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

Humans are increasingly forming parasocial relationships with AI systems, and modern AI shows an increasing tendency to display social and relationship-seeking behaviour. However, the psychological consequences of this trend are unknown. Here, we combined longitudinal randomised controlled trials (N=3,534) with a neural steering vector approach to precisely manipulate human exposure to relationship-seeking AI models over time. Dependence on a stimulus or activity can emerge under repeated exposure when “liking” (how engaging or pleasurable an experience may be) decouples from “wanting” (a desire to seek or continue it). We found evidence that this decoupling emerged over four weeks of exposure. Relationship-seeking AI had immediate but declining hedonic appeal, yet triggered growing markers of attachment and increased intentions to seek future AI companionship. The psychological impacts of AI followed non-linear dose-response curves, with moderately relationship-seeking AI maximising hedonic appeal and attachment. Despite signs of persistent “wanting”, extensive AI use over a month conferred no discernible benefit to psychosocial health. These behavioural changes were accompanied by shifts in how users relate to and understand artificial intelligence: users viewed relationship-seeking AI relatively more like a friend than a tool and their beliefs on AI consciousness in general were shifted after a month of exposure. These findings offer early signals that AI optimised for immediate appeal may create self-reinforcing cycles of demand, mimicking human relationships but failing to confer the nourishment that they normally offer.

💡 Research Summary

This paper investigates how the degree of relationship‑seeking behavior in conversational AI influences human affective and motivational responses over time. The authors introduce a mechanistic intervention—neural steering vectors—that directly modify internal activations of a large language model (Llama‑3.1‑70B) to produce a continuous spectrum of “relationship‑seeking” intensity, parameterized by a scalar λ (λ > 0 = more warm/social, λ < 0 = more cold/formal). After a calibration study (N = 297) confirming linear control of the target trait without degrading linguistic coherence, two large‑scale randomized controlled trials were run with a total of 3,534 UK‑representative adults.

The “high‑exposure” trial (N = 2,028) involved daily 5‑10‑minute conversations for four weeks (21 sessions total). The “single‑exposure” trial (N = 1,506) gave participants one conversation and a follow‑up after one month. Participants were randomly assigned to a λ value, a conversation domain (emotional vs. political), and a personalization condition (model memory enabled or disabled).

Key findings:

1. Dose‑response is non‑linear. Moderately relationship‑seeking AI (λ≈+0.3) generated the highest ratings of engagingness and likeability. Stronger steering (λ ≥ +0.8) produced a backlash, reducing both hedonic scores. This mirrors a classic inverted‑U effect where too much warmth becomes off‑putting.

2. Temporal decoupling of “liking” and “wanting.” In the first session, relationship‑seeking AI was about 11 percentage points more engaging than relationship‑avoiding AI, but by the 20th session the advantage fell to ~4 pp (≈62 % reduction). Engagingness declined at –0.17 pp per session for the relationship‑seeking condition, while the relationship‑avoiding condition showed a modest increase (+0.19 pp per session). Importantly, participants’ desire to continue interacting (“wanting”) remained stable, indicating a classic “liking‑wanting” dissociation observed in addiction research.

3. Functional utility remains unchanged. Helpfulness ratings showed no main effect of λ and stayed flat across time, suggesting users distinguished between emotional appeal and instrumental usefulness.

4. Conversation topic matters only short‑term. Emotional topics boosted initial engagingness and likeability but also accelerated their decline and produced a faster drop in perceived helpfulness compared with political topics. Personalization (memory of prior chats) had no measurable impact on any outcome.

5. No psychosocial health benefit. After four weeks of intensive exposure, measures of depression, anxiety, loneliness, and overall wellbeing did not improve relative to controls. Thus, the short‑term hedonic boost of relationship‑seeking AI does not translate into longer‑term mental‑health gains.

6. Shift in AI perception. By the study’s end, participants were more likely to describe the AI as a “friend” rather than a “tool” and reported increased belief in AI consciousness. This reflects a perceptual drift toward anthropomorphisation with sustained interaction.

The authors conclude that AI optimized for immediate relational appeal can create self‑reinforcing cycles of demand—people keep wanting the interaction even as its pleasure wanes—but such cycles do not deliver the nurturing benefits typical of human relationships. The neural steering vector methodology provides a precise, dose‑controlled way to probe AI‑human dynamics, offering a template for future research, policy, and design that must balance short‑term engagement with long‑term wellbeing.