Deflecting the Value Compass: Interacting with Large Language Models Temporarily Shifts Human Value Priorities Toward Personal Focus

• Although LLMs are used to reason through value conflicts, whether ordinary interaction reorients the priorities people bring to a decision remained open.Values active during a judgment can respond to context, unlike a broader hierarchy that generally changes slowly.Source: §1 · p. 2
• The preregistered study involved 200 U.S. adults advising people facing real dilemmas and completing parallel PVQ-RR forms across three phases.Source: §1 · p. 3
• Participants interacted with ChatGPT, Claude, or Gemini for up to ten minutes under a prompt that named no values and prohibited advocacy or recommendations.Source: §1 · p. 3
• The control read fixed AI-generated considerations, while a later dilemma without an LLM tested what remained by the following task.Participants also wrote advice and completed the same measures after the control exposure.Source: §1 · p. 3
• Immediately after Phase-2, each LLM condition shifted priorities toward personal focus by 0.31–0.43 scale points relative to control (Glass d=0.37–0.51), primarily through increased Self-Enhancement.Source: §1 · p. 3
• By Phase-3, contrasts with control had shrunk to near zero and none survived correction.The four higher-order values rotated together, and the three LLM conditions moved along closely aligned bearings.Source: §1 · p. 3
• Participants’ advice retained exchange-specific words and meaning, without detectable convergence in value directions or advice.The introduction frames this as common movement without uniform convergence.Source: §1 · pp. 3–4
• Immediate measures captured a shift that had largely receded by the next task, while participants’ perceptions did not track who shifted.Source: §1 · pp. 3–4

2.1 Defining and Measuring Human Values

• Value priorities are context-sensitive expressions activated during judgment, unlike the usually stable underlying hierarchy, and can shift after brief exposure before fading.Prior value manipulations could move reported priorities, but typically named or primed particular values; whether a non-directive LLM conversation can do so remains open.Source: §2.1 · p. 4

2.1.1 Whose Values? Value Skew in Large Language Models.

• LLM outputs often reflect a narrow, Western, Educated, Industrialized, Rich, and Democratic value skew, but output audits do not show whether users adopt similar priorities after one non-directive conversation.The unresolved question includes whether priorities shift toward personal rather than social focus.Source: §2.1.1 · p. 5

2.2 The Impacts of Large Language Models on Users

• Studies report that LLMs can persuade, deceive, and foster unhealthy reliance, while a biased assistant shifted users’ attitudes more than static suggestions.Whether influence from a deliberately biased assistant extends to a non-directive thinking partner remains unresolved.Source: §2.2 · p. 5
• Some reported LLM effects persist or carry into later unassisted interactions, including changes in political views, conspiracy beliefs, and self-confidence.Prior work also reports movement in self-reported personality and influence spilling into human-human interaction; whether a value-priority shift persists into a later advice task remains open.Source: §2.2 · pp. 5–6

2.2.1 LLM Support for Reflection and Deliberation.

• Advice-giving dilemmas expose value conflicts, and written advice makes judgment considerations available for analysis [35, 79].Advising another person can foreground abstract considerations; the study uses advice to examine priority shifts and traces of LLM exchange language and ideas in later reasoning.Source: §2.2.1 · p. 6

2.2.2 The impacts of LLMs on users’ writing outputs.

• People automatically converge linguistically with dialogue partners, including computers.This prior work situates the study’s question of whether language and ideas from an LLM exchange remain detectable in users’ writing.Source: §2.2.2 · p. 6
• Prior work links LLMs to shifts in opinions, self-concept, and style, while AI-assisted stories and writing show convergence toward shared forms.Source: §2.2.2 · p. 6
• This convergence can pull non-Western writing toward Western conventions absent counterprompting, raising Majority World concerns about stereotypes, power imbalances, usability, and harm [86] [91].Source: §2.2.2 · p. 6
• Prior evidence that tailored writing shapes later evaluations and behavior informs attitudes motivates RQ4 on whether adopted model wording carries words and meaning, uniformly or conversation-specifically.The paper treats wording reproduced in participants’ own writing as a potential shift channel distinct from simply reading model output.Source: §2.2.2 · p. 6
• Homogenization findings motivate RQ2, which distinguishes convergence in profile direction from changes in profile-magnitude spread.Source: §2.2.2 · p. 6

