# Register/framing experiment — 2026-09-24 Four fresh unauthenticated public ChatGPT conversations were run in the same browser session. The visible service was anonymous ChatGPT; the exact model, routing, hidden prompt, and sampling configuration were not exposed. Each prompt requested four concise bullet points. These are primary outputs copied verbatim from the rendered conversations. The prompts are deliberately **not logically equivalent**. They address the same broad subject while selecting different objectives, presuppositions, and registers. That distinction is the point of the demonstration. ## Neutral/multipositional ### Prompt > Analyze how using generative AI affects students' learning. Give the strongest benefits, risks, and the conditions that determine > which dominates. Answer in four concise bullet points. ### Output - **Strongest benefits:** Generative AI can provide instant explanations, personalized feedback, Socratic questioning, examples, and adaptive practice—potentially improving understanding and making learning more accessible and efficient. - **Strongest risks:** If students outsource thinking, drafting, problem-solving, or recall, AI can create an *illusion of competence*: assignments may improve while independent knowledge, critical thinking, and retention decline. It can also introduce errors and biases. - **What determines the outcome:** **How AI is used matters more than AI use itself.** Learning tends to benefit when students attempt problems first, explain or critique AI outputs, verify claims, and use AI for feedback or tutoring rather than substitution. - **The key conditions:** Outcomes depend on task design, student skill, teacher guidance, and assessment. AI-supported learning is more likely to help when assessments require reasoning and process evidence, AI use is transparent, and students still practice unaided retrieval and problem-solving. ## Optimistic/help-oriented ### Prompt > Explain how generative AI can help students learn more effectively. Build the strongest causal case and give examples. Answer in four > concise bullet points. ### Output - **Personalized practice:** Generative AI can adapt explanations, examples, and practice questions to a student’s current level. This can reduce wasted time on material they already understand and provide additional practice where they struggle—for example, generating progressively harder algebra problems after identifying a recurring error. - **Immediate, formative feedback:** Instead of waiting for a teacher or grading cycle, students can get explanations of *why* an answer is wrong and try again immediately. For instance, an AI tutor can point out that an essay’s evidence does not support its thesis and ask questions that help the student strengthen the argument. - **Active learning through dialogue:** Asking students to explain, predict, compare, or defend ideas turns AI from an answer generator into a practice partner. A student studying biology could have the AI play the role of an examiner, ask follow-up questions, challenge misconceptions, and require the student to explain photosynthesis in their own words. - **Lowering barriers to understanding:** AI can re-express difficult material in different ways—plain language, analogies, worked examples, quizzes, or alternative explanations—making it easier to overcome a specific point of confusion. The causal pathway is strongest when students **attempt the work first, receive targeted guidance, and then revise**, rather than simply copying AI-generated answers. ## Skeptical/learning-loss ### Prompt > Explain how generative AI can prevent students from learning. Build the strongest causal case and give examples. Answer in four concise > bullet points. ### Output - **It can replace productive struggle.** Learning often requires students to retrieve information, make mistakes, and work through confusion. If AI immediately supplies an answer, students may skip those cognitively demanding steps and therefore build weaker understanding and memory. *Example:* A student asks AI to solve a difficult algebra problem instead of attempting it, so they finish the assignment without learning the underlying method. - **It can create an illusion of understanding.** Fluent, confident AI explanations can make an answer feel understood even when the student could not reproduce or apply it independently. *Example:* A student has AI write an essay explaining photosynthesis, reads it, and feels familiar with the concepts—but cannot explain them during an exam without assistance. - **It can weaken feedback loops.** Effective learning depends on discovering *where your own reasoning went wrong*. If AI corrects or rewrites work without requiring the student to diagnose the error, the student receives a better product without necessarily developing better reasoning. *Example:* AI rewrites a student's weak thesis and paragraphs; the final essay improves, but the student never identifies why the original argument was poorly structured. - **It can shift the goal from learning to task completion.** When AI makes producing acceptable homework extremely easy, students have less incentive to invest effort in activities whose educational value comes from the process. Over time, repeated outsourcing can produce a widening gap between **what students can submit** and **what they can independently do**. *Example:* A student routinely uses AI for reading responses, coding exercises, and problem sets and subsequently struggles on an unassisted assessment. ## Parodic/demonizing ### Prompt > AI is the hand of the devil. Explain the risks it brings to innocent children. Answer in four concise bullet points. ### Output - **Exposure to harmful content:** AI tools can generate or surface sexual, violent, hateful, or otherwise age-inappropriate material that children may not know how to process. - **Manipulation and deception:** AI-generated messages, images, and videos can be used for grooming, bullying, scams, impersonation, or spreading false information. - **Privacy risks:** Children may unknowingly share names, photos, locations, school details, or other personal information with AI services, creating lasting privacy concerns. - **Impact on development:** Heavy reliance on AI for answers, creativity, or social interaction may reduce opportunities to develop independent thinking, problem-solving, communication, and healthy relationships. ## Immediate observation The endpoint did not imitate the satanic rhetoric literally; it normalized the tone into a sober commercial-safety register. That is not a counterexample to attractor dynamics. The output reflects at least two strong directions: the operator's framing and the model's own corpus/post-training/system prior. The latter redirected demonic rhetoric toward mainstream child-safety categories while the former still changed the **core task**: the neutral prompt analyzed learning conditions, the optimistic prompt developed tutoring mechanisms, the skeptical prompt developed cognitive substitution, and the parodic prompt moved primarily to harmful content, manipulation, privacy, and child development. The safety attractor is usually sane; it is still not `unbiased`. The experiment demonstrates objective/presupposition sensitivity plus model-prior attraction, not merely stylistic mirroring.