Short answer: STEM heutagogy is self-determined learning applied to STEM so that the learner retains ownership of the question, the first attempt, the evidence, and the revision. AI can support that process, but it should not become the origin of the learner’s reasoning.
That distinction matters more now because generative AI can remove friction faster than any previous classroom tool. A student can move from question to polished answer in seconds. The result may look efficient while hiding whether the student ever formed a model, tested an assumption, noticed an error, or changed their mind.
Key Finding
The goal is not to keep AI out of learning. It is to keep the learner inside the reasoning. A useful AI interaction should leave the student with more capability, not merely a better-looking output.
What is STEM heutagogy?
Heutagogy is generally described as self-determined learning. Stewart Hase and Chris Kenyon introduced the term in 2000 as a move beyond teacher-directed pedagogy and learner-directed andragogy toward a model in which learners take greater responsibility for deciding what and how they learn. Their emphasis was not simply independence for its own sake, but the development of capability: the ability to use knowledge adaptively in unfamiliar situations.
That idea is especially compatible with STEM. Scientific inquiry begins with questions. Engineering advances through constraints, prototypes, failure and revision. Mathematics depends on constructing and checking relationships. Technology becomes meaningful when a person understands enough of a system to use, inspect and challenge it.
In STEM.org’s current human-agency framework, STEM Heutagogics extends this tradition into system-rich learning environments. The learner should not be reduced to the last person in a chain that begins with a platform, a prompt template or an algorithmic recommendation. The learner remains the primary agent who frames the problem, produces an initial model, evaluates evidence and decides what must change.
This develops earlier writing on heutagogy and the future of STEM education into an operational classroom protocol.
AI changes the cost of an answer—not the work of knowing
There is a strong reason to explore AI as a learning support. Recent research on self-regulated learning shows that AI systems can provide adaptive prompts, feedback and individualized support that would be difficult for one teacher to deliver simultaneously to every learner.
But the evidence does not support treating AI as an automatic learning accelerator. A 2025 systematic mapping review identified 84 studies at the intersection of AI and self-regulated learning and found that the research was concentrated heavily in higher education, with much less attention to primary education and educators. The literature also spans very different tools and purposes, making broad conclusions difficult.
A 2026 randomized controlled trial brought the question into real secondary-school classrooms. Researchers assigned 371 students in Grades 7–9 to two targeted GenAI support conditions or a standard ChatGPT control across six 45-minute sessions. The targeted support helped preserve one motivational measure—perceived utility value—but did not produce significant advantages over the control condition for effort, domain knowledge or elaboration-based strategy use. The authors concluded that GenAI support may help some aspects of self-regulated learning, but that stronger designs are still needed for cognitive strategies and learning outcomes.
Research Note
AI can scaffold self-regulation. It can also over-scaffold it. The practical question is not whether a learner used AI, but which parts of the learning process remained cognitively owned by the learner.
A six-step learner-first protocol
The following sequence is designed for classrooms, programs and independent learners. It is not a claim that every STEM task should use AI. It is a way to preserve human agency when AI is available.
Choose the question
The learner states the problem in their own words and explains why it matters. The teacher can define the domain, safety limits and constraints, but the learner should be able to identify the question without copying it from an AI response.
Make a first model
Before prompting a system, write a prediction, sketch a mechanism, estimate an answer, build a rough explanation or identify what would count as evidence. The first model can be wrong. Its purpose is to give the learner something real to test.
Attempt the work
Calculate, measure, prototype, retrieve from memory, inspect a source or work through the problem far enough to expose uncertainty. This is where productive struggle and useful friction become educational assets rather than inconveniences.
Use AI as a critic, not a substitute
Only after an independent attempt should AI enter the loop. Ask for a counterexample, a missing variable, a competing explanation, a hint, a test case or questions that reveal weak reasoning. A prompt that asks “What am I overlooking?” is often more educational than “Give me the answer.”
Verify outside the conversation
Check a calculation, inspect a primary source, run the experiment, compare measurements, reproduce the result or use another independent method. The AI output is a claim to examine, not evidence by itself. The same principle underlies estimating before trusting a precise-looking output.
Revise and explain
The learner records what changed and why. If the evidence contradicts the first model, revise it. If the anomaly is weak or the measurement is questionable, test again. Good STEM learning includes the discipline of knowing when evidence is strong enough to change a model.
Pass Condition
Close the AI window. Can the learner still explain the question, the first model, the evidence, the system’s contribution, and the reason for the final revision? If not, the system probably carried too much of the cognitive load.
Self-determined does not mean unsupported
Heutagogy is sometimes misunderstood as the teacher disappearing and the learner being left alone. That is not the useful interpretation for STEM education.
Teachers remain essential because they design the environment in which self-determination becomes possible. They protect safety, set meaningful constraints, introduce tools, expose learners to disciplinary standards, ask better questions, challenge weak evidence and help students recognize when confidence has outrun understanding.
The shift is in the location of intellectual ownership. The teacher does not need to decide every question, sequence every move or pre-author every conclusion. Instead, the teacher creates conditions in which learners can increasingly make those decisions with competence.
Use AI to
- generate counterexamples and test cases;
- surface assumptions the learner may have missed;
- offer hints after a genuine attempt;
- compare alternative explanations;
- support reflection after evidence has been gathered.
Do not let AI
- choose every meaningful question;
- produce the first model by default;
- replace measurement, calculation or source checking;
- turn uncertainty into instant confidence;
- write the reflection that is supposed to reveal the learner’s thinking.
How assessment changes when agency matters
If the final answer is the only artifact that counts, generative AI creates an obvious problem: the most polished output can come from the least visible thinking.
A learner-first STEM environment therefore needs more inspectable evidence. Instead of grading only the finished response, educators can ask for a short trail: the original question, first prediction or model, one independent attempt, the AI prompt that mattered most, the verification step, and a final explanation of what changed.
This does not require turning every lesson into paperwork. A photograph of a prototype, a three-line estimate, a lab notebook entry, a marked-up diagram, a short oral explanation or a version history can make the reasoning visible. The goal is to preserve enough of the process that the learner and teacher can still see where understanding was formed.
This approach also complements STEM.org’s five tests for preserving human agency in responsible AI use. The principle is the same: tools should remain inspectable, bounded and subordinate to human judgment.
Why this matters beyond school
The long-term value of self-determined learning is not that every learner becomes completely independent of teachers, institutions or tools. No serious STEM practice works that way. Engineers depend on standards and teams. Scientists depend on instruments and peer criticism. Technicians depend on documentation. Mathematicians depend on shared notation and prior results.
Capability means something more demanding: a person can enter a new situation, identify what they do not know, choose a way to learn it, test what they find and revise their approach without surrendering judgment to the first system that offers an answer.
That is precisely the capacity that becomes more important as AI becomes more capable.
The easier it becomes to obtain an output, the more education must protect the learner’s ability to originate a question, inspect a process and defend a conclusion with evidence. STEM Heutagogics is not a rejection of intelligent tools. It is a design rule for keeping intelligence human where it matters most.
Use the system. Keep the reasoning.
Research cited: Hase & Kenyon, “From andragogy to heutagogy” · 2025 systematic mapping review of AI and self-regulated learning · 2026 randomized school trial of GenAI-supported self-regulated learning
