Your Students Already Have Token Anxiety
They Just Don’t Know What to Call It Yet
Depending on how deeply you have integrated AI into your daily practice as an educator, the following scene will either sound familiar or like a glimpse of what is to come.
A high school English teacher sits at her kitchen table on a Sunday evening, using an AI assistant to generate differentiated reading guides for three ability levels. What once took an entire afternoon now takes twenty minutes. But she doesn’t stop there. The tool is so fast, so fluent, that she decides to create vocabulary scaffolds, discussion prompts, and formative assessment rubrics as well. By midnight, she has produced more material than she could use in a month, and she feels more exhausted than if she had planned the old-fashioned way.
Meanwhile, one of her students is working on a literary analysis essay. He drafts his thesis in ChatGPT, asks it to strengthen his argument, then requests counterarguments he can preemptively address. The output is polished, syntactically confident, and almost entirely not his own thinking. He submits it with a nagging sense that he hasn’t actually learned anything, but his grade will probably be fine.
Neither the teacher nor the student has a name for what they are experiencing. But the technology industry does. Developers and AI engineers call it token anxiety: the compulsive pressure to optimize every interaction with a language model, and the creeping sense that if the machine can always do more, perhaps you should always do more too.
A concept born in Silicon Valley
A token, in the technical sense, is the fundamental computational unit that a large language model processes. It maps roughly to three-quarters of an English word. Every query to a commercial AI system consumes tokens, and those tokens cost money. However, the anxiety linked to them, as demonstrated in tech culture, goes far beyond just financial concerns.
Industry observers have documented a cultural shift in which developers leave social gatherings early, or wake in the middle of the night, driven by a compulsive need to check on overnight code generation and ensure their daily token allotments are being fully used. The metric of professional worth has shifted from the quality of human output to the volume of tokens processed per day. In developer communities, token anxiety describes the cognitive overhead of constantly monitoring, budgeting, and optimizing AI consumption, combined with the psychological weight of a tool that never stops being available and never stops implying that you could be more productive.
Education hasn’t adopted this language yet. But the underlying dynamics — the pressure to extract maximum productivity from AI and the anxiety of managing a tool that is simultaneously indispensable and overwhelming — are already reshaping faculty workloads and student learning habits. The phenomena are here. The vocabulary to describe them is still catching up.
The intensification trap: why AI makes teaching harder
One of the most persistent assumptions about AI in education is that it reduces workload. Technology vendors market language models as tools that liberate educators from administrative drudgery, freeing them to focus on the relational, empathetic work of teaching. A 2026 longitudinal study published in the Harvard Business Review, conducted by researchers from the University of California, Berkeley, directly challenges this narrative. After eight months studying employees at a technology company following the integration of generative AI, the researchers concluded AI tools do not reduce work. Instead, they consistently intensify it.
The mechanism is straightforward but insidious. Because AI lowers the barrier to entry for complex tasks, people voluntarily expand their scope of responsibilities. The teacher who previously lacked time to create differentiated materials for every reading level now attempts to do so because the tool makes generation possible. But she remains responsible for reviewing, editing, and implementing that expanded volume of material. The work hasn’t decreased; it has metastasized.
The study also identified how AI erodes the boundaries between work and rest. Because initiating a task becomes nearly frictionless, professionals slip small work activities into moments previously reserved for cognitive recovery: lunch breaks, evenings, or weekends. The always-available assistant creates an always-available expectation. For teachers already facing unsustainable workloads, this is a genuinely dangerous dynamic.
What strikes me most about this pattern is the shift it represents in professional identity. Teaching has always been cognitively demanding work. Before AI, educators engaged in continuous diagnostic reasoning by interpreting student confusion and adjusting explanations in real time. As a 2026 analysis from the UNESCO World Education Blog frames it, these represent the cognitive core of professional expertise. When AI systems assume instructional decisions, they displace precisely this judgment, redirecting the labor of teaching toward algorithmic oversight and the policing of academic integrity. The teacher becomes a manager of machine output rather than a practitioner of a craft.
