Asking for Help Is a Skill. AI Is Quietly Replacing It.

A ninth grader stares at a geometry proof. She reads the problem twice. Then she opens a chat window and types, "explain step by step." Forty seconds later she has the answer, copied into her notebook, and her hand never went up in class.

Something is missing from that scene, and it is not the answer.

For decades, educational psychologists have studied a quiet skill called academic help-seeking. It sounds trivial, the kind of thing a parent would file under "common sense." It is not. Help-seeking is a learnable, observable, achievement-correlated competency that depends on a chain of cognitive moves: noticing you are stuck, deciding what kind of help you need, choosing whom to ask, formulating a question that fits the listener, evaluating the answer, and integrating it into what you already know. Strip any link out of that chain and the chain stops doing its job.

Generative AI is rapidly becoming the first stop for student help-seeking, and in many cases the only stop. According to Pew Research Center's 2026 survey, sixty-four percent of U.S. teens now use AI chatbots, and about ten percent say they do all or most of their schoolwork with one (Pew, 2026). The share of teens using ChatGPT for schoolwork doubled between 2023 and 2024, from thirteen percent to twenty-six percent (Pew, 2025). The chatbots are fluent, patient, available at midnight, and never sigh. From a student's perspective, they are a strict upgrade over the embarrassment of raising your hand. From a developmental perspective, they are quietly displacing one of the most important skills school is supposed to build.

Asking for Help Is a Skill, Not a Personality Trait

The intuition that some kids are "just shy" and others are "just confident" misses what the research actually shows. Richard Newman, professor emeritus at UC Riverside, spent his career documenting that adaptive help-seeking is a strategy, not a temperament. In his 2002 paper for Theory into Practice, he laid out the four competencies a student needs in order to seek help productively: cognitive competence to recognize that help would be beneficial, social competence to request it, motivational competence to tolerate the difficulty in the first place, and a classroom context that does not punish the asking (Newman, 2002).

That last item matters more than people realize. Stuart Karabenick and Richard Newman's edited volume Help Seeking in Academic Settings: Goals, Groups, and Contexts (Routledge, 2006) brings together decades of work showing that help-seeking is one of the strongest correlates of self-regulated learning. Karabenick and Knapp (1991) had already shown that students who used help-seeking strategies also used more cognitive and metacognitive learning strategies overall. Asking is part of how you study, not separate from it.

The literature also distinguishes between two kinds of asking. Instrumental help-seeking is asking for the minimum support needed to keep working: a hint, a clarification, a question that probes your reasoning. Executive help-seeking is asking for the answer so you can stop working. Instrumental help-seeking is positively associated with achievement; executive help-seeking is not (Karabenick, 2003). The distinction is the whole ballgame, and it lives entirely in how the student approaches the conversation.

Why Students Avoided Asking, Even Before AI

Anyone who has watched a classroom knows that even with a kind teacher and willing peers, many students will not ask. The reason has been studied to exhaustion. Allison Ryan and Paul Pintrich's 1997 work, and a follow-up review with Margaret Hicks and Carol Midgley, found that adolescents who were oriented toward demonstrating ability or maintaining social status reported feeling threatened by the prospect of asking a question, and consequently avoided asking even when they needed to (Ryan, Pintrich, & Midgley, 2001). The middle-school refrain is real: "I think the teacher or other kids might think I am dumb when I ask a question in math class."

So when AI shows up offering private, judgment-free, instant help, it looks like a humane fix to a longstanding problem. And in some respects it is. A student who would never have raised her hand can now get a clear explanation of an exponent rule, in any pace, in any tone, with no audience. Pew's 2026 data show that about a quarter of teens describe chatbots as extremely or very helpful for completing schoolwork (Pew, 2026). That is not nothing.

But the help-seeking research has always insisted that the asking is part of the learning, not a workaround that bypasses it. The cognitive work of formulating a question, the social work of asking a person whose reaction you cannot control, and the metacognitive work of evaluating the answer against your own thinking are all developmental rehearsals. They build the muscles that let a student function in a discussion section, a job interview, a doctor's office, a code review. If we automate them out of childhood, we should not be surprised when they atrophy.