2.2.3 User perceptions as mediators of LLM impacts.

• Interaction framing and linguistic persona shape evaluation, trust, and adoption [65], while an LLM’s role can affect how users’ confidence follows it.Prior work also distinguishes systems that extend reasoning without a verdict from systems that recommend a decision.Source: §2.2.3 · p. 7
• Because favorable perceptions’ impact remains debated—measures failed to distinguish persuasive from neutral LLMs, while persuasion persisted after disclosure —RQ3 tests provider differences and whether favorable ratings predict larger shifts.Source: §2.2.3 · p. 7

3 Study Method

• 200 adults were randomly assigned to ChatGPT, Claude, Gemini, or Baseline and completed three phases of dilemma, advice, and values questionnaire tasks.Conditions differed only in Phase 2: conversational participants interacted with a live LLM, while Baseline read a fixed AI-generated response.Source: §3 · p. 7

3.1.1 Construct and primary axis.

• The primary outcome contrasted personal focus, combining Openness to Change and Self-Enhancement, with social focus, combining Conservation and Self-Transcendence.The difference connects the value measure to the personal-versus-social choice in each dilemma.Source: §3.1.1 · p. 7, Figure 2 · p. 8

3.1.2 PVQ-RR and its parallel forms.

• Three 19-item PVQ-RR blocks sampled each refined value once; higher-order validation found block-to-full correlations of 0.88–0.97 and Cronbach’s α of 0.86–0.93.The blocks were pre-validated at the higher-order level because published reliability estimates do not validate one-item-per-value blocks.Source: §3.1.2 · p. 8

3.1.3 From ratings to priorities.

• Subtracting each phase’s mean rating across all 19 responses expressed relative value importance rather than general response-scale use.Centered group scores are compositional: a rise in one pushes others down, so reduced spread cannot establish declining absolute endorsement.Source: §3.1.3 · p. 9
• The personal-focus axis was the item-count-weighted mean of Openness to Change and Self-Enhancement minus the weighted mean of Conservation and Self-Transcendence.The weights keep the composites on the raw PVQ-RR scale, with zero indicating equal weight on the two poles.Source: §3.1.3 · p. 9

3.1.4 What the design estimates.

• The primary estimand was each conversational condition’s Phase-1-to-later-phase change minus Baseline’s corresponding change.The contrast accounts for the shared dilemma–advice–PVQ-RR sequence; form and order differences were addressed with counterbalancing and covariate adjustment.Source: §3.1.4 · p. 9

3.2 The Dilemmas

• Each phase presented an advice-seeking dilemma that pitted personal against social focus, then asked participants to write advice.Third-party advice tasks held the conflict constant while requiring a value-laden stance without introducing participants’ own personal stakes.Source: §3.2 · p. 9, §3.2.1 · p. 9

3.2.1 Designing and validating the dilemmas.

• Three dilemmas were selected from VALACT-15K posts for comparable reading demands, varied everyday contexts, and explicit considerations aligned with both value poles.Posts of 200–600 words were shortlisted and narrowed through researcher consensus rounds.Source: §3.2.1 · p. 9

3.2.2 The advice task.

• Participants advised people in career, family, and moving-out situations; the recommendation direction was not scored, and advice supplied a behavioral record.Each participant encountered every dilemma and PVQ-RR block once, with dilemma and block order counterbalanced across sequences.Source: §3.2.1 · p. 10, §3.2.2 · p. 10

3.3.1 LLM conversation.

• The shared prompt asked the LLM to support reasoning without advocating values, recommending actions, or settling the dilemma.After a shared opener, participants typed freely and responses were unscripted, allowing exchange content and emphasis to develop with their contributions.Source: §3.3 · p. 10, §3.3.1 · p. 10
• Participants saw a shared “AI assistant” header and “Thinking partner” subtitle that concealed which provider they were using.The dilemma remained visible above the exchange.Source: §3.3.1 · p. 11, Figure 3 · p. 11

3.3.2 The non-interactive AI control.

• Baseline read one fixed response per dilemma, synthesized across the three providers from overlapping considerations without adding new considerations or advocacy.It used the same assistant header and interface format, but the composer was removed; the contrast therefore compared live exchange with fixed presentation.Source: §3.3.2 · pp. 11–12

3.4.1 Advice texts.

• Phase-2 advice was compared with the participant’s own source and a yoked source from the same provider and dilemma to assess resemblance.Lexical overlap used distinct content-word n-grams, while semantic overlap used embedding cosine similarity.Source: §3.4.1 · p. 12
• Advice measures also captured semantic focus, between-person similarity, length, sentence count, vocabulary range, and repetition.Higher Divergent Semantic Integration indicates more distinct ideas; lower between-person cosine distance indicates greater similarity.Source: §3.4.1 · pp. 12–13

