💻 computer science

Self-Regulated Personal Contracts as a Harm Reduction Approach to Generative AI in Undergraduate Programming Education

This study evaluates a non-binding, self-regulated GenAI contract in an undergraduate programming course, finding that while it successfully fostered self-awareness and accountability for many students, the cognitive burden of sustaining intentional guidelines often led to abandonment under academic pressure, highlighting the challenges of aligning student agency with the ease of unrestricted AI use.

Original authors: Aadarsh Padiyath, Jessica Shen, Barbara Ericson

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aadarsh Padiyath, Jessica Shen, Barbara Ericson

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are learning to cook. You have a new, incredibly powerful kitchen assistant (Generative AI) that can chop vegetables, mix sauces, and even plate the dish in seconds. The problem is, if you let the assistant do everything, you never actually learn how to cook. But if you ban the assistant entirely, you might burn out trying to chop onions for three hours when you're already tired.

This is the dilemma facing students learning to program today. They have access to AI tools that can write code instantly, but they struggle to decide when to use them and when to do the work themselves.

This paper describes an experiment where researchers tried to solve this not by banning the AI, but by giving students a "Personal Contract." Think of this not as a legal document signed in blood, but more like a fitness plan or a dietary pact you make with yourself.

Here is how the experiment worked, broken down simply:

The Setup: A "No-Judgment" Zone

The researchers took 217 students in an intermediate Python programming class. Instead of saying "No AI allowed" or "Use AI however you want," they asked the students to write their own rules.

The contract had four main steps over the course of a semester:

The Goal: "What do I actually want to get out of this class?" (e.g., "I want to understand how loops work," not just "I want an A").
The Plan: "Here is my personal rule for using AI." (e.g., "I will only ask AI to explain an error message if I've been stuck for 30 minutes," or "I will never let AI write the whole function for me.")
The Check-in: Halfway through the term, students looked back. "Did I follow my rules? Did my rules help me learn?"
The Final Review: At the end, they reflected on what they learned about themselves.

Crucially, this contract was not enforced. No one was punished for breaking their own rules. The only requirement was to fill out the paper. It was designed to be a mirror, not a police officer.

What Worked: The "New Driver" Effect

The results showed that this approach worked very well for students who were new to managing AI.

The Analogy: Imagine a new driver who has never driven before. They don't know the rules of the road yet. If you give them a map and a checklist of "When to stop, when to go," they become much safer drivers.
The Result: About 58% of the students said the contract changed how they thought. They realized they were using AI mindlessly (like a crutch) and started using it intentionally (like a tool). They created specific "If-Then" rules, like: "If I'm stuck for 20 minutes, THEN I can ask AI for a hint, but not the answer."
Why it worked: For these students, the contract acted as a scaffold (like training wheels). It helped them turn vague feelings into concrete plans.

What Didn't Work: The "Willpower Trap"

However, the contract had a major flaw. It relied entirely on the students' memory and willpower.

The Analogy: Imagine you write a diet plan on a piece of paper on Monday. By Wednesday, you are hungry, tired, and stressed. You forget the paper exists. You eat the cookie. The paper didn't stop you because it wasn't in the way of the cookie.
The Result:
- Forgetting: Many students admitted they forgot about the contract between check-ins. It was just a file on a computer they didn't look at.
- The Pressure Cooker: When deadlines loomed and students were exhausted, 11% of them admitted they abandoned their own rules. The AI offered a "fast lane" to finish the homework, and the "slow lane" of learning felt too hard. Even though they knew they shouldn't use the AI, the pressure was too strong.
- The "Already Good" Students: About 25% of students already had their own strict rules. For them, the contract was just a formality. They didn't need the training wheels; they were already riding the bike.

The Big Takeaway

The paper concludes that awareness alone isn't enough.

Knowing you should eat vegetables doesn't stop you from eating cake when you're stressed. Similarly, knowing you should learn to code doesn't stop you from using AI when you're tired and behind on homework.

The "Personal Contract" was a great tool for helping students start thinking about their choices. But it couldn't force them to stick to those choices when life got hard. The researchers suggest that for this to work long-term, we might need to change the environment (like putting the cookies in a locked box) rather than just relying on the student's willpower.

In short: The contract helped students realize they had a choice, but it couldn't always help them make the hard choice when the clock was ticking.