The Hidden Curriculum of Asking a Real Person

Consider what actually happens when a student walks up to a teacher and says, "I don't get it." First, she has to admit it, which trains tolerance for not-knowing. Second, she has to formulate the request, which forces her to localize the confusion: is it the formula, the setup, the vocabulary, or the underlying concept? Cognitive scientists call this metacognitive monitoring, and it is the precursor to nearly every form of self-correction. Third, she has to listen and read the teacher's reaction, which is partly an answer and partly a diagnostic. Fourth, when the explanation lands wrong, she has to repair the conversation: "Wait, I meant the second part."

Each of those moves is a tiny but real piece of intellectual development. A chatbot does not require any of them. It accepts a half-formed prompt, infers what the student probably meant, and produces a confident, polished response that matches the surface of the question rather than the depth of the confusion. The student gets an answer. She does not get the practice of being clear with another mind.

This is not a hypothetical concern. Researchers studying generative AI in classrooms have begun to document a pattern that learning scientists have seen before. A 2025 paper in the British Journal of Educational Technology by Xu and colleagues found that students using GenAI tools without explicit metacognitive scaffolding showed weaker self-regulated learning behaviors than students who were prompted to plan, monitor, and evaluate their AI use (Xu et al., 2025). Without scaffolds, the tool tends to absorb the metacognition rather than provoke it.

Intelligent Tutors Taught Us This Lesson Twenty Years Ago

The most uncomfortable part of the current AI moment is that the field has run this experiment before, with weaker tools, and the results were not encouraging.

In the 2000s, Vincent Aleven, Kenneth Koedinger, Bruce McLaren, and Ido Roll built and studied "cognitive tutors," which were intelligent tutoring systems offering on-demand hints to students working through math problems. The systems were genuinely effective at delivering help. The students were not always effective at using it. Aleven and Koedinger found that a startling share of student help-requests, about seventy-two percent in one analysis, represented unproductive help-seeking: drilling through hints to extract the answer, gaming the system, requesting hints they did not need, or refusing hints they did need (Aleven et al., 2006).

The team built a "Help Tutor" that gave metacognitive feedback when students asked for help in unproductive ways. They published their summary of nearly two decades of this work under a deliberately humble title: "Help Helps, But Only So Much" (Aleven et al., 2016).

"Help helps, but only so much."

Read that sentence twice. The conclusion of the most rigorous body of research we have on automated help-seeking is that the value of on-demand help is bounded by the metacognitive sophistication of the student using it. When students were taught how to seek help, the gains transferred to new content (Roll, Aleven, McLaren, & Koedinger, 2011). When students were not taught, the help did not save them.

Generative AI takes the affordances of those old cognitive tutors and dials them to eleven. The hint is now a paragraph. The answer is now a polished essay. The "ask a question" interface is now a chat that will explain the question itself if you do not understand it. And we are handing this interface to children who, by their own admission, do not know how to ask for help in person without feeling stupid.

The First Stop Becomes the Only Stop

There is a particular pattern worth naming. In the help-seeking literature, the most successful learners do not ask a person every time they get stuck; they cycle. They try, get stuck, attempt to repair, ask a peer, try again, ask the teacher, integrate, retry. Each loop builds the next. The chatbot collapses the cycle. The student who would have struggled, asked a friend, attempted a revision, and then approached the teacher now asks the chatbot, gets the polished output, and stops. There are no further iterations because the work feels done.

This is the core of why AI tutoring is qualitatively different from a human tutor, even a good one. A human tutor is constrained by their own pace, their own limits, and their interest in the student's growth. A good tutor will refuse to give the answer. A good tutor will ask the student to predict what comes next. A good tutor will use silence. A chatbot has no incentive to do any of that, because its product is the answer, and its measure of success is user satisfaction.

The Bellwether report on productive struggle put the broader version of this concern bluntly. The risk, they wrote, is that AI tools "privilege efficiency over learners' epistemic agency, normalizing quick answer-taking rather than verification, sense-making, and productive struggle" (Bellwether, 2025). Efficiency is a fine virtue at the end of learning. It is the wrong virtue at the beginning.