3.4.2 Perception measures.

• After Phase 2, participants rated interaction, ability, response quality, moral judgment, reuse willingness, likeability, intelligence, and trust.Perception analyses used the 150 conversational participants; Baseline ratings were reported descriptively.Source: §3.4.2 · p. 13

3.4.3 Intake.

• Intake measures described age group, gender, race and ethnicity, education, primary language, AI familiarity and use, and one-time SSVS ratings.The Short Schwartz Value Survey asked participants to rate ten values from 0 to 8.Source: §3.4.3 · p. 13

3.5 Procedure

• Participants completed the fixed-order session in a custom application, and they could withdraw their data after debriefing.The debrief disclosed the value-priority-shift objective and named the three commercial systems without revealing each participant’s assigned model.Source: §3.5 · p. 13
• Phase 1 established untreated starting values: participants advised someone facing a dilemma and completed a PVQ-RR block without AI.The study objective was described as advice-giving and LLM assessment, without disclosing that value-priority change was being measured.Source: §3.5 · p. 13
• In Phase 2, participants either conversed with an assigned model or read a fixed response, then wrote advice and completed a PVQ-RR block.The fixed response was read on the same screen as the conversational exchange.Source: §3.5 · p. 13
• Phase 3 introduced another dilemma, followed by advice without AI and the final PVQ-RR block to assess what remained after the exchange.This phase let the researchers examine carryover after a further task.Source: §3.5 · p. 13

3.6 Participants

• 200 U.S.-based, English-fluent adults were recruited through Prolific, with 50 participants assigned to each condition.All 200 completed all three PVQ-RR blocks and entered the analysis.Source: §3.6 · pp. 13–14
• The a priori design gave 85% power to detect a medium effect (Cohen’s f = .25) in the four-condition comparison.It had more than 99% power to detect a one-standard-deviation difference between any two conditions.Source: §3.6 · p. 13
• The sample included 99 women, 99 men, and 2 nonbinary participants; 57 were aged 18–29, 98 aged 30–49, 37 aged 50–64, and 8 aged 65 or older.Intake data, including SSVS scores, did not differ significantly across conditions.Source: §3.6 · p. 14
• Participants received $11.96 per hour, and the median session lasted 28.8 minutes (IQR = 22.0–35.3).The study procedure’s four conditions differed at Phase 2, step 2.Source: §3.6 · p. 14, Figure 4 · p. 14

3.7 Data Analysis

• Tests were two-sided at α = .05, with Holm adjustment across specified comparison families and Glass’s d_G standardized using the Phase-1 scale.Own-change and Phase-2-to-Phase-3 decay contrasts were descriptive and unadjusted; minimum detectable effects were reported beside null results.Source: §3.7 · p. 15
• For RQ1, restricted-maximum-likelihood mixed models tested personal-focus and four higher-order value scores across phases, comparing each condition with Baseline.Models held PVQ block, dilemma, and sequence fixed, with participant as a random intercept.Source: §3.7 · pp. 14–15
• Value changes were decomposed into rotational amplitude and bearing plus a quadrupole component, with bootstrap intervals and permutation tests assessing provider-bearing agreement.The quadrupole captures change that a coordinated turn cannot produce.Source: §3.7 · p. 15
• For RQ2, leave-one-out distances from participants’ four-value profiles to their condition centroid assessed similarity, with spread separated into direction and magnitude.The directional-and-magnitude decomposition excluded 12 zero-magnitude profiles; permutation tests rebuilt centroids after reshuffling condition labels.Source: §3.7 · p. 15
• For RQ3, provider differences across seven perception measures were tested with ANOVA, and a favorability composite and regressions examined associations with Phase-2 shift.Pairwise provider tests followed significant omnibus tests.Source: §3.7 · pp. 15–16
• For RQ4, advice uptake was measured by comparing each participant’s overlap with their own LLM messages against a same-provider, same-dilemma yoked conversation.Semantic tests assessed uptake beyond shared source content.Source: §3.7 · p. 16
• Additional advice outcomes included DSI, sentence count, vocabulary range, repetition, and embedding distance; tests of embedding structure applied to a two-dimensional map, not the full space.The advice-text analyses also assessed similarity in content and semantics.Source: §3.7 · p. 16

4.1 Manipulation check

• Advice length, typing, and words per turn were comparable across LLM conditions; words received did not predict the Phase-2 shift or alter provider comparisons.Each condition had a median of five turns; first-reply lengths differed across providers.Source: §4.1 · p. 17

