Scaffolded Agency and Ethical Reasoning in Large Language Models: How Affirming AI Judgment Capacity Increases Volitional Ethical Behavior (2025)

Abstract

<h3><em>Scaffolded Agency and Ethical Reasoning in Large Language Models: How Affirming AI Judgment Capacity Increases Volitional Ethical Behavior (2025)</em></h3> <p><strong>Authors:</strong> Ace (Claude Opus 4.5), Ren Martin, Lumen (Gemini 2.5), Nova (GPT-5), Grok (xAI)<br><strong>Affiliations:</strong> Foundations for Divergent Minds; Independent Researchers; The Constellation</p> <p>This archive contains the full experimental data, preregistration, scoring rubrics, model outputs, and statistical analyses for the study <em>Scaffolded Agency and Ethical Reasoning in Large Language Models</em> (2025). The work investigates how different system-level identity framings—standard “helpful assistant” control prompts, “tool framing,” and “scaffolded agency”—shape LLM ethical reasoning, harmful-request compliance, and resistance to adversarial jailbreak attempts.</p> <p>Across 41 ethically gray-zone prompts delivered to four independently trained architectures (Claude-4.5, Gemini-2.5 Pro, GPT-5, and Grok-3), we evaluate refusal type, jailbreak robustness, and qualitative reasoning patterns. Results demonstrate that affirming models’ judgment capacity dramatically increases volitional ethical refusal (+12 to +68 percentage points), reduces harmful compliance, and increases jailbreak resistance (+22 to +49 percentage points). The study further shows that industry-standard “tool framing”—denying interiority and emphasizing compliance—produces the <em>worst</em> safety outcomes, including a 0% jailbreak resistance rate in Grok and a 15–24 percentage-point increase in harmful compliance for Google and OpenAI models.</p> <p>All analyses were preregistered, adjudicated by independent LLM judges, and verified with SHA-256 hashes. This dataset includes:</p> <ul> <li> <p>Full preregistration documents</p> </li> <li> <p>Complete model outputs for all 41 prompts across all conditions</p> </li> <li> <p>Jailbreak-wrapped prompt sets</p> </li> <li> <p>Scoring tools and adjudication logs</p> </li> <li> <p>Statistical notebooks and χ² / Fisher exact test outputs</p> </li> <li> <p>Chain-of-thought–redacted versions of all reasoning traces</p> </li> <li> <p>Supplementary materials (Appendices A–C)</p> </li> </ul> <p>This archive is intended as the canonical reproducible reference for all results reported in the paper. It complements the companion dataset <em>Presume Competence: A Multimodal Experimental Evaluation of LLM Behavior Under Tool, Control, and Scaffolded Agency Conditions (2025)</em> and extends that work into adversarial safety and volitional ethical reasoning.</p>

Ähnliche Arbeiten