Technical Summary: Self-Regulated Personal Contracts as a Harm Reduction Approach to Generative AI in Undergraduate Programming Education

Problem Statement

Undergraduate programming students face a constant stream of implicit decisions regarding the use of Generative AI (GenAI) tools like ChatGPT and Copilot. These decisions are frequently driven by external pressures—such as deadline stress, peer behavior, and career anxieties—rather than explicit alignment with personal learning goals. While instructors often adopt binary policies (total prohibition or unrestricted acceptance), these "one-size-fits-all" approaches fail to address the diverse motivations and skill levels of the student population. Furthermore, the frictionless affordances of GenAI often encourage passive, productivity-focused usage that undermines deep learning and self-efficacy. The core problem is a lack of student agency and self-regulatory capacity to navigate the tension between the tool's ease of use and the cognitive effort required for genuine skill acquisition.

Methodology

The authors implemented a "GenAI Contract" intervention grounded in Self-Regulated Learning (SRL) theory and Harm Reduction (HR) philosophy within an intermediate-level Python course ( $N=217$ ) at a large public research university.

Theoretical Framework: The design operationalizes Zimmerman's three-phase SRL model (forethought, performance, self-reflection) and implementation intention theory (if-then planning). The Harm Reduction approach acknowledges that students will use GenAI regardless of policy; thus, the goal shifts from abstinence to supporting intentional, informed decision-making.
Intervention Design: The contract was a non-binding, unenforced reflection tool graded solely for completion. It consisted of four prompts distributed over an eleven-week semester:
1. Week 2 (Steps 1 & 2): Students articulated specific personal learning goals and created "if-then" usage guidelines (e.g., "I will use GenAI only after 30 minutes of independent struggle").
2. Week 5 (Step 3): Post-Midterm 1 reflection on actual usage patterns, alignment with goals, and guideline adjustments.
3. Week 11 (Step 4): Post-Midterm 2 reflection on the semester's usage, confidence levels, and final planning statements for future learning.
Data Collection: Analysis relied on qualitative review of student contract submissions and self-reported data. The research team identified patterns in engagement, goal alignment, and behavioral adherence without calculating inter-rater reliability, positioning the work as an experience report.

Key Contributions

Operationalizing Harm Reduction in CS Education: The paper demonstrates a practical application of harm reduction principles in programming education, moving away from policing behavior toward scaffolding student agency and metacognition.
Differentiated Impact of SRL Interventions: The study provides evidence that self-regulation scaffolds are most effective for students initiating new behaviors (novices without established GenAI habits) but less effective for modifying existing strong habits, whether those habits are productive or unproductive.
Identification of the "Willpower Gap": The research highlights a critical limitation in purely reflective interventions: awareness does not automatically translate to sustained behavioral change. Students frequently abandoned their self-set guidelines when faced with time pressure, revealing that self-regulation alone is insufficient against the low-cost pull of GenAI's default affordances.

Results

High Engagement: Completion rates were high (97% for initial steps, 85% for the final reflection), indicating students were willing to engage with the reflective process.
Behavioral Shifts: 58% of students reported that the contract changed their thinking, helping them establish accountability structures and make more deliberate choices.
Habit Persistence: 25% of students arrived with established GenAI practices (either highly restricted or highly permissive) and found the contract confirmatory rather than transformative.
Attrition of Guidelines: 13% of students forgot about the contract between checkpoints, and 11% explicitly abandoned their guidelines under deadline pressure.
Student Feedback: Students suggested that the intervention failed to integrate into their actual workflow. They requested more frequent check-ins (weekly vs. midterm), concrete examples of responsible use, and technical scaffolding (e.g., AI tools that enforce learning steps) to reduce the cognitive load of self-regulation.

Significance and Claims

The paper claims that while self-regulated personal contracts are a valuable tool for making implicit choices explicit and supporting students who lack established GenAI strategies, they are not a panacea. The authors argue that intentional use of GenAI requires more than just self-awareness; it requires environmental and structural support.

The significance of this work lies in its modest conclusion: for many students, the "most agentic choice" may be deliberate non-use of GenAI, rather than a struggle to use it "correctly" against the tool's design. The authors suggest that future educational interventions must move beyond individual reflection to include technical and pedagogical designs that reduce the burden of self-regulation, potentially by embedding learning constraints directly into the tools themselves. The study serves as a call for educators to recognize the limits of willpower-based interventions in the face of highly accessible, productivity-optimized AI tools.