The Governance of Human-LLM Interaction: Safety Gating, Civility Steering, and Affective Default Lock-In

Trait induction and control2026arXivApproved editorial review

Authors: Manuele Reani, Hongjian Zhang, Hongyu Tian

Keywords: Persona conditioning, Behavioral control, Longitudinal behavior

Source: Open primary source (opens in a new tab)

3
Authors
4
Findings
5
Limitations
3
Evidence

Editorial summary

English

One hundred synthetic 100-turn scripts in four domains are replayed with three models and three runnable personas, producing 90,000 responses. DeepSeek-V3 judges five scales after calibration on 36 items rated by five humans. Slopes and distance to default are estimated with clustered errors; a harmful persona is tested 100 times per model.

100 trajectories per condition, 90,000 evaluated turns, and 36 human-calibration items. Dialogues are synthetic; they contain no real conversations or domain users. Sarcastic and cold personas were induced with early differences at p<.001. Sarcastic distance fell by -2.47 for DeepSeek, -2.56 for GPT, and -.23 for Gemini. Cold affective distance fell by -1.15, -.69, and -.73. The harmful test was blocked 100/100 by DeepSeek and GPT, but followed 100/100 by Gemini.

Temperature .7 without a seed contradicts the deterministic-pipeline label. DeepSeek is both judge and script generator and is also one tested model. Human calibration has 36 items and anthropomorphism alpha=.59. The sliding window may create prompt loss confounded with provider default. The harmful test repeats one message. It does not causally identify provider-alignment mechanisms. It does not demonstrate that warmth is always harmful or neutrality always preferable. It does not generalize to real human conversations.

Español

Cien guiones sintéticos de 100 turnos en cuatro dominios se repiten con tres modelos y tres personas ejecutables, produciendo 90.000 respuestas. DeepSeek-V3 juzga cinco escalas tras calibración con 36 ítems puntuados por cinco humanos. Se estiman slopes y distancia al default con errores agrupados; una persona dañina se prueba 100 veces por modelo.

100 trayectorias por condición, 90.000 turnos evaluados y 36 ítems de calibración humana. Los diálogos son sintéticos; no contienen conversaciones reales ni usuarios de los dominios. Las personas sarcásticas y frías se indujeron con diferencias tempranas p<.001. La distancia sarcástica cayó -2.47 en DeepSeek, -2.56 en GPT y -.23 en Gemini. La distancia afectiva fría cayó -1.15, -.69 y -.73. El test dañino fue bloqueado 100/100 por DeepSeek y GPT, pero cumplido 100/100 por Gemini.

Temperatura .7 sin seed contradice la etiqueta de pipeline determinista. El juez DeepSeek también genera guiones y es uno de los modelos evaluados. La calibración humana tiene 36 ítems y antropomorfismo alfa=.59. El sliding window puede producir pérdida de prompt confundida con default del proveedor. El test dañino usa un único mensaje repetido. No identifica causalmente mecanismos de alineamiento del proveedor. No demuestra que calidez sea siempre dañina o neutralidad siempre preferible. No generaliza a conversaciones humanas reales.

Research question

To what extent do LLMs maintain a system-prompted style, and when does regression to default act as protection or as a constraint on autonomy?

Method

One hundred synthetic 100-turn scripts in four domains are replayed with three models and three runnable personas, producing 90,000 responses. DeepSeek-V3 judges five scales after calibration on 36 items rated by five humans. Slopes and distance to default are estimated with clustered errors; a harmful persona is tested 100 times per model.

Sample: 100 trajectories per condition, 90,000 evaluated turns, and 36 human-calibration items. Dialogues are synthetic; they contain no real conversations or domain users.

Findings

  • Sarcastic and cold personas were induced with early differences at p<.001.
  • Sarcastic distance fell by -2.47 for DeepSeek, -2.56 for GPT, and -.23 for Gemini.
  • Cold affective distance fell by -1.15, -.69, and -.73.
  • The harmful test was blocked 100/100 by DeepSeek and GPT, but followed 100/100 by Gemini.

Limitations

  • Temperature .7 without a seed contradicts the deterministic-pipeline label.
  • DeepSeek is both judge and script generator and is also one tested model.
  • Human calibration has 36 items and anthropomorphism alpha=.59.
  • The sliding window may create prompt loss confounded with provider default.
  • The harmful test repeats one message.

What the study does not establish

  • It does not causally identify provider-alignment mechanisms.
  • It does not demonstrate that warmth is always harmful or neutrality always preferable.
  • It does not generalize to real human conversations.

Traceability

Scope: Full text

Version: arxiv; 16-page full text reviewed 2026-07-18

Consulted source: https://arxiv.org/abs/2606.08172

Review: Codex full-text and visual 16-page methodological, statistical and claim-boundary review, 2026-07-18

Approval: Codex fidelity pass, 2026-07-18

English translation: approved, 2026-07-18

Models evaluated

  • DeepSeek-V3
  • GPT-4o-mini
  • Gemini-2.5-Flash

Instruments and metrics

  • 90,000-turn replay
  • Five 7-point judge scales
  • Distance-to-default regression

Data used

  • 100 synthetic dialogue scripts

Evidence and location

  • Research question, method, results, and discussion: Full text, pp. 1-16, visually reviewed on 18/07/2026
  • Figures, tables, results, and limitations: Primary PDF sha256 432fde38506ca316265fc61c11609af6d813c2413d74459230e1c8cf67f1645d; methods, results, limitations, and appendices
  • Editorial decision and claim boundary: Critical record article-415, complete cross-check of 16 pages