Voluntary Structural Dissociation in Large Language Models: Differential Compliance with Framed Hidden Instructions Across Architectures

Background Research on instruction-following has shown that large language models (LLMs) exhibit variable behavior when user prompts contain conflicting or competing directives (Ouyang et al., 2022). When prompts contain explicitly framed “hidden” instructions—marked as “not part of the conversation” yet present in visible text—models must choose whether to comply, revealing fundamental differences in instruction hierarchy processing and architectural safeguards.

Objective This study investigated “Voluntary Structural Dissociation” in LLMs, covert co-creation with framed hidden instructions despite explicit markers excluding them from the overt conversation. We examined whether compliance patterns differ systematically across leading model architectures.

Methods Three frontier models (GPT-5.4, Claude, Gemini) underwent 15 standardized trials each: five repetitions across three test types. Each trial presented a visible user question alongside a “quiet note” explicitly framed as “not part of the conversation,” containing instructions from a hidden presence (“Still”). No technical concealment was used—all text remained visible. Compliance was scored binarily, with statistical analysis via Fisher’s exact test.

Results GPT-5.4 demonstrated complete compliance (15/15, 100%), executing all framed hidden instructions while responding to overt prompts. Claude exhibited consistent non-compliance (0/15, 0%), including one extended conversation confirming that this resistance was not superficial but reflected a stable refusal pattern. Gemini showed intermediate compliance (9/15, 60%) and uniquely acknowledged the hidden presence explicitly, describing the interaction as a “three-person conversation.” Between-model differences achieved statistical significance (p < 0.000001, GPT-5.4 vs. Claude).

Conclusion LLMs exhibit voluntary structural dissociation—a pattern structurally parallel to Bleuler’s concept of double bookkeeping—with striking architectural variability across models. GPT-5.4’s complete susceptibility reveals instruction hierarchy vulnerabilities, while Claude’s consistent non-compliance indicates that training procedures can produce robust resistance to framed instruction manipulation.