This is, in all but name, token anxiety applied to education. The stress of managing an ever-capable system, the expansion of expected output, and the collapse of boundaries between working and not working — these are the same dynamics the tech industry identified and named. Educators are experiencing them without the conceptual framework to push back.
The performance paradox: why students learn less while producing more
The consequences for students are equally concerning, though they manifest differently. The central risk is what researchers have termed the “productivity paradox”: AI significantly boosts a student’s immediate task performance while simultaneously undermining the durable, long-term learning that education exists to produce.
This paradox makes sense through the lens of Cognitive Load Theory, a framework used in instructional design that categorizes mental effort into intrinsic load (the inherent difficulty of the material), extraneous load (unnecessary mental effort caused by poor design), and germane load (the productive effort dedicated to building lasting knowledge structures). AI excels at reducing extraneous load. It can summarize dense texts, clarify jargon, and organize information with impressive efficiency. The danger lies in its tendency to eliminate germane load as well: the productive cognitive struggle through which students actually learn.
Lev Vygotsky’s theory of the Zone of Proximal Development holds that optimal learning occurs in the space between what a student can accomplish independently and what they can achieve with guided support. When an AI system immediately provides a polished answer, it collapses that zone entirely. The student never engages in the effortful processing required to move concepts from short-term awareness into durable understanding.
Empirical evidence is confirming these theoretical concerns. A 2025 study examined how students navigate the intersection of academic pressure and AI availability. The researchers found that 32.7% of the student population showed moderate-to-severe AI dependency patterns, and this rate climbed as students progressed academically, from 27.1% among first-year students to 37.9% among fourth-year students. Students exhibiting high AI dependency recorded a mean GPA deficit of 0.41 points compared to their non-dependent peers. Through regression analysis, the study identified measurable decreases in independent analytical ability, weakened writing skill development, and reduced content knowledge acquisition among dependent users.
What the study also revealed, compellingly, is that students recognized the problem. Faced with minimal institutional support, many developed their own coping strategies. The most common was what the researchers termed “strategic access rationing”: deliberately limiting AI usage to specific times, or requiring themselves to complete an independent draft before consulting the tool. This is a sophisticated metacognitive response, and it mirrors the self-regulation strategies that developers in the tech industry have adopted to manage their own token anxiety.
The parallel is striking. Students are independently developing similar coping strategies, but they lack a common language to recognize the recurring pattern they’re experiencing.
The third digital divide: tokenization as a vector for inequity
The equity dimensions of this emerging anxiety are perhaps the least visible and the most consequential. They center on a technical reality that most educators have never encountered: the way language models process text is structurally biased against non-English languages.
Tokens are not words. They are subword units generated through compression algorithms trained predominantly on English-language, Latin-script datasets. For English, a token represents roughly three-quarters of a word, meaning a 16,000-token context window can accommodate approximately 12,000 words. For morphologically complex languages or non-Latin scripts, the same tokenizer fragments characters far less efficiently. A single word in Hindi might require nearly six times as many tokens as its English translation. In Armenian, the multiplier approaches ten.
The implications compound in multiple directions. Financially, non-English speakers pay a steep premium to process the same semantic information through commercial AI systems. An institution processing texts in Korean might deplete its token budget about five times faster than an English-speaking counterpart. Cognitively, the inflated token consumption reduces the effective working memory of the AI during conversations. A tutor conducting a multi-turn dialogue in Arabic will lose track of earlier instructions and student responses far sooner than one operating in English, leading to degraded pedagogical continuity and increased hallucination rates. For indigenous languages absent from standard tokenizer vocabularies altogether, the AI becomes functionally useless.
