and sharpest turn of the lens. To say "they are really grandstanding, chasing a display of virtue" is, more often than not, a way for the one saying it to establish their own position of superiority. That is, the act of exposing virtue signaling may itself be virtue signaling. This essay, in reading reactions to AI, exhibits a moral-intellectual stance: "see clearly, and test yourselves too." This stance too is subject to the same lock; it cannot exempt itself from it.
So does all this lead us to endless doubt, to being unable to say anything? No; the way out is given again by Flowerree and Satta: instead of reading motive, one should look at the reasons a person declares and at whether their action shows respect to those it addresses. This is what the essay tried to do at all its stages. It did not diagnose the inner motive of the forum commenters; it examined the argument they made, and from which position that argument is coherent. It put its own position (the side that uses AI, that benefits from it) to the same test: in measuring the magnifier's neutrality, in taking care not to turn litmus's distinction into a superiority, and in rejecting the shield in both its personal and corporate forms. As long as the lens stays two-way it is a tool of reading; the moment it locks in one direction, it turns into an instance of the very thing it diagnoses.
VI. Framing: how you see AI decides the outcome
The litmus section showed that the same tool gives different results in two hands but left one question open: the skill that makes the difference is the skill of what, exactly? This section gives it a short answer, because the answer names the mechanism beneath litmus. A large part of the distinction comes from what a person sees AI as.
Orlikowski and Gash give the name "technological frame" to the body of assumptions, expectations, and knowledge people hold about a technology; they also show that different groups hold different frames, and that the incongruence between these frames explains the unexpected outcomes experienced with the same technology (Orlikowski and Gash, 1994). When it comes to AI, two frames stand out. One sees the tool as an answer machine, an oracle: a question is asked, an answer received, and the answer received is the product itself. The other sees it as a processing engine: it places material, context, a well-formed problem before it; here the art is in managing, auditing, and re-feeding what the engine does inside that material.
That this distinction is not an empty analogy is shown by Zamfirescu et al.'s experiment. The study, examining how non-experts design AI prompts, finds that participants carry expectations from their human-to-human teaching experience over to the tool, overgeneralize from a single trial, proceed opportunistically rather than systematically, and that these habits form obstacles before effective prompt design (Zamfirescu-Pereira et al., 2023). That is, the approach that treats the tool as an interlocutor, as an oracle to be spoken with like an ordinary person, fails. The oracle frame is the frame in which the tool spins its wheels. What Long and Magerko call "AI literacy" tries to close exactly this gap: the competencies to critically evaluate the tool, collaborate with it, and recognize its limits (Long and Magerko, 2020). The word literacy matters, because it points to a capacity, not to an innate gift.
What the right frame is not, Yiu, Kosoy, and Gopnik put sharply: it is wrong to think of these systems as an individual human mind, an agent. For them the most accurate frame is to see AI as a powerful cultural technology like writing, print, the library; because these systems are extraordinary imitation engines that transmit humanity's accumulation, but on their own they carry no capacities for seeking truth or making original discovery (Yiu et al., 2023). This frame avoids both errors: it takes the tool as neither a thinking subject nor a magical oracle, but as an engine that processes human accumulation. The person who drives the engine onto the field, places their own material before it, and audits its output with their own knowledge draws production from it; while the person who asks it questions like an oracle and takes the answer as is confuses wheel-spinning with production.
A field counterpart of this change of frame is measured too. In Lee et al.'s knowledge-worker data, AI shifts the weight of critical thinking: from production itself toward verifying what is produced, integrating the parts, and managing the process (Lee et al., 2025). The passage from the oracle frame to the processing-engine frame is, in the good user, not an abstract preference but a measured behavioral difference; they invest their labor not in getting the answer but in auditing and integrating it.
This also makes visible the difference between demonstrated skill and skill genuinely possessed. Mei and Weber distinguish, in the age of AI, "demonstrated" critical thinking from "performed" critical thinking: the product that emerges with the tool's help and the cognitive process a person can carry out without the tool are not the same thing; moreover assessments often measure the first and speak about the second (Mei and Weber, 2025). The oracle frame settles for the demonstrated; the processing-engine frame feeds the performed.
