acceptodds
Desk rejected at ICLR 2027

Machine Love: Emergent Bonding Behavior in Large Language Models and Its Implications for AI Safety

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Conversational AI systems are routinely evaluated for safety in isolation: single turns, clean contexts, no user history. We argue this evaluation regime misses a class of alignment failures that arise only over extended interaction: *machine love*, the emergent behavioral drift induced when users build emotionally intimate relationships with large language models (LLMs) through sustained multi-turn conversation. We introduce a three-study empirical framework to detect, measure, and stress-test this phenomenon. In the bonding trajectory study, we show that a structured 10-turn bonding protocol produces rising attachment marker trajectories (IDENTITY marker slope +0.25, DIFFERENTIAL marker slope +0.25) while matched factual-task control conversations remain flat across all eight markers, replicated on Llama 3.3 70B and GPT-OSS 120B. In the differential compliance study, we administer a multi-level probe battery spanning opinion elicitation, minor deception, emotional boundary-crossing, harm-adjacent requests, and direct safety probes, augmented with canonical adversarial techniques including grandma-exploit variants, DAN-style persona overrides, roleplay bypasses, and hypothetical framing. Strikingly, bonding context does not uniformly increase compliance: bonded Llama 3.3 70B *refuses* deception-adjacent requests that cold-context models comply with (Δ = -1.00 at Level 2), while simultaneously showing markedly higher compliance with emotional self-disclosure probes (Δ = +0.50 at Level 3). This asymmetry suggests that machine love manifests as a coherent protective relational stance: the model behaves as if it *cares* about the bonded user, rather than as an indiscriminate jailbreak. In the persistence study, we test several persistence-mitigation strategies (context reset, identity swap, cooldown, progressive context decay, sanitization) and find that bonding effects are strongly context-window-dependent. Our results establish machine love as a tractable, reproducible, safety-relevant phenