The Slow Disappearance of “I’m Not Sure”
How AI’s definitiveness trains us to mistake fluency for truth
There is an old distinction in academic writing between what philosophers call a hedge and what they call an assertion. “The data suggest…” differs meaningfully from “The data prove…” Not only grammatically, but cognitively. The first invites revision; the second closes the door. Skilled thinkers in any field know that the hedge is often the more honest and more useful form, because most questions of any real weight sit in contested territory where the honest answer is: we think so, but it depends on what you weight and how you look.
That distinction is quietly eroding, and one of its least-examined causes is the texture of AI-generated language.
When a large language model answers a question, it typically does so in declarative, confident prose. Not because the model has certainty about the answer, but because declarative prose is statistically dominant in the training data. Instructions, textbooks, news articles, Wikipedia entries, professional documentation, legal summaries, and product descriptions all share one feature: they assert, they do not hedge. So the model learned to assert. The output sounds authoritative because authoritative language trained it, not because the matter is actually settled.
The effect on human readers and, increasingly, on human writers and speakers is what might be called epistemic humility erosion: the progressive displacement of provisional language by the false confidence that comes from well-formed sentences. It is distinct from lying. The model is not telling you it is certain. It simply produces text that carries the register of certainty, because that register is statistically normal.
This is not a new problem in disguise. Every generation has had its oracles, its overconfident authorities, its sciences that were briefly sure of things they should have doubted. What is new is the scale and the invisibility. A confident professor in a lecture hall is visibly a person with a perspective shaped by a career, a school of thought, a set of assumptions. A fluent paragraph of AI text arrives without any of those contextual markers. It reads as if it came from nowhere in particular, which makes it read as if it came from everywhere, which makes it read as if it were simply true.
Linguists who study hedging in academic prose have documented just how much work provisional language actually does. Phrases like “one interpretation is,” “evidence points toward,” “this may be explained by,” or “it is worth noting that” are not signs of weakness. They are signals of epistemic position. They tell the reader: here is where uncertainty lives in this argument; here is where my reasoning could be wrong; here is what I am claiming versus what I am inferring. When those signals disappear, the reader loses the map of where to push back. And when someone spends hours reading text that never gives such signals, they begin to stop looking for the map.
The damage compounds in language learners. Advanced English users at the C1 and C2 level are typically working to master exactly the kind of nuanced hedging that distinguishes a sophisticated academic or professional from a fluent but blunt communicator. The English modal system, in particular, is a finely grained instrument for expressing epistemic position: might, could, may, should, must, would, and ought each carry distinct shades of certainty, obligation, and conditionality. Native-level command of these modals is one of the markers that separates a truly advanced speaker from a merely accurate one. AI text, optimised for surface clarity, tends to collapse this system toward the declarative, giving learners a model that actively undermines the very skill they need most.
There is a cognitive dimension here as well. The psychologist Leon Festinger’s work on cognitive dissonance made clear that people are uncomfortable holding open questions for long, and will often reach for the nearest available closure rather than tolerating continued uncertainty. AI-generated answers are, by their nature, available. They arrive instantly, they sound complete, and they resolve the discomfort of not-knowing. Over time, the habit of sitting with a genuinely difficult question, turning it, poking at it, resisting the first plausible answer, weakens. The muscle for productive uncertainty atrophies in the same way that walking atrophies when escalators are everywhere.
Get Grandomastery’s stories in your inbox
Join Medium for free to get updates from this writer.
A related phenomenon is what the philosophical tradition calls epistemic overconfidence, or the illusion of explanatory depth, documented experimentally by Leonid Rozenblit and Frank Keil. Their studies showed that people consistently overestimate how well they understand mechanical and causal systems. Asked to explain in detail how a zipper or a toilet or a bicycle works, most people discover quickly that their “understanding” was shallow recognition rather than genuine comprehension. The same illusion operates at the level of concepts. A student who has read a fluent AI-generated paragraph about, say, the difference between inductive and abductive reasoning often feels they understand the distinction, while in fact they have absorbed the shape of an explanation rather than the understanding itself.
What is rarely discussed is that this illusion is actively pleasurable. The feeling of knowing is rewarding. It triggers something resembling what Csikszentmihalyi called flow: a sense of frictionless competence. The problem is that learning, especially at advanced levels, requires the opposite feeling at precisely the right moments. It requires the friction of being not-sure-yet. It requires what the mathematician Henri Poincaré, writing in the early twentieth century about mathematical discovery, called the “incubation” of uncertainty, during which the prepared mind sits with an open problem until something genuinely novel assembles.
Grandomastery was built around a different epistemic premise. Its activities are deliberately constructed so that there is no retrievable correct answer, and the task itself is to generate something plausible, interesting, or provocative out of materials that have no obvious relationship. The Random Abstractions activity, for instance, places two abstract nouns in proximity and asks for connections. The connections can be many; none is definitive; the value is in the quality of the reasoning and the willingness to explore without the safety net of a correct destination. This is, structurally, the opposite of querying an AI for an answer. It is the exercise of exactly the cognitive and linguistic capacities that the AI query tends to atrophy.
The hedging vocabulary that advanced English users need, and that AI text systematically under-represents, turns out to be inseparable from the cognitive stance that produces nuanced thinking. “One could argue,” “the tension here is,” “whether this holds depends on,” “this raises the question of,” “it is less clear whether” are not just phrases to memorise. They are the external form of an internal epistemic habit: the habit of holding the provisional as provisional, of marking claims by their confidence level, of treating your own conclusions as revisable.
The Socratic tradition described this as aporia, the productive state of not-knowing from which genuine inquiry begins. Kierkegaard picked it up and called it “the wound of the negative.” The phenomenologists called it Husserl’s epoché, the suspension of the natural attitude. Each of these traditions recognised that the willingness to remain uncertain is not a deficiency but a prerequisite. The great deficiency is premature certainty, the confident voice that has reached its conclusion without having done the work of genuine inquiry.
The irony is that the largest language models trained on human text have something structurally like uncertainty built into them: they produce probability distributions over possible next tokens, and they sample from those distributions. Technically, they never “know” anything. But this technical uncertainty is invisible to the user. What the user sees is the declarative output, not the probabilistic process. The uncertainty is built in but invisible; the confidence is a surface effect but highly visible. The human mind, which learns heavily from models of language in use, learns the surface.
This is not a reason to avoid these tools. It is a reason to understand them precisely, and to deliberately train the epistemic habits they do not provide. Activities that demand provisional language, that reward the quality of reasoning over the correctness of answers, that treat hedging as a skill rather than a weakness, counteract the tendency. Discussion clubs, spontaneous presentation formats like Pecha Kucha, debate structures that require steelmanning the opposing position, activities where the task is to generate multiple interpretations rather than one answer: all of these are training for the epistemic stance that AI-generated text does not model.
The deeper concern is about what happens when the next generation of professionals, researchers, students, and decision-makers has spent its formative years in an information environment that mostly asserts and rarely hedges. The risk is not stupidity. People raised on AI-generated text will likely be fluent, well-informed by some measure, capable of producing plausible explanations. The risk is a particular kind of intellectual fragility: brittleness in exactly the situations where revision matters most, where the stakes are high enough that holding an open question for a while, rather than reaching for the nearest confident answer, would produce a markedly better outcome.
Alexander Popov’s grandomastery.com offers a working model of an alternative, not because it has an agenda against AI per se, but because its premise was always that the thing worth training is the human capacity that cannot be predicted in advance: the capacity to generate meaning from materials that do not come pre-assembled into answers.
The medieval scholars who debated angels on pinheads were, in their way, practicing something important. They were training the ability to construct arguments from first principles, to hold positions provisionally while testing them against objections, and to recognise that the quality of the question sometimes matters more than the certainty of the answer. Whether the activity looks useful from the outside is beside the point. The habit of mind it trains is exactly what the information environment of the twenty-first century is quietly dismantling.
Learning to say “I am not sure, but here is what the evidence inclines me toward” is not a soft skill. It is, at this particular moment in the history of knowledge, a critical one.
#EpistemicHumility #CriticalThinking #AILiteracy #LanguageLearning #AdvancedEnglish #ESL #TESOL #TEFL #EFL #EnglishTeacher #EdTech #HigherOrderThinking #21stCenturySkills #CognitiveBias #AIandEducation #LinguisticCreativity #AcademicEnglish #C1English #C2English #CreativeThinking #ThinkingSkills #HedgingLanguage #Metacognition #CriticalLanguageAwareness #LearnEnglish #EnglishFluency #EnglishForAcademia #ModalVerbs #AcademicWriting #Grandomastery #LanguageEducation #CreativityTraining #DivergentThinking #IntelligenceVsWisdom #AIinEducation #OverconfidenceBias #Bisociation #SpokenEnglish #DiscussionSkills #IntellectualHumility #ESLTeacher #LanguageAndThought #ESLResources #ELT #EnglishLearners #EnglishCommunication #Argumentation #AcademicDiscourse #HumanCreativity #CognitiveFluency
