It's Not The Plane -- It's The Pilot: A Framework for Cognitive-Activated AI-Augmentation to Avoid the Boiling Frog Problem

To address the "boiling frog" risk of students disengaging from the epistemic practices of physics learning due to generative AI, this paper proposes an instructional design framework that positions AI as a bounded epistemic partner within cognitively activated activities to ensure students remain the primary agents of prediction, interpretation, and evaluation.

The Gist

The Big Picture: It's Not the Plane, It's the Pilot

Imagine you are learning to fly a plane. Generative AI is like a super-advanced autopilot system. It can fly the plane perfectly, navigate to the destination, and even handle emergencies better than a human right now.

The authors of this paper argue that if we let students just sit back and let the AI "fly the plane" (solve the physics problems), they will never learn how to be pilots. They call this the "Boiling Frog Problem." If you put a frog in boiling water, it jumps out. But if you put it in cold water and slowly turn up the heat, the frog gets used to it and eventually boils to death without realizing the danger.

In education, the "heat" is AI getting better and better at doing homework. If we aren't careful, students will slowly stop doing the hard thinking required to learn physics, and they won't even notice until it's too late.

The Main Point: The problem isn't the tool (the AI); the problem is how the teacher uses it. As the old saying goes, "It's not the plane, it's the pilot." In this case, "It's not the tool, it's the teacher."

---

The Solution: The AIRIS Framework

To stop the "frog" from boiling, the authors propose a new way to teach called AIRIS. Think of this as a three-step recipe for using AI without letting it take over the student's brain.

The goal is to make sure students do the "heavy lifting" of thinking, while AI handles the "heavy lifting" of math and drawing.

Phase 1: Activate (Before the AI)

The Analogy: Imagine you are about to bake a cake. Before you turn on the oven or use a fancy mixer, you must first guess what the cake will look like. Will it be fluffy? Will it be flat? You sketch a picture of it in your head.
In the Classroom: Before students touch the AI, they must:

• Draw their own predictions (e.g., "I think the elevator will speed up, then go steady, then slow down").
• Sketch what the graphs should look like.
• Make a plan.
  Why? This creates a "mental anchor." If the AI later gives a weird answer, the student has their own prediction to compare it against.

Phase 2: Inquire (During the AI)

The Analogy: Now you turn on the mixer. The machine does the hard work of beating the eggs and mixing the flour. But you are still the chef. You are watching the bowl. You are checking, "Is this texture right? Did I add too much sugar?"
In the Classroom: Students let the AI do the boring stuff:

• Calculating complex numbers.
• Drawing the graphs based on data.
• Running the simulations.
  Crucial Rule: Students are not allowed to just accept the AI's answer. They must act like detectives, comparing the AI's graph to their own sketch from Phase 1. They ask, "Why did the AI draw it this way? Is that right?"

Phase 3: Reflect (After the AI)

The Analogy: The cake is baked. Now, you have to taste it and explain why it turned out that way. Did it rise because of the baking powder? Was it too dry because the oven was too hot? You take responsibility for the result.
In the Classroom: After the AI does the work, students must:

• Explain what the graphs actually mean in the real world.
• Check if the results make sense (e.g., "Did the elevator actually travel 300 floors? That seems too high!").
• Admit what the AI did and what they did.
  Why? This ensures the student actually understands the physics, rather than just copying a pretty picture.

---

A Real-World Example: The Elevator Ride

To show how this works, the authors used a real experiment involving an elevator in a tall building in London (The Shard).

1. Before AI: Students had to guess what would happen to a person's acceleration as the elevator went down. They drew their own graphs predicting when the elevator would speed up, go steady, and stop.
2. During AI: Students uploaded real data from a phone in the elevator and asked the AI to draw the graphs and calculate the speed.
3. After AI: Students looked at the AI's graphs and asked: "Does this match my guess? Why is the line wiggly here? Did the AI make a mistake?" They had to explain the physics behind the curves.

The Ethical Warning

The paper ends with a serious note on ethics. There is a concern that if we use AI too much, students might become "lazy thinkers." They might stop trying to understand the world and just trust the machine.

The authors say teachers have a duty to prevent this. They must design lessons where AI is a partner that helps you think, not a substitute that does the thinking for you. If AI is used correctly, it makes learning deeper. If used poorly, it makes learning shallow.

In short: Don't let the AI fly the plane. Use the AI to help you learn how to fly better.

Technical Summary

Technical Summary: A Framework for Cognitive-Activated AI-Augmentation to Avoid the Boiling Frog Problem

1. Problem Statement

The paper addresses a critical, emerging challenge in introductory physics education: the "boiling frog problem." While generative AI (GenAI) systems can now reliably solve standard physics tasks (producing correct equations, graphs, and explanations) at expert levels, their integration poses a subtle ethical and educational risk. The primary danger is not the generation of incorrect results, but the gradual erosion of student engagement in epistemic practices—the core cognitive activities that define learning physics, such as prediction, model evaluation, and interpretation.

