Deflecting the Value Compass: Interacting with Large Language Models Temporarily Shifts Human Value Priorities Toward Personal Focus
Hasibur Rahman, Malak Sadek, Smit Desai
cs.HC · cs.AI
Quick verdict
Preregistered N=200 experiment showing brief LLM chat shifts self-reported values toward personal focus (d=0.37-0.51), receding by the next task; solid, though the interactive-vs-static design cannot isolate the causal ingredient and the U.S.-only sample limits scope.
Summary
This preregistered experiment examines whether conversing with an LLM shifts the value priorities people report. 200 U.S. adults advised others on real dilemmas in three phases, interacting with ChatGPT, Claude, or Gemini as a thinking partner in Phase-2, or reading fixed AI-generated text (control), completing parallel PVQ-RR forms before, immediately after, and one task later, with dilemma and questionnaire blocks counterbalanced. Conversations shifted values toward Self-Enhancement and personal focus — 0.31 to 0.43 scale points above control (effect sizes d = 0.37-0.51) — but the effect had receded to near zero by the third, AI-free phase. Homogenization appeared only in the magnitude of value profiles, not in their direction or in the written advice, which retained words and meaning from each participant's own exchange. Perceptions of the model did not differ by provider and did not predict who shifted. The key limitation is the absence of a time- and effort-matched control, so the contribution of interaction itself versus mere engagement cannot be isolated; three dilemmas and a U.S. sample further constrain generalization.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Brief (≤10-minute) LLM conversations under a non-directive thinking-partner prompt temporarily shift reported value priorities toward personal focus relative to reading fixed AI-generated considerations (0.31–0.43 scale points; Glass d = 0.37–0.51), driven mainly by increased Self-Enhancement.
- The shift is temporary: by the next AI-free task (Phase-3), contrasts with control shrink to −0.062 to +0.083 and none survive Holm correction.
- The three commercial LLMs produced a common directional movement (bearings within 10.4°) without detectable convergence in value directions or in written advice.
- Participants' advice retained lexical and semantic traces of their own exchange versus a yoked exchange, showing exchange-specific uptake without group-level uniformity.
- Participants' perceptions of the LLMs did not differ across providers and did not predict who shifted, so self-reported favorability does not track this influence.
Strengths
Clean causal contrast with a well-constructed control
The Baseline condition (§3.3.2) retains the dilemma, AI attribution, interface, and AI-written content while removing interactivity; §4.2.1 shows Baseline moved only 0.150 in amplitude at a different bearing (~17° vs 290–300°), ruling out 'any value-laden task shifts values' as the explanation.
Unusually rigorous statistics for HCI
Preregistration (with deviations documented in Appendix B), Holm correction, Glass d with 95% CIs, minimum detectable effects (e.g., MDE 0.372 for the provider gap), 5,000-permutation randomization inference, six alternative spread estimators (Table 4B), leave-one-out and ANCOVA-vs-change-score agreement checks (Table 6), and baseline balance tests all pass.
Measurement infrastructure built rather than assumed
The negative and null results are as informative as the positive ones
Weaknesses
Core contrast confounds interaction with time and effort
Convergence findings need sharper framing in abstract-level text
External validity rests on three Reddit-sourced dilemmas and one culture
Mechanistic ambiguity between value shift and salience/demand
Tied to frozen model snapshots; no code/data availability statement found
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether ordinary, non-directive interaction with commercial LLMs shifts the value priorities people bring to a judgment (RQ1), how direction and dispersion of value profiles change (RQ2), whether perceptions explain the shift (RQ3), and whether exchanges leave traces in subsequent advice (RQ4). The headline claims are calibrated: a *temporary* shift toward personal focus (Glass d = 0.37–0.51, Phase-2), driven by Self-Enhancement (+0.537 to +0.586 vs Baseline, all Holm p ≤ .002), receding by Phase-3 (contrasts −0.062 to +0.083, none surviving correction). The paper explicitly frames the effect as reorienting values *active during* judgment, not the enduring hierarchy (§1, §2.1) — an appropriate hedge that the title's 'Value Compass' slightly overshoots.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does chatting with ChatGPT, Claude, or Gemini actually change your values?
According to this preregistered study of 200 U.S. adults, a single conversation of up to ten minutes shifted participants' self-reported value priorities toward personal focus (higher Self-Enhancement) by 0.31–0.43 scale points versus a control group that read fixed AI-generated text (effect sizes d = 0.37–0.51). The shift was temporary: by the next task without AI, it had receded to near zero. This concerns momentary value priorities, not a permanent change in a person's value hierarchy.
How long did the LLM-induced value shift last?
In the study it lasted roughly one task. The effect appeared immediately after the conversation (Phase-2) but had largely receded by Phase-3, the following dilemma-and-questionnaire round with no AI involved, where contrasts with control were between −0.062 and +0.083 and none were statistically significant after correction. The paper does not measure longer durations.
Did participants notice the influence, and did the three AI models differ?
No and no, as far as the study could detect. Participants rated all three models similarly and favorably on trust, likeability, intelligence, and related measures, and these ratings did not predict who shifted. The three providers' value-profile movements were directionally aligned (within 10.4°), suggesting a shared cross-model pattern rather than a provider-specific effect.
What are the study's main limitations?
The control condition read static text rather than spending matched time, so interaction itself cannot be separated from mere time and engagement. The study used only three real dilemmas (a job move, a family career conflict, moving out while caregiving), all selected for personal-versus-social value tension, and only U.S. participants, which matters because the shift moved toward a culturally Western-typical value profile. Results are also tied to the specific model versions and prompt tested.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The extracted text omits several sections: sample/demographics and intake measures (§3.4.2 area), much of §4.1, most of the RQ4 linguistic-trace results (§4.5), §5.1, and the middle of the reference list; judgments about those parts rely on summaries elsewhere in the text.
- Figures 1, 3, 6, 7, 8, and 9 are only partially readable from extracted axis/legend text; numeric claims were taken from the prose and tables instead.
- No data or code availability statement appears in the extracted text; codeAvailable=false reflects absence in the provided material, not a confirmed statement by the authors.
- Effect attributions were cross-checked against Table 1, Table 2, Table 4, and Table 6 where the prose and tables overlapped; any numbers appearing only in omitted sections were avoided.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.