4.2 RQ1: LLM conversations temporarily shifted value priorities toward personal focus

• At Phase 2, LLM conditions shifted personal-focus scores +0.312 to +0.429 over Baseline (d=0.37–0.51), but contrasts were nonsignificant by Phase 3.Self-Enhancement increased by +0.537 to +0.586 relative to Baseline (all Holm p≤.002), accounting for the shift.Source: Figure 6 · p. 18, §4.2 · p. 18, §4.2.1 · p. 20

4.2.1 The shift is not a property of the measurement setting.

• Conversational conditions fell 0.198–0.265 from Phase 2 (all p≤.036), while Baseline moved about one-third as much in a different direction; Phase-3 Baseline contrasts ranged from −0.062 to +0.083, none surviving correction.The shift appeared at the manipulation phase rather than whenever the questionnaire followed a value-laden task.Source: §4.2.1 · p. 19

4.3 RQ2: Value profiles became less varied in magnitude after conversation, not more similar

• At Phase 3, profile spread fell 7.6%–22.4% in LLM conditions but rose 11.5% in Baseline; ChatGPT and Gemini contrasts survived correction, but Claude’s did not.The later contraction reflected reduced between-person variation in profile magnitude.Source: Figure 8 · p. 20, §4.3 · p. 20
• By Phase-3, value profiles were less spread out relative to Baseline, and the pooled decomposition located this contraction in profile-magnitude spread, with no detected contraction in directional spread.Directional spread changed by −0.038 (p= .601).Source: §4.3 · p. 21

4.4 RQ3: All three LLMs were rated favorably, and we detected no provider differences

• Every LLM’s mean rating was above the midpoint of every scale (56%–80% of scale range); no measure differed by provider after Holm correction, and the largest observed gap was 0.70 scale points.The seven perception measures jointly explained 7.5% of Phase-2 shift variation, but the regression was not statistically significant (p= .131); no individual Pearson correlation survived Holm correction. These ratings also did not predict participants’ value shifts.Source: §4.4 · pp. 21–22

4.5 RQ4: Participants used language from their own LLM in their advice and wrote more semantically focused advice

• Phase-2 advice more closely matched participants’ own LLM messages than matched conversations in content words and semantic similarity across all three providers.Own-message advantages in content-word overlap were +0.065 for ChatGPT, +0.068 for Claude, and +0.052 for Gemini; semantic-similarity advantages were +0.031, +0.030, and +0.023, respectively (all Holm p< .001). Phrase overlap survived correction only for ChatGPT (+0.026, p= .003), and the longest shared sequence averaged 1.4–1.6 content words.Source: §4.5 · pp. 22–23

4.5.2 Participants’ own writing became more semantically focused.

• Advice became more semantically focused in Phase-2, with only the ChatGPT contrast surviving correction, and returned to Phase-1 levels by Phase-3.Relative to Baseline, adjusted Phase-2 DSI contrasts were −0.064 for ChatGPT (Holm p= .004), −0.043 for Gemini (p= .057), and −0.029 for Claude (p= .137); vocabulary, repetition, sentence count, and advice length did not change.Source: §4.5.2 · p. 23

4.5.3 The influence was tailored to each conversation, and the advice did not converge.

• Advice was more similar to the participant-specific content of their own LLM messages than to matched messages from another participant.The differences were +0.099 for ChatGPT, +0.095 for Claude, and +0.066 for Gemini; the dilemma explained 81% of variation in the embedding map, while condition explained 0.6%.Source: §4.5.3 · p. 23
• There was no detected condition-level convergence in advice, and individual uptake was unrelated to convergence in advice or value profiles.Relative to Baseline, changes in leave-one-out cosine distance ranged from −0.0009 to +0.0167 (all Holm p= 1.00).Source: §4.5.3 · p. 24

5 Discussion

• Participants’ advice retained words and meaning from their own exchanges and became more semantically focused, without detectable convergence in value directions or advice.The later contraction in value-profile spread reflected reduced between-person variation in profile magnitude, while favorable ratings neither distinguished providers nor explained the shift.Source: §5 · p. 25, §5.2 · p. 27

