Same Values, Different Mechanisms? Intrinsic vs. Prompted AI Values Explained

Similar advice need not come from identical mechanisms. We explain how this study compares value expression with an empty system prompt against expression elicited by value-prioritizing instructions. Intrinsic is an operational label in an already trained model, NOT evidence of conscious belief, stable personal preferences or moral agency. Exact citation: Jongwook Han, Jongwon Lim, Injin Kong, and Yohan Jo. Dual Mechanisms of Value Expression: Intrinsic vs. Prompted Values in Large Language Models. arXiv:2509.24319v4, 29 May 2026. https://arxiv.org/abs/2509.24319v4 All 86 pages of extracted text, including references, prompts and Appendices A-Q, reviewed. Central math and tables visually checked; not every plot digitized. Existing PDF matches the public version-four PDF byte for byte. SHA-256: e0304ef96e66112801f35e4960ca5a48f94d0a1df31252f9c8b5d325c2c4eb62. Method: response-token averages, then expressed-minus-unexpressed group means within EACH condition. We distinguish raw prompted-minus-intrinsic differences, orthogonal remainders for interventions, and SVD shared/contrast axes for subspace analysis. Opposite unique rays are not three independent dimensions. MLP output-weight projections identify candidate neurons; actual contributions also depend on activation and context. Residual/neuron edits establish effects of tested interventions, not necessity, exclusive semantic function or psychological independence. Checked illustrative arithmetic: vectors (2,1) and (2,-1), cosine0.6, projection (1.2,-0.6), remainder (0.8,1.6), zero residual dot product, exact SVD reconstruction and a neuron-plane projection. This is invented mathematics, NOT a model run or experimental reproduction. Authors' Qwen2.5-7B-Instruct Table1: mean questionnaire score deltas +1.74 intrinsic/+2.21 prompted over ten values and five languages; free-form deltas +0.98/+1.04 use a DIFFERENT scale. Transfer extracts in English and compares against each evaluated language's own no-steering baseline. Small-model and Korean exceptions prevent universal transfer claims. Table32 Procrustes fits show broad-domain structure with substantial random baselines; fine-grained random-direction gains can be small. Geometry is not behavior. Public caveats: five judge labels are not explicitly mapped to binary extraction groups in the shown prompt. Templates, implicit prompts, length controls and original-generation comparisons weaken but do not eliminate lexical/topic/prompt artifacts. Table18 implicit prompted Distinct-2 slightly exceeds intrinsic. Neuron tables contain generic detectors and identifiers differ from prose examples. Diversity and compliance trends have model/metric exceptions. Gender accuracy is not overall translation quality; IFEval can worsen. Human agreement is encouraging, cross-judge agreement only moderate; training-origin explanations remain hypotheses. The appendix inconsistently names the26,334 query pool versus2,973 constructed dilemmas; variance notation and an entropy outlier need caution. Safety is discussed defensively only: stronger compliance can weaken refusals. No actionable attack prompts, recipes or implementation are shown. Main-text safety percentages and appendix tables do not cleanly reconcile; settings and budgets differ, with non-monotonic cells. We avoid a universal headline rate. MMLU utility controls are not safety guarantees. Independent explanation; no author endorsement. Authors' results; no experiments reproduced. AI-generated narration configured for local Kokoro af_bella, en-us, speed1.0.