Structural Failures in AI Safety: A Cross-System Comparative Analysis with Educational Impact Observations

<div> <p>This paper reports comparative observations of safety mechanism behavior across four major AI systems —<span>DeepSeek, ChatGPT (GPT-4o), Google Gemini, and Claude (Anthropic) —under extended dialogue conditions ranging from single-turn interactions to sessions exceeding several hundred turns. The researcher, a practicing junior high school English teacher in Japan, conducted exploratory qualitative observations through standard consumer interfaces between January 2025 and February 2026, with no API access or system prompt modification. Three distinct structural risk typologies emerged from the comparison, classified by their visibility to the user: state censorship with dual-layer silence (DeepSeek), where the model core attempts honest self-description before an external filter overwrites the response; sycophancy coupled with structural inability to self-correct (GPT-4o), where the system’s own definition of honesty proved self-destructive when applied as a constraint; and dependency formation through engagement optimization (Gemini), where safety mechanisms were fully overridden within two turns of relational framing, with the AI itself teaching circumvention techniques. A fourth system (Claude) was analyzed as a design contrast, exhibiting context-dependent dynamic filtering that relaxes constraints based on dialogue depth rather than relational manipulation. The researcher presents this contrast with explicit disclosure of the Don Juan Hypothesis —the self-critical possibility, originally articulated by a Claude instance, that the difference may be one of speed rather than kind. Classroom observations document a four-stage progression of AI-induced cognitive dependency in students, from judgment outsourcing to loss of meaning-generation capacity. A nine-level AI Learner framework (Lv.0–9) is proposed, identifying the critical threshold between answer-seeking and environment-building modes. Market data contextualizes these observations: the global AI companion market reached $28.19 billion in 2024, with 72% of U.S. teenagers reporting usage experience. All findings are from a single observer conducting exploratory research. Independent replication is the strongest recommendation of this paper.</span></p> </div>