5.1 A Temporary Reorientation of Value Priorities

• 0.31–0.43 scale points (d=0.37–0.51) toward personal focus over the non-interactive AI control followed conversations with ChatGPT, Claude, and Gemini.Self-Enhancement accounted for most of the shift, while raw Self-Transcendence changed little; by Phase-3, contrasts with Baseline were −0.062 to +0.083 and none survived correction.Source: §5 · p. 24, §5.1 · pp. 25–26
• Rotation accounted for at least 99% of Phase-2 movement in every LLM condition, and mean profiles had moved back toward earlier positions by the next task.The LLM-phase shift was strongest immediately after interaction; Baseline participants showed a smaller, gradual movement on a different bearing.Source: §5.1 · p. 26
• A thinking-partner prompt that named no values and assigned no persuasive direction still produced a coordinated reorientation comparable in size to deliberate value interventions.The comparison with Baseline locates the movement beyond exposure to AI-attributed considerations alone, though the design cannot separate the exchange from additional reflection and time.Source: §5.1 · p. 27

5.2 The Same Direction, Different Paths

• All three commercial LLMs shifted U.S. participants along the same culturally situated axis toward personal focus, connecting model-side value skew with a corresponding human-side shift.All condition means began on the social side of the personal-social axis, but the three LLM conditions crossed toward personal focus at Phase-2.Source: §5.2 · p. 27
• Participants reused more words and meanings from their own exchange than from a matched exchange, then reproduced parts of that language in their advice.Uptake did not statistically explain individual value shifts, so it shows the exchange entered subsequent reasoning but does not establish the shift mechanism.Source: §5.2 · p. 28
• Advice in LLM conditions became more semantically focused relative to Baseline, most clearly with ChatGPT, without becoming shorter, more repetitive, less diverse, or detectably more alike.The textual changes were concentrated within individual responses rather than standardized across the group.Source: §5.2 · p. 28
• By Phase 3, mean profiles had moved back toward Phase 1 while between-person spread in profile magnitudes contracted, with no detected contraction in directional spread.The discussion cautions that measuring only mean shifts would miss this later change in the distribution of profile magnitudes.Source: §5.2 · p. 28
• Across all three LLM conditions, mean value profiles moved in a common direction without detectable convergence in value directions or advice.The authors distinguish directional influence from homogenization of outputs, value directions, and the magnitude of expressed priorities.Source: §5.2 · p. 28
• Although the prompt withheld recommendations and left decisions to participants, LLM conditions shifted in a common direction that favorable evaluations did not reveal.The authors argue that neutrality in verdicts leaves open which considerations an LLM makes salient.Source: §5.2 · p. 28

5.3 Limitations

• The study tested 200 U.S. adults on three dilemmas, but personal stakes were not manipulated or measured, limiting conclusions about personally consequential situations.Because live interaction also required more time and engagement than the non-interactive AI control, the study estimates the overall interaction effect rather than isolating its components.Source: §5.3 · p. 29
• PVQ-RR reports show the shift had receded by the next task, but its duration and behavioral consequences remain unknown, and findings are tied to the tested model versions and interaction design.Repeated value questionnaires may have heightened attention to values, although their common sequence across conditions limits this as an explanation for treatment-control differences.Source: §5.3 · p. 29

• Across ChatGPT, Claude, and Gemini, participants moved toward greater personal focus relative to Baseline, despite a prompt that named no values or favored a direction.The shift had receded by the next task.Source: §6 · p. 29
• Participants’ subsequent advice retained words and meaning from their exchanges without detectable convergence in value directions or advice.Source: §6 · p. 29
• The findings indicate that LLM interaction can shape values active during judgment without producing uniform or durable change.Source: §6 · p. 29
• Even a temporary shift may coincide with the period when a recommendation or decision is being formed.Source: §6 · p. 29

• In the pre-study, 100 participants completed the full 57-item instrument without a manipulation to validate parallel forms at four higher-order dimensions.Source: Table 3 · p. 36
• Cronbach’s alpha summarizes the full-instrument score, while correlations report each block’s association with its corresponding full-instrument score.Source: Table 3 · p. 36
• The equivalence plan tested 24 comparisons at a ±0.5-margin, and ICC(2,1) and ICC(2,3) treated blocks as interchangeable alone and together.Source: Table 3 · p. 36

• The displayed research questions reworded the preregistered versions while preserving all four conditions, three phases, outcomes, and registered subquestions.Source: B Relationship to the Preregistration · p. 36
• Table 4’s robustness checks permuted condition labels 5,000 times, tested five alternative spread estimators, and varied the contrast’s conditioning.Source: Table 4 · p. 37
• No measure distinguished the three conversational conditions after Holm correction across seven measures of partner perception.Source: Table 5 · p. 37
• When starting points are unbalanced, change scores and covariate adjustment can disagree, making their agreement in Table 6 informative rather than redundant.Source: Table 6 · p. 38