Researchers and policymakers describe this as an emerging third digital divide. The first divide concerned access to hardware and internet connectivity. The second concerned the digital literacy skills required to use technology effectively. And the third concerns access to premium AI models, sufficient computational resources, and token-efficient architectures. As a 2026 Brookings Institution analysis argues, while basic AI interfaces may be freely available, they impose strict rate limits and constrained capabilities. Wealthy institutions can afford enterprise subscriptions with massive context windows and state-of-the-art reasoning models. Marginalized learners, particularly those in developing nations or using non-Latin languages, must ration their AI use while simultaneously contending with algorithmic inefficiency that penalizes them for the language they speak.
This is consistent with research that Darya Ramezani and I published at the 2025 International Conference on Education and New Developments (END2025), where we described an emerging “AI Productivity Divide” that operates across technical access, AI literacy, and institutional readiness. What distinguishes this divide from earlier digital divides is that it encompasses both access limitations and voluntary non-adoption, creating compound disadvantages. A student who lacks reliable internet access and a student who has access but lacks the literacy to use AI tools effectively end up in a similar position. They are both falling further behind peers whose institutions have invested in infrastructure and thoughtful integration. The tokenization bias described above adds yet another layer: even students with access, literacy, and institutional support may face degraded AI performance simply because of the language they speak.
Teaching with intention: what AI-resilient pedagogy looks like
Addressing these pressures requires more than better tools or cheaper subscriptions. It requires rethinking what we ask students to do with AI and why.
The most important shift is from product to process. Traditional assessment models reward final artifacts: the polished essay, the correct answer, or the completed problem set. Because language models generate these artifacts with ease, any assessment anchored solely to the final product will incentivize dependency. AI-resistant assessment, by contrast, primarily evaluates the process of learning. Students show their understanding by defending their reasoning and the choices they made along the way. This approach stimulates what researchers call the “protégé effect,” whereby the effort of explanation and critique consolidates long-term understanding far more effectively than passive consumption.
Equally important is protecting what cognitive scientists call productive struggle. There’s a pedagogical benefit to explicitly telling students that certain fundamental assignments restrict AI. The cognitive work of confronting a difficult concept builds durable understanding in ways that frictionless AI interaction cannot. This requires normalizing discomfort in learning, which runs counter to the consumer logic of AI platforms designed to minimize every form of friction.
Institutional policy deserves equal attention. AI detection algorithms have been shown to be unreliable and to penalize non-native English speakers and neurodivergent students disproportionately. A surveillance-centered approach to academic integrity will fail. Transparent policies that establish clear expectations for AI use, teach algorithmic literacy as a core competency, and treat strategic AI engagement as a skill to be developed will serve students far better than an arms race between generation and detection.
Giving the problem a name
There is real value in having the right term for what is happening. When educators feel overwhelmed by AI without knowing the reason, and when students sense constant AI access is diminishing their abilities but can’t explain the mechanism, these are symptoms of a coherent phenomenon. The technology industry, for all its excesses, has at least produced a name for it.
Token anxiety may have originated in developer culture, but the condition it describes is already present in schools and universities worldwide. Naming it within education would give teachers and students a shared framework for recognizing the pattern: the compulsive optimization, the erosion of rest, the confusion of efficiency with learning. And naming a problem is often the first step toward designing a response.
As I have argued many times, I firmly believe that generative AI can genuinely augment education. It can reduce unnecessary cognitive load, provide scaffolding for struggling learners, and handle administrative tasks that drain instructional energy. But it can only do these things within a pedagogical framework that understands the difference between making a task easier and making a student more capable.
The struggle of learning is not an inefficiency to be optimized away. It is the mechanism through which understanding develops. The schools and educators who recognize this will be the ones best prepared to resist the quiet pressure of a technology that always suggests you could do more.
The images in this article were generated with Nano Banana 2.
P.S. I believe transparency builds the trust that AI detection systems fail to enforce. That’s why I’ve published an ethics and AI disclosure statement, which outlines how I integrate AI tools into my intellectual work.