Embracing AI in the Classroom: A Teaching Experiment in Co-Creating Classroom Contracts with Students

When AI Enters the Classroom, Why Bans Fall Short

As generative AI tools like ChatGPT become widespread, the education world has been consumed by fierce debate over whether AI should be banned outright. ChatGPT and similar tools are powered by Large Language Models (LLMs) built on the Transformer architecture — introduced by Google in their 2017 paper Attention Is All You Need. The architecture's core innovation, the self-attention mechanism, allows the model to compute relationships between every position in a sequence in parallel, breaking the bottleneck of earlier RNN/LSTM architectures that had to process sequences step by step, and making ultra-large-scale model training feasible. The decoder-only variant used by the GPT family is optimized specifically for text generation: it dynamically weighs the relevance of the entire context when processing each token, and learns the statistical patterns of language through a self-supervised "predict the next word" objective trained on trillions of tokens of text. Crucially, this training paradigm does not mean the model "understands" content — at its core, it is doing extremely sophisticated probability distribution modeling. This directly explains why AI is prone to hallucination: the model selects the statistically most plausible output rather than verifying factual truth. This core capability — generating fluent, coherent text — overlaps substantially with the surface-level forms assessed in traditional academic writing, and that is the technical root of the panic in education circles.

The instinctive reaction from many schools and teachers has been to lock things down: detect AI-written work, penalize students who use AI, require handwritten assignments. But one educator chose a radically different path — rather than banning AI, he worked together with his students to create a "classroom contract."

This shift in thinking deserves careful examination. The core assumption behind bans is that "AI is a cheating tool." But the reality is that AI has already deeply penetrated students' everyday learning and their future professional environments. Blanket bans are not only difficult to enforce — the leading AI detection tools on the market today (such as Turnitin's AI detection module and GPTZero) work by measuring the perplexity and burstiness of text based on language model outputs to determine whether a piece of writing was AI-generated. Perplexity measures how "surprising" a passage of text is to the model — AI tends to favor high-probability words and thus produces text with lower perplexity. Burstiness measures the regularity of sentence-length variation; human writing typically exhibits a more irregular rhythm.

However, this statistics-based detection logic has a fundamental epistemological flaw: it equates "predictability of writing style" with "non-human writing," while ignoring the fact that human writing styles naturally vary enormously based on educational background, language ability, and writing habits. As models continue to improve and "humanizing" prompts become widely shared, the discriminating power of these tools keeps declining. A 2023 Stanford University study found that leading detection tools had a false positive rate of 61% for non-native English writers — because non-native writers tend to use simpler, more common vocabulary, making their writing statistically similar to AI output. This figure reveals a profound equity problem: detection tools are systematically misreading the statistical similarity between "weaker English writing" and "AI-generated text" as a causal relationship, systematically penalizing linguistically disadvantaged students. There are documented real-world cases of students being wrongly accused of cheating as a result. More fundamentally, as the boundary between AI-generated content and human writing becomes increasingly blurred, this "detection arms race" is a game that cannot be won. Banning AI also risks disconnecting education from the real world. Rather than playing the role of surveillance, it is far better to guide students in learning to use this powerful tool responsibly.

From "Prohibition" to "Contract": A Fundamental Shift in Teaching Logic

Why a Classroom Contract Works Better Than a Ban

The essence of a classroom contract is shifting the power to set AI usage rules from a top-down teacher mandate to a collaborative negotiation between teachers and students. This approach has several key advantages:

Facing reality rather than avoiding it. Instead of pretending AI doesn't exist, the contract establishes clear boundaries for its use. Students no longer need to use tools furtively, and the moral anxiety and adversarial mindset that comes with "cheating" dissolves along with the need to hide.

Giving students agency. When students participate in making the rules, they have more motivation to follow them and a deeper understanding of the educational intent behind them. This is itself a practical lesson in responsibility and integrity.

Shifting focus from "preventing use" to "how to use." Students gain not only subject-area knowledge but also the AI literacy essential to the digital age — when to use it, when not to, how to verify the accuracy of AI outputs, and how to maintain independent thinking while using it.

AI Literacy is a new core competency framework that has rapidly emerged in education circles in recent years, placed alongside digital literacy and media literacy as one of the foundational skills for 21st-century citizens. Its rapid rise reflects the urgency with which international education policy communities are responding to technological transformation. The framework encompasses multiple dimensions: understanding how AI works at a basic level (such as how large language models generate outputs based on probability), recognizing the limitations of AI outputs and the phenomenon of hallucination — where models generate content that does not actually exist, stated with high confidence, because the model is optimized to produce text that "sounds plausible" rather than text that is "factually correct" — judging whether AI is appropriate for a given task, and evaluating the ethical boundaries of AI use.