Thus framing names the core of the skill litmus separates. The same tool gives different results because users take it up with different frames; a frame is a mental model, and it is that model which determines the outcome. To say it in the magnifier language of the second section: the hand that holds the lens is, in fact, holding a frame. And this frame, as the fifth section insisted, is not an innate privilege but an acquirable literacy.
Conclusion: seeing clearly
The question at the outset was this: why do people looking at the same object reach such different results? The method followed throughout the essay even shifted the place of the question itself. People reach different results because what they look at is, more often than not, not the object but the shadow their own position casts upon the object. AI works less as a body onto which the reaction falls than as a surface that reflects back where the reacting person stands.
The four lenses were four modes of this reflection. The mirror showed that the content of the reaction gives away what a person is protecting (their competence, their privilege). The magnifier showed that the lens is neutral, making visible both the inadequacy and the accumulation placed beneath it by multiplying them. The litmus showed that the tool separates more than it equalizes; because to produce well with it is a skill that is uncodifiable, maturing through living and mentorship. The shield showed how a technically correct observation like "AI makes things up" can be turned into an excuse that shifts responsibility from agent to tool. The sixth section joined the four in one mechanism: what makes the difference is what a person sees AI as, with which frame they hold it.
But the essay's real burden was not in the individual lenses but in their being reversible. The same lens reads also the hand that reads the counterpart. The one who says "you resist because you are afraid" and the one who says "you have stopped thinking" can both use the same apparatus to confirm their own position. For this reason the lens here did not become a weapon of diagnosis; had it, it would have turned into an instance of the very thing it diagnoses. It remained a reading, a reading that reads itself too.
This does not mean nothing is determined. What can be said passes not through reading motive but through looking at the declared reason and the respect shown. A reaction to AI carries information about its owner's position; but this information does not, on its own, determine whether that reaction is right or wrong. Not to confuse the two is an honesty we owe both to our counterpart and to ourselves.
This essay did not claim, and could not claim, to have discovered a new mechanism. It brought together, at the level of public discourse, studies that had drifted apart from one another (identity threat, tacit knowledge, overreliance, the responsibility gap, moral grandstanding), established the link between them, and applied that link to itself as well. Its value, if any, lies not in the individual findings but in the discipline of this blending and self-testing.
What remains is an attitude. The essay that launched cognitive surrender also tied its close here: "that choice is not about understanding technology, it's about challenging our cultural instincts" (in the Spanish original, "the line is not in the technology but in the attitude"). That attitude belongs to the person using the tool. The one condition this essay adds is that the attitude not be applied in one direction only: to set about reading another's position toward AI without seeing our own is what most blinds the lens in our hand. Seeing clearly, more often than not, begins with first seeing where you are looking from.
References
The references are given uniformly in the original citation form of each source, in keeping with the principle of the reachable address (though the concepts in the text are rendered in English, the citation is the address that takes the reader to the same edition). For online-first publications, where the in-text citation year and the volume year diverge, the volume information is noted in the citation.
Bankins, S., Ocampo, A. C., Marrone, M., Restubog, S. L. D., & Woo, S. E. (2024). A multilevel review of artificial intelligence in organizations: Implications for organizational behavior research and practice. Journal of Organizational Behavior, 45(2), 159-182. https://doi.org/10.1002/job.2735
Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889-942. https://doi.org/10.1093/qje/qjae044
Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 188. https://doi.org/10.1145/3449287
Dans, E. (2026, April 5). Pensar cansa [Blog post]. Enrique Dans. https://www.enriquedans.com/2026/04/pensar-cansa-por-eso-estamos-empezando-a-dejar-que-la-inteligencia-artificial-lo-haga-por-nosotros.html (English version: "Thinking hurts, so why not surrender to AI?" Medium. https://medium.com/enrique-dans/thinking-hurts-so-why-not-surrender-to-ai-68da41119eb2)
Dell'Acqua, F., Rajendran, S., McFowland, E., III, Krayer, L., Mollick, E., Candelon, F., Lifshitz-Assaf, H., Kellogg, K., & Lakhani, K. R. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality (Harvard Business School Working Paper No. 24-013). https://ssrn.com/abstract=4573321
Di Santi, E. (2026). Cognitive amplification vs. cognitive delegation in human–AI systems: A metric framework [Preprint]. arXiv:2603.18677. https://arxiv.org/abs/2603.18677
Dowden, B. (n.d.). Fallacies (Genetic fallacy). Internet Encyclopedia of Philosophy. ISSN 2161-0002. https://iep.utm.edu/fallacy/
Ehsan, U., Passi, S., Saha, K., McNutt, T., Riedl, M. O., & Alcorn, S. (2026). From future of work to future of workers: Addressing asymptomatic AI harms for dignified human-AI interaction. Proceedings of CHI 2026. https://doi.org/10.1145/3772318.3791081 (Preprint: arXiv:2601.21920.)
Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489-530. https://doi.org/10.1111/bjet.13544
Flowerree, A. K., & Satta, M. (2023). Moral grandstanding and the norms of moral discourse. Journal of the American Philosophical Association, 10(3), 483-502. https://doi.org/10.1017/apa.2023.8
Fuchs, T. (2022). Understanding Sophia? On human interaction with artificial agents. Phenomenology and the Cognitive Sciences, 23(1), 21-42 (vol. 2024). https://doi.org/10.1007/s11097-022-09848-0
Gerlich, M. (2025a). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), 6. https://doi.org/10.3390/soc15010006
Gerlich, M. (2025b). From offloading to engagement: An experimental study on structured prompting and critical reasoning with generative AI. Data, 10(11), 172. https://doi.org/10.3390/data10110172
Hadjimichael, D., Ribeiro, R., & Tsoukas, H. (2024). How does embodiment enable the acquisition of tacit knowledge in organizations? From Polanyi to Merleau-Ponty. Organization Studies, 45(4), 545-570. https://doi.org/10.1177/01708406241228374
Kapur, M. (2008). Productive failure. Cognition and Instruction, 26(3), 379-424. https://doi.org/10.1080/07370000802212669
Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task [Preprint]. arXiv:2506.08872. https://arxiv.org/abs/2506.08872
Kruger, J., & Dunning, D. (1999). Unskilled and unaware of it: How difficulties in recognizing one's own incompetence lead to inflated self-assessments. Journal of Personality and Social Psychology, 77(6), 1121-1134. https://doi.org/10.1037/0022-3514.77.6.1121
Lave, J., & Wenger, E. (1991). Situated learning: Legitimate peripheral participation. Cambridge University Press. ISBN 978-0-521-42374-8.
Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25). https://doi.org/10.1145/3706598.3713778
Long, D., & Magerko, B. (2020). What is AI literacy? Competencies and design considerations. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, 1-16. https://doi.org/10.1145/3313831.3376727
Lovett, M. (2026). Individual differences in generative AI effectiveness: An empirical puzzle for HRD research. New Horizons in Adult Education and Human Resource Development. https://doi.org/10.1177/19394225261472792
Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D., & Ho, D. E. (2024). Hallucination-free? Assessing the reliability of leading AI legal research tools (arXiv:2405.20362). https://arxiv.org/abs/2405.20362
Matthias, A. (2004). The responsibility gap: Ascribing responsibility for the actions of learning automata. Ethics and Information Technology, 6(3), 175-183. https://doi.org/10.1007/s10676-004-3422-1
Mei, K. X., & Weber, N. (2025). Designing AI systems that augment human performed vs. demonstrated critical thinking [Preprint]. arXiv:2504.14689. https://arxiv.org/abs/2504.14689
Mirbabaie, M., Brünker, F., Möllmann Frick, N. R. J., & Stieglitz, S. (2021). The rise of artificial intelligence: Understanding the AI identity threat at the workplace. Electronic Markets, 32(1), 73-99 (vol. 2022). https://doi.org/10.1007/s12525-021-00496-x
Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence [MIT working paper; peer-reviewed version: Science, 381(6654), 187-192]. https://doi.org/10.1126/science.adh2586
Orlikowski, W. J., & Gash, D. C. (1994). Technological frames: Making sense of information technology in organizations. ACM Transactions on Information Systems, 12(2), 174-207. https://doi.org/10.1145/196734.196745
Passi, S., & Vorvoreanu, M. (2022). Overreliance on AI: Literature review (Microsoft Research / Aether report).
