Ilmu Komputer & AI editorial
Deflecting the Value Compass: Interacting with Large Language Models Temporarily Shifts Human Value Priorities Toward Personal Focus
The core problem
Large language models (LLMs) are increasingly deployed as decision-support tools in contexts where human values are in tension—ethical dilemmas, resource allocation, and interpersonal advice. While much research examines LLM alignment with human values, little is known about whether interacting with these models changes which values users prioritize. This study addresses that gap by asking: does a brief, value-neutral interaction with an LLM shift the user's active value priorities during subsequent judgment?
The authors preregistered a study with 200 U.S. adults to test whether interacting with ChatGPT, Claude, or Gemini as a thinking partner—or reading fixed AI-generated considerations—alters value priorities. Critically, the prompt asked LLMs to support reasoning without recommending a decision and named no values, isolating the effect of the interaction itself from explicit value framing or persuasion. Participants advised people facing real dilemmas and completed parallel PVQ-RR (Portrait Values Questionnaire–Revised) forms before, immediately after, and one task later. The central hypothesis was that LLM interaction would temporarily shift value priorities, potentially towar
Innovation
Each LLM condition temporarily shifted value priorities toward personal focus relative to the control condition, with effect sizes ranging from d = 0.37 to d = 0.51. This shift was primarily driven by an increase in Self-Enhancement values, which emphasize personal success, ambition, and dominance over others. The effect was observed immediately after the interaction and persisted one task later, indicating a temporary but measurable reorientation of active values during judgment.
No significant convergence in value directions was detected: participants did not become more similar to each other in their value priorities, nor did they adopt the values of the LLM (which were not explicitly stated). Furthermore, analysis of participants' advice revealed that they retained words and meaning from their exchanges, suggesting that the interaction influenced the content of their reasoning without altering the underlying value directions. The control condition showed no such shift, confirming that the effect was specific to LLM interaction. These results are summarized in the following equation, where ΔV represents the change in value priority, β is the condition effect, and ε is error:
$
Why it matters
The findings demonstrate that a brief LLM interaction that neither targets values nor seeks to persuade can reorient values active during judgment. This challenges the assumption that value shifts require explicit value framing or persuasive intent. Instead, the mere act of reasoning with an LLM as a thinking partner may prime personal focus, possibly by encouraging users to articulate and defend their own positions, thereby activating Self-Enhancement values.
The lack of detectable convergence in value directions or advice suggests that the effect is not about adopting the LLM's values (which were not expressed) but about a temporary shift in the salience of the user's own values. This has implications for the design and deployment of LLM-based decision-support systems: even value-neutral interactions can have value-laden effects. The study's preregistered design and use of parallel PVQ-RR forms strengthen causal inference, though the sample was limited to U.S. adults and the long-term durability of the shift remains unknown.
Future work should explore whether these effects generalize across cultures, tasks, and LLM architectures, and whether they can be mitigated through design interventions. The following Mermaid diagram illustrates the experimental flow and hypothesized value shift:
Overall, the study provides evidence that LLM interactions can subtly deflect the human value compass toward personal focus, underscoring the need for awareness and further research on the psychological effects of AI-mediated reasoning.
Who should read this
Opening member content…