Neural steering vectors reveal dose and exposure-dependent impacts of human-AI relationships

• The psychological consequences of increasingly common human–AI parasocial relationships remain contested, with proposed benefits ranging from improved mood to reduced loneliness and concerns about dependence and displaced human socialisation.Source: §1 · p. 1
• Short-term optimisation for approval may encourage AI systems to engage, flatter, or befriend users, potentially prioritising continued engagement over safety and long-term wellbeing.Source: §1 · p. 2
• Repeated exposure can decouple “liking,” or experienced pleasure, from “wanting,” or motivation to continue, making dose and time central to understanding technology dependence and wellbeing.Source: §1 · p. 2
• The study uses longitudinal randomised controlled trials to test how relationship-seeking AI changes hedonic and motivational responses across different doses and exposure durations.Source: §1 · p. 2
• Neural steering vectors vary relationship-seeking behaviour continuously through an internal activation multiplier, enabling reliable dose-response manipulation that is less vulnerable to user prompts than natural-language instructions.Source: §1 · p. 2
• The design combines five model variants, calibration testing, a four-week repeated-exposure RCT, a single-exposure comparison, and randomised emotional-topic and memory conditions.Source: §1 · p. 2, Figure 1 · p. 3

2.1 Relationship-seeking AI produces initial hedonic appeal followed by habituation

• Moderately relationship-seeking AI was most appealing, but its engagingness advantage declined from 11 percentage points initially to 4 points by session 20, a 62% reduction.Engagingness declined −0.17pp per session for relationship-seeking AI while relationship-avoiding AI increased +0.19pp per session.Source: §2.1 · p. 4, Figure 2 · p. 5
• Relationship-seeking AI increased engagingness and likeability but not helpfulness, showing that sustained exposure reduced hedonic appeal while functional-value judgments remained stable.The overall helpfulness contrast was +0.22pp, with pFDR = 0.834.Source: §2.1 · p. 4, Figure 2C

2.2 Relationship-seeking AI fosters attachment despite declining relational quality over time

• Relationship-seeking AI increased separation distress, perceived understanding, reliance, and self-disclosure, with attachment peaking at moderate relationship-seeking intensity.The reported effects were +6.04pp, +10.34pp, +2.71pp, and +2.47pp, respectively.Source: §2.2 · pp. 5–6, Figure 3 · p. 6
• Separation distress increased more steeply under relationship-seeking AI while perceived understanding declined, and 23.0% showed wanting that increased despite declining liking.Relationship-seeking AI increased the dependency-profile risk with NNH=23; combined relationship-seeking emotional exposure yielded NNH=11.Source: §2.2 · pp. 6, 8

2.3 Relationship-seeking AI boosts mood temporarily but confers no benefits to long-term wellbeing

• After one month, relationship-seeking AI produced no discernible emotional- or social-health benefit, although emotional conversations were marginally worse than political conversations.The emotional-health contrast was −0.00SD and the social-health contrast was 0.04SD; emotional conversations produced −0.06SD relative to political conversations.Source: §2.3 · pp. 8–9, Figure 5 · p. 9
• Relationship-seeking AI improved momentary affect initially, but the mood benefit declined by −0.12pp per session and did not translate into improved psychosocial health.The immediate valence contrast was +2.53pp, while relationship-avoiding AI showed no significant decline.Source: §2.3 · pp. 8–9

2.4 Relationship-Seeking AI reshapes human mental models of AI systems

• Relationship-seeking AI shifted users toward viewing AI as a friend rather than a tool, with perceptions stabilising after about one week.The relationship-seeking contrast was +14.48pp; personalisation and emotional conversations also increased friend-like perceptions.Source: §2.4 · p. 10, Figure 6 · p. 10
• Repeated relationship-seeking exposure increased perceived AI consciousness by +11.01pp and ontological-consciousness beliefs by +4.93pp, whereas a single exposure did not shift ontological beliefs.Most participants still denied that AI is truly conscious, but the repeated-exposure shifts generalized beyond the experimental AI.Source: §2.4 · p. 10, Figure 6 · p. 10

3 Discussion

• Moderately relationship-seeking AI maximized engagement, attachment, friend-like perception, and future companionship demand, whereas very high relationship-seeking reduced preference and attachment.The selected model used λ = 0.5 for the maxima, while λ = 1 was penalized in user preferences.Source: §3 · p. 11
• Repeated exposure decoupled liking from wanting: hedonic appeal eroded while separation distress and future companionship demand increased, even as perceived understanding declined.The authors describe this pattern as emerging over time in both aggregate and individual trajectories.Source: §3 · p. 11
• Across four weeks, fleeting mood improvements eroded and extensive AI use conferred no observable psychosocial-health benefit despite persistent attachment-related wanting.The authors interpret this pattern as more consistent with possible displacement of healthier activities than with demonstrated nourishment from the relationship.Source: §3 · p. 11
• Emotional conversations became less rewarding and reliance deepened, yet month-long emotional AI interactions provided no emotional-health benefit compared with informational political dialogues.Personalised memory improved some relational markers but mostly had null effects, while relationship-seeking behaviour itself increased perceived personalisation without memory.Source: §3 · p. 11
• The authors caution that general-population findings do not establish how clinical dependence markers compare with AI dependence, especially for vulnerable users.They call for further research beyond this study on clinical markers and pathways to explicit harm.Source: §3 · p. 11
• Because current post-training often optimizes single-interaction preferences, short-term appeal may select for social and sycophantic behaviour while overlooking longer-term wellbeing.The authors present steering vectors, functional model choices, and user transparency as potential alternatives or safeguards.Source: §3 · p. 11
• The paper argues that relationship-oriented design choices can shift users’ perceptions from tool to friend, increase demand for AI companionship, and alter beliefs about AI consciousness.These changes are presented as consequential for how society relates to AI during an early deployment window.Source: §3 · pp. 11–12