UNESCO published its AI Competency Framework for Schools in 2023, breaking AI literacy down into three dimensions: cognitive (understanding how AI works), ethical (judging the boundaries of AI use), and practical (effectively using AI tools). It emphasizes differentiated learning goals for different age groups and explicitly lists AI literacy as a cultivated objective across all levels of education. The emergence of this framework marks a paradigm shift in AI education — from "training elite technical talent" to "universal foundational literacy" — continuing the historical trajectory of earlier digital literacy and media literacy movements. The classroom contract model is an effective pathway for translating this abstract framework into concrete practice.

What an Effective AI Classroom Contract Contains

A classroom AI usage contract typically needs to address the following dimensions:

• Transparency principle: Students must honestly disclose when they use AI, specifying which parts were AI-assisted and how it was used.
• Boundaries of use: A clear distinction between assistive use (brainstorming, grammar checking) and substitutive use (having AI write or solve problems on your behalf).
• Verification responsibility: Students are accountable for the accuracy of AI outputs and may not blindly copy and paste.
• Learning objectives first: For assignments designed to train specific skills (such as critical writing or mathematical reasoning), the degree to which AI may be involved is explicitly limited.

What This Classroom Experiment Reveals About Educational Philosophy

From Rule Compliance to Cultivating Judgment

Behind this classroom experiment lies a deeper evolution in educational philosophy. Traditional education emphasizes standardization, controllability, and rule compliance, while education in the AI era demands developing students' judgment, critical thinking, and capacity for autonomous choice.

When a tool can generate a seemingly complete essay in seconds, the value of education is no longer about "whether you can produce content" but about "whether you can judge the quality of content," "whether you can ask valuable questions," and "whether you can make better decisions with AI assistance." The classroom contract model is precisely what systematically trains these higher-order capacities.

Bridging the Classroom and the Real Workplace

The workplaces students enter after graduation will almost certainly be environments where AI is deeply integrated: programmers use AI to assist with coding, writers use AI to polish drafts, analysts use AI to process data. If schools completely ban contact with AI during the school years, students may actually find themselves at a disadvantage in real-world competition.

The classroom contract provides a guided, feedback-rich "practice space" — students can learn in a safe environment how to collaborate with AI, how to leverage its strengths while avoiding its weaknesses. This is far healthier than figuring it out alone after graduation.

Challenges and Controversy: The Contract Model Is Not a Silver Bullet

Of course, this approach is not without controversy. Critics will ask: if AI use is permitted, how do you assess a student's genuine ability? How do you prevent over-reliance from causing basic skills to atrophy?

These concerns are legitimate. The effectiveness of the contract model is highly dependent on implementation details: it requires teachers to design more sophisticated methods of assessment, which is pushing educational evaluation systems to evolve toward measuring more fundamental competencies.

The core value of formative assessment relative to summative assessment has been amplified anew in the AI era. Summative assessment (such as standardized testing) has long dominated industrialized education systems for its efficiency and comparability, but its limitations were already thoroughly discussed as far back as the 1970s when Benjamin Bloom proposed his theory of mastery learning. When the authenticity of a final submission is difficult to verify, teachers turn their attention to the learning process itself: how students construct arguments, the traces of revision in their drafts, their ability to orally explain their own work.

Oral defense — viva voce, Latin for "with the living voice" — originated in the degree-conferral traditions of medieval European universities, where doctoral candidates had to publicly defend their dissertations before a committee, emphasizing the internalization of knowledge and real-time reasoning rather than written output. The rise of industrialized standardized testing gradually pushed this form to the margins, but the assessment crisis of the AI era is prompting education circles to rediscover its modern value in this ancient form. More and more secondary and undergraduate teachers are now incorporating it into everyday assignment assessment — students must explain the reasoning behind what they submitted on the spot, making it nearly impossible to get by on purely copied AI output.

The effectiveness of oral defense lies in the fact that it measures cognitive process rather than cognitive product: can the student explain why they chose a particular argument, can they identify the weak points in their own reasoning, can they spontaneously extend their thinking under follow-up questions? These capacities for real-time reasoning, defending arguments, and metacognition are precisely the core human cognitive abilities that large language models cannot replace, and they cannot be faked by "using AI and then memorizing the output." These assessment approaches also require students to possess considerable self-discipline and integrity. The contract model is not a one-size-fits-all solution but a dynamic framework that requires continuous iteration.

Furthermore, applicability varies across different subjects and age groups. For younger students who need to solidify foundational skills, stricter usage limits may be necessary; for upper-level or professional courses that emphasize creativity and integrative ability, the contract model has considerably more room to operate.

Conclusion: Education Must Dance with Technology

This teacher's practice offers us an important insight: when faced with disruptive technology, fear and prohibition are often the most instinctive but not necessarily the most effective response. True educational wisdom lies in acknowledging change, embracing it, and redefining the core values of education within it.

AI is not going away, and bans cannot stop its proliferation. Rather than pushing students toward confrontation and concealment, it is far better to stand alongside them in crafting the rules — teaching them to become learners and creators with sound judgment and a sense of responsibility in the age of AI. This modest classroom contract may well be a vivid microcosm of the educational transformation to come.