What Generative AI Outputs Can Teach Us About Physiology Misconceptions
In the early days of using generative AI, sometimes I would read an output that made my soul go, Bless your heart.
I was experimenting with using generative AI for drafting formative assessment questions, when I got this response after asking for a question about the sliding filament model:
ChatGPT Output:
What is the mechanism of muscle contraction and relaxation?
a. Myosin filaments slide past actin filaments, powered by ATP; calcium ions bind to troponin to initiate relaxation
b. Myosin filaments slide past actin filaments, powered by calcium ions; ATP binds to troponin to initiate relaxation
c. Actin filaments slide past myosin filaments, powered by ATP; calcium ions bind to troponin to initiate relaxation
d. Myosin filaments bind to actin filaments, powered by ATP; calcium ions unbind from troponin to initiate relaxation
e. Myosin filaments slide past actin filaments, powered by ATP; ATP binds to troponin to initiate relaxation
f. Actin filaments slide past myosin filaments, powered by calcium ions; calcium ions bind to troponin to initiate relaxation
Problem. None of those answers were quite right. When I asked which answer was correct, the output read:
ChatGPT Output:
Answer c. Actin filaments slide past myosin filaments, powered by ATP; calcium ions bind to troponin to initiate relaxation is the correct answer.
Nope. I tried again:
My Input:
But calcium ions bind to troponin to initiate contraction. How does a muscle relax, in relation to the sliding filament model
ChatGPT Output:
You are correct, I apologize for my mistake. The correct answer for question 10 should be: b. Myosin filaments bind to actin filaments, powered by ATP; ATP unbinds from myosin to initiate relaxation. In the sliding filament model, the contraction of a muscle occurs when the myosin filaments bind to the actin filaments, powered by ATP. This binding forms cross-bridges between the two types of filaments, allowing for the sliding of the actin filaments past the myosin filaments, which shortens the sarcomere and results in muscle contraction. To relax the muscle, ATP unbinds from the myosin head, causing the cross-bridges to release and the myosin filaments to return to their original position. This process allows the actin filaments to slide back to their original position, which lengthens the sarcomere and results in muscle relaxation.
I prompted again, but got the same response. Still incorrect.
Granted, this was GPT-3, which may not have much training data including the word hydrolysis (which, to be fair, could also be assumed of incoming anatomy and physiology students). Though, reading the syntax, it is reminiscent of a novice explanation of ATP and the sliding filament model: incorrect cause-and-effect relationships, and oversimplified binding of molecules without a discussion of energy transfer.
I wondered at first if the output was regurgitating misinformation from training data, but then I wondered if it was a different error. Maybe conventions of the English language conflict with biological mechanisms. Simplified explanations available in training data may subvert processes and amplify oversimplifications, mirroring novice-level understanding.
Why Generative AI Makes Physiology Mistakes
Language-based generative AI tools are trained to predict the next best word in a sequence. Domain-specific knowledge requires its own logic, and mastering physiology requires cause-and-effect thinking of understanding a structure/function that is next in sequence, not just a word. Current models produce more accurate physiological explanations than they used to, but the fundamental way of knowing is the same: the output is based on language, not on logic.
English arguments lean into teleology, emphasizing purpose. It makes sense for an LLM, trained to predict words in a sequence, would generate responses structuring language in that way. This, coupled with the training or settings a generative AI tool has in terms of user interaction, can further tailor the output based on user expectations. I hear from students frequently that AI-generated questions I provide during tutoring sessions are much harder than the ones they get when they use the same models, but that is an exploration for another day.
Where This Mirrors Student Thinking
Predicting the next word in a sequence is not the same operation as reasoning through a mechanism, but the strategies a novice uses to make initial sense of physiology is not that different. When a concept is not understood, terminology gets applied on the surface, for both LLMs and students. The outputs and answers end up incorrect in the same ways.
Every A&P teacher has heard a student explain a process by saying the body "wants" something. Teleological reasoning is an embedded psychological tendency, and some researchers argue high school and college students students cannot fully abandon it. What students can develop is the metacognitive skill to notice when they use teleologically reasoning and pivot to think mechanistically instead. That's where the human and the language model diverge. Current LLMs have some limited capacity to monitor their own outputs, but not in a way that catches domain-specific conceptual errors like teleological framing. Students, with the right instruction, can.
Implications for Teaching Practice
Can generative AI help students learn physiology?
If it cannot be a physiology expert, can it be a reliable tool for students to use to develop their own understanding? The first challenge is that generative AI reflects back what students already know. Thinking in terms of a zone of proximal development, this does not scaffold for students the next step towards mastery (which, if anyone has examples of generative AI developing this scaffold well, I would love to hear more, so please reach out). The second challenge is AI sycophancy. My own chat history (from a time when I was less AI-literate and trying to correct the model) shows illustrates this failure. ChatGPT did not evaluate the output for accuracy when I pushed back. It moved toward whatever answer I seemed to want. This is a dangerous thing for a student who would not know what is false in an output, especially without an understanding of how to prompt a generative AI tool to double-check.
This is where effective teaching bridges the gap. Generative AI cannot distinguish physiological language from its causal structure, or reliably catch its own teleological framings. Students, given the right scaffolding, can.
Understanding the source of the error can help us design better solutions.
While often we consider the high vocabulary load with the anatomy side of A&P, physiology carries a heavy vocabulary load as well. However, fluency with the vocabulary is not the same thing as understanding the physiology. ChatGPT had all most of the right words and arranged them in fluent-sounding sentences, but the output lacked a correct explanation. Students can do exactly the same.
Research on biology vocabulary illustrates the trap. When students self-assess their understanding of biology terms, they are least accurate on the words that sound most familiar (e.g., words like "fitness," "adaptation," "energy," and "theory"). These words carry technical meanings that diverge from everyday use. Students are more accurate at judging their grasp of unfamiliar molecular terms, because unfamiliarity signals that this is new vocabulary requiring careful learning. Familiarity is the confidence trap.
Helping students develop metacognition addresses both the traps of language familiarity and teleological reasoning. Rather than only correcting misconceptions when they appear on paper, teaching practice can build in structured moments where students evaluate the gaps in their reasoning. Exercises that ask students to translate between everyday language, precise scientific terminology, and the underlying mechanisms put the metacognitive move into practice. Confronting a teleological explanation requires helping the student notice when their own reasoning has slipped from cause and effect into purpose language, and giving them the practice to catch it in themselves.
Try In Your Classroom
I utilize Writing-to-Learn (WTL) as formative assessment in my own teaching. When a student put an idea into their own words before they reach for the vocabulary, both of us get a clearer picture of what they actually understand. Here is one way to incorporate such an exercise into any organ system unit.
Ask students to explain a physiological process at three levels:
1. Explain Like I'm 5 (ELI5). Write an explanation a young sibling would understand. No jargon allowed. Instead, explain the concept using everyday language.
2. Explain to your teacher. Now use the correct vocabulary. Include the structures, molecules, and cause-effect relationships expected in an A&P class.
3. Mechanism audit. Check both the ELI5 and Explain to your teacher drafts. Underline every place you gave the body/a structure a wish, a goal, or a purpose. Check for terms like wants, tries to, needs to, or decides to. Then, for each one, rewrite it as a cause-and-effect statement.
A few notes about this exercise.
Don't skip the ELI5 step. It looks unscientific, but the point of this step is to check that students understand the concept before adding the jargon. Students who go straight to the vocabulary version can parrot back the expected words on a spectrum of accuracy. Students who write both and then audit see the gap in their own reasoning.
Expect discomfort in the audit step. Some students will realize their teacher version also used teleological framing, writing things like "the hypothalamus wants to bring the temperature down" without noticing. Good. That's the moment where they see that the technical vocabulary did not eliminate purpose reasoning; it just dressed it up. Developing metacognition is uncomfortable, but this revision step gives students practice translating purpose statements into mechanisms.
If incorporating AI literacy into the curriculum, an extension of this exercise could include evaluating an AI-generated explanation of the same topic. Ask students to evaluate how the mechanisms are explained, or to compare the output against a textbook. In the output, they will likely find teleological framings, purpose language, and moments where the AI sounds confident without describing a cause. The goal: the same tool students use to diagnose their own reasoning becomes the tool they use to evaluate generative AI outputs.