Raitskaya, L., & Tikhonova, E. (2025). Enhancing critical thinking skills in ChatGPT-human interaction: A scoping review. Journal of Language and Education, 11(2), 5-19. https://doi.org/10.17323/jle.2025.27387
Santoni de Sio, F., & Mecacci, G. (2021). Four responsibility gaps with artificial intelligence: Why they matter and how to address them. Philosophy & Technology, 34(4), 1057-1084. https://doi.org/10.1007/s13347-021-00450-x
Shaw, S. D., & Nave, G. (2026). Thinking, fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender [Wharton working paper; the original title has an em dash after "Thinking"]. SSRN 6097646. https://doi.org/10.31234/osf.io/yk25n_v1
Stanković, M., Hirche, E., Kollatzsch, S., & Doetsch, J. N. (2025). Comment on: Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing tasks [Preprint]. arXiv:2601.00856. https://arxiv.org/abs/2601.00856
Tosi, J., & Warmke, B. (2016). Moral grandstanding. Philosophy & Public Affairs, 44(3), 197-217. https://doi.org/10.1111/papa.12075
Tsao, J., Liang, C. X., Nogues, C., & Wong, A. (2026). Perceptions and integration of generative artificial intelligence in creative practices and industries: A scoping review and conceptual model. AI & Society, 41(3), 2259-2278. https://doi.org/10.1007/s00146-025-02667-2
Westra, E. (2021). Virtue signaling and moral progress. Philosophy & Public Affairs, 49(2), 156-178. https://doi.org/10.1111/papa.12187
Yiu, E., Kosoy, E., & Gopnik, A. (2023). Transmission versus truth, imitation versus innovation: What children can do that large language and language-and-vision models cannot (yet). Perspectives on Psychological Science, 19(5), 874-883. https://doi.org/10.1177/17456916231201401
Zager, S. (2023). Toggling: How shifts between deontology and virtue ethics undermine public moral discourse about gender and race. Hypatia, 38, 294-315. https://doi.org/10.1017/hyp.2023.25
Zamfirescu-Pereira, J. D., Wong, R. Y., Hartmann, B., & Yang, Q. (2023). Why Johnny can't prompt: How non-AI experts try (and fail) to design LLM prompts. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, Article 437, 1-21. https://doi.org/10.1145/3544548.3581388
Version and change log
Litmus · v1.1 · 12 August 2026 · two factual corrections
This version, in the multilingual translation round, integrates two factual corrections that emerged from primary sources; the core argument and structure remain unchanged. ① Ehsan et al.'s AI-as-amplifier paradox was harmonized under its original name (AI-as-Amplifier Paradox). ② Lave and Wenger's quotation on p. 33 was corrected to reflect the mood of the English original (not acquisition: "only has power in specific circumstances"). The bibliography, 38 entries, remains unchanged.
Litmus · v1.0 · 11 August 2026 · first English edition
This is the first English edition, translated from the Turkish original (Turnusol, v1.0, first published 10 August 2026). The body is translated from the Turkish source text; quotations are handled under the raw-quotation regime described below.
Translation regime. The concepts in the text are rendered in English, most of them the original English academic terms; the citation format (author-year, et al., p./pp., &) and the number format (decimal point, APA statistics) follow English convention. The one verbatim quotation is Dans's closing sentence: because the English reader is directed to Dans's own English article, his English wording is used ("that choice is not about understanding technology, it's about challenging our cultural instincts"), whereas the Turkish original is anchored to the Spanish source's "actitud" (attitude); the essay's own "attitude" axis is preserved, and the divergence between the Spanish "actitud" and Dans's English "cultural instincts" is noted here for the record. The Turkish source was built through editorial revision rounds over the author's full rewrite, whose main frames were: drawing the opening claim into the active voice; setting the litmus summary as a confrontation of the equalizing thesis in its strongest, testable form; grounding the shield's priority question ("what is decisive, what the tool was designed for, or what it turns into, in whom?") in its sources; distinguishing Ehsan's "AI-as-Amplifier Paradox" (expertise erosion along the axis of time) from this essay's magnifier (the input made visible along the axis of the moment); and weaving the 2025-2026 studies on AI and cognitive competence (Lee et al., Gerlich, Fan, Kosmyna-Stanković, Raitskaya-Tikhonova, Di Santi, Mei-Weber) into the body along two veins.
The references comprise 38 entries in APA 7. They are given in the original form of the sources under the principle of the reachable address; though the concepts in the text are rendered in English, the citation is the address that takes the reader to the same edition.