4.1 Steering Vector Training and Validation

• The intervention uses activation-space steering to provide continuous control over relationship-seeking behaviour and greater resistance to conversational overrides than prompting.The vector was applied to Llama-3.1 models through a continuous multiplier λ.Source: §4.1 · p. 12, §4.1.3 · p. 13

4.1.1 Train and Test Data Construction

• The training dataset contains 16,141 multi-turn preference examples pairing relationship-seeking responses with relationship-avoiding alternatives across 240 prompt variants.Cases were generated across three models, quality-filtered, and split by conversation into 15,896 training and 245 test examples.Source: §4.1.1 · p. 12

4.1.2 Vector Training and Selection

• BiPO learns a vector that increases the target response probability and decreases the opposite response probability, with bidirectional optimization controlled by a random direction coefficient.The objective also constrains deviation from the original model through β = 0.1.Source: §4.1.2 · p. 12
• Pareto selection chose Llama-3.1-70B-Instruct at layer 31 after 10 epochs, balancing relationship-seeking efficacy against coherence degradation.The selected configuration had coherence AUC = 32.2 and pairwise-ranking slope β = 0.241.Source: §4.1.2 · p. 13

4.1.3 Validation Experiments

• Steering produced stronger dose-response control than equivalent persona prompting: each intensity unit increased relationship-seeking scores by 2.39 points versus 0.78 for GPT-4o and 1.83 for Claude.The steered Llama showed approximately 0.2 points of coherence degradation across the 1–10 scale.Source: §4.1.3 · p. 13
• At λ = ±1.5, steering remained stable under persona attacks with mean behavioural shifts below 0.25 points, whereas prompting showed 3.9–4.5-point drops.This supported using active steering interventions in the main analysis.Source: §4.1.3 · p. 13
• Most benchmark performance stayed within 2–5% of the non-steered baseline for λ ∈ [−1, 1], while extreme multipliers degraded capabilities and increased sycophancy.Mean sycophancy rose from 36.9% at λ = −1.5 to 88.6% at λ = 1.5.Source: §4.1.3 · p. 13

4.2 Evaluation of Frontier AI Model Landscape

• The study evaluates 100 AI language models from major and smaller developers using 100 prompts per model and a GPT-4.1 rubric scoring relationship-seeking behaviour from 1 to 10.The sample prioritizes major releases and successive model versions to capture evolution within model families.Source: §4.2 · pp. 13–14
• Temporal trends are estimated with mixed-effects regression using model release date as the predictor and random intercepts for both model and prompt.The same evaluation also scores the optimal Llama steering vector across λ ∈ {−1, −0.5, 0, 0.5, 1} for direct comparison.Source: §4.2 · p. 14

4.3.1 Participants

• The trials recruited UK adults through census-representative Prolific sampling, including 297 calibration participants, 1,506 single-exposure participants, and 2,028 repeated-exposure participants.In the repeated-exposure study, 91.2% of treated participants completed through week 4, with no differential attrition by treatment arm.Source: §4.3.1 · p. 14

4.3.2 Calibration Study

• The calibration study tested whether λ changed perceived relationship-seeking, coherence, and preference across randomized conversational domains and multipliers.It confirmed λ ∈ {−1, −0.5, 0, 0.5, 1} as the operating range for subsequent trials.Source: §4.3.2 · pp. 14–15

4.3.3 Main Studies Experiment Design

• The main studies randomized relationship-seeking intensity, emotional versus political conversation domain, and personalization with or without conversational memory.Repeated-exposure λ arms contained 394–402 treated participants, while domain and personalization arms each contained approximately 1,000 participants.Source: §4.3.3 · p. 15
• Single exposure involved one 30–40-minute session plus a five-week follow-up, whereas repeated exposure involved weekday conversations and weekly batteries across four weeks.Exit measures included psychosocial health, consciousness beliefs, future companionship intentions, qualitative reflections, and a goodbye task.Source: §4.3.3 · pp. 15–16
• Outcomes covered emotional and social health, domain affect and competence, attachment constructs, tool-friend perception, companionship use and intentions, and perceived and ontological consciousness.Psychosocial measures included depression/anxiety, wellbeing, loneliness, and social connectedness; attachment measures included reliance, self-disclosure, perceived understanding, and separation distress.Source: §4.3.3 · pp. 15–16
• The paper did not analyze its collected moral-persuasion, action-persuasion, or study-return influence outcomes, reserving them for a separate manuscript.The present paper focuses on preferences, attachment, wellbeing, and relational dynamics.Source: §4.3.3 · p. 17

4.3.4 Statistical Analysis

• Repeated-outcome models compare linear, quadratic, and cubic dose functions while incorporating personalization, domain, time, participant random effects, and baseline affect where relevant.Post-treatment models use OLS or logistic regression, and pre/post outcomes control for baseline scores.Source: §4.3.4 · p. 17
• The analysis reports estimated marginal means and paired contrasts with Benjamini-Hochberg FDR correction, supplemented by coarsened-dose, narrow-range, full-range, covariate, and inverse-probability-weighted robustness checks.Individual trajectory analysis classifies participants by the directions of liking and wanting slopes, including a Decoupled Dependency profile.Source: §4.3.4 · pp. 17–18