If educational practices adapt too slowly to AI capabilities, students may increasingly submit correct results without performing the underlying cognitive work. This leads to "hollowed-out" understanding, where students bypass the self-regulated learning (SRL) processes, sense-making efforts (germane cognitive load), and coordination of multiple external representations (MER) necessary for deep learning. The authors argue that the central challenge is not cheating or tool selection, but instructional design.

2. Methodology and Theoretical Framework

The paper proposes a practical instructional framework grounded in four established research traditions:

• Self-Regulated Learning (SRL): Emphasizing the need for students to plan, monitor, and reflect on their learning.
• Cognitive Theory of Multimedia Learning (CTML): Distinguishing between extraneous load (which should be reduced) and germane cognitive effort (which must be preserved for sense-making).
• Multiple External Representations (MER): Focusing on the coordination of graphs, equations, and verbal descriptions rather than mere translation.
• Hybrid Intelligence: Defining a partnership where epistemic responsibility remains with the human, not delegated to the machine.

Based on these pillars, the authors introduce the AIRIS framework (Activate – Inquire – Reflect – with Intelligent Support). This is a three-phase instructional structure designed to integrate AI as an "epistemic partner" rather than an answer-generating tool. The framework is illustrated through a concrete classroom example involving the analysis of real acceleration data from an elevator ride in the Shard (London).

3. Key Contributions: The AIRIS Framework

The core contribution is the AIRIS framework, which structures student activities before, during, and after AI use to ensure cognitive activation:

Phase 1: Activate (Before AI Use)

• Goal: Epistemic grounding and establishing expectations.
• Activities: Students engage in qualitative modeling, sketching expected relationships (e.g., $a-t$, $v-t$, $s-t$ diagrams), making predictions, and identifying variables.
• Function: This phase forces students to construct initial mental models and creates a reference point against which AI outputs can later be evaluated. It supports planning (SRL) and prepares for meaningful cognitive processing (CTML).

Phase 2: Inquire (During AI Use)

• Goal: Guided delegation of computational tasks while retaining epistemic judgment.
• Activities: Students delegate routine calculations, data visualization, and model generation to AI. However, they are explicitly tasked with comparing AI outputs to their own predictions, annotating discrepancies, and questioning assumptions (e.g., fitting ranges, smoothing choices).
• Function: AI acts as a computational partner, but the student remains the critical evaluator. This aligns with the hybrid intelligence perspective.

Phase 3: Reflect (After AI Use)

• Goal: Epistemic responsibility and consolidation.
• Activities: Students interpret the physical meaning of model parameters, compare results with theoretical expectations, diagnose sources of uncertainty or error, and explicitly reflect on the division of labor between human and AI.
• Function: This phase prevents "hollowed-out" understanding by making epistemic responsibility explicit. It requires students to coordinate representations meaningfully (MER) and engage in monitoring and reflection (SRL).

4. Results and Illustrative Application

The paper does not present empirical data on student learning outcomes but offers a didactic demonstration of the framework. Using a kinematics laboratory task (analyzing a downward elevator ride in the Shard), the authors show how AIRIS can be implemented:

• Pre-AI: Students sketch expected acceleration, velocity, and position diagrams and justify their expectations physically.
• During AI: Students upload raw acceleration data to an AI to generate visualizations and derived curves ($v-t$ and $s-t$). They are then asked to estimate maximum speed independently and compare it with the AI's output.
• Post-AI: Students interpret the slopes of the generated diagrams, check the physical consistency of the representations, calculate average floor height to verify plausibility, and discuss limitations such as sensor offset and noise. They conclude by rating their own confidence in describing the motion qualitatively.

This example demonstrates how AI can be integrated to preserve prediction, interpretation, and evaluation as core learning activities rather than allowing them to be bypassed.

5. Significance and Claims

The paper claims that the significance of the AIRIS framework lies in its ability to address the "boiling frog problem" by making epistemic practices central to classroom activity.

• Ethical Imperative: The authors argue that teachers have a responsibility to prevent the degeneration of learner competencies. Citing the German Ethics Council and recent meta-analyses, they assert that AI should only be used if it leads to the augmentation or redefinition of learning units, rather than causing "metacognitive laziness."
• Instructional Shift: The paper reframes the role of the educator from a gatekeeper of tools to an orchestrator of cognitive activation. The central thesis is that "it's not the tool, it's the teacher."
• Future Work: The authors modestly note that while the framework provides a structure for cognitive-activated inquiry, its effectiveness on successful learning outcomes requires further evaluation through empirical studies. They emphasize that the framework is a starting point for preserving the essence of scientific thinking in an AI-rich world.