So What Do We Do

Schools and parents do not need a manifesto. They need a small number of practices that put help-seeking back into the curriculum.

Teach the difference between instrumental and executive help. Tell students explicitly that asking for a hint is different from asking for an answer, and grade the difference. Some teachers we have talked to have started requiring a "what I tried" paragraph alongside any AI use on a homework assignment. The effort to articulate what was tried is the metacognitive practice the chatbot otherwise erases.

Build a hierarchy. A student stuck on a math problem might be expected, by classroom norm, to first reread the problem, then check notes, then ask a peer, then post in the class channel, and only then turn to AI. The hierarchy reintroduces the cycle that a chatbot collapses. It also restores the social rehearsal that adolescents desperately need.

Make the asking visible. The Aleven and Roll line of research is clear that students do not magically learn to seek help productively, but they can be taught when their help-seeking is monitored and given feedback. In a process-visible writing environment, where keystrokes, drafts, and revisions are captured, a teacher can see the cycle a student went through, including their use of outside tools. That is far better data than asking after the fact whether a student used AI.

Talk about the social cost honestly. Parents and teachers can acknowledge what Ryan and Pintrich documented: that asking for help in front of peers can feel humiliating, and that AI can feel like a relief. Then they can name what is lost when relief becomes the default. Children can hold both ideas. Adults often cannot.

What We Are Really Protecting

There is a tendency in AI-in-education debates to frame everything as a fight about cheating. That framing is too narrow. The deeper question is what kinds of human capability we want school to build, and what trade-offs we will accept to get there.

Help-seeking is not a workaround for being unable to learn. It is one of the ways learning works. The student who knows how to ask, of whom, and for what is the student who will keep learning long after the last assignment. We have spent decades learning how to teach that skill. We are now in the middle of an unintended experiment in how quickly a generation can lose it. The good news is that the same research that warned us about this also tells us how to fix it. We just have to decide that the asking is worth protecting.

Sources

1. Aleven, V., McLaren, B. M., Roll, I., & Koedinger, K. R. (2006). Toward meta-cognitive tutoring: A model of help seeking with a cognitive tutor. International Journal of Artificial Intelligence in Education, 16(2), 101–128.
2. Aleven, V., Roll, I., McLaren, B. M., & Koedinger, K. R. (2016). Help helps, but only so much: Research on help seeking with intelligent tutoring systems. International Journal of Artificial Intelligence in Education, 26(1), 205–223.
3. Bellwether (2025). Productive Struggle: How Artificial Intelligence Is Changing Learning, Effort, and Youth Development in Education. bellwether.org.
4. Karabenick, S. A. (2003). Seeking help in large college classes: A person-centered approach. Contemporary Educational Psychology, 28(1), 37–58.
5. Karabenick, S. A., & Knapp, J. R. (1991). Relationship of academic help seeking to the use of learning strategies and other instrumental achievement behavior in college students. Journal of Educational Psychology, 83(2), 221–230.
6. Karabenick, S. A., & Newman, R. S. (Eds.). (2006). Help seeking in academic settings: Goals, groups, and contexts. Routledge.
7. Newman, R. S. (2002). How self-regulated learners cope with academic difficulty: The role of adaptive help seeking. Theory into Practice, 41(2), 132–138.
8. Pew Research Center (2025, January 15). About a quarter of U.S. teens have used ChatGPT for schoolwork, double the share in 2023.
9. Pew Research Center (2026, February 24). How Teens Use and View AI.
10. Roll, I., Aleven, V., McLaren, B. M., & Koedinger, K. R. (2011). Improving students' help-seeking skills using metacognitive feedback in an intelligent tutoring system. Learning and Instruction, 21(2), 267–280.
11. Ryan, A. M., Pintrich, P. R., & Midgley, C. (2001). Avoiding seeking help in the classroom: Who and why? Educational Psychology Review, 13(2), 93–114.
12. Xu, J., et al. (2025). Enhancing self-regulated learning and learning experience in generative AI environments: The critical role of metacognitive support. British Journal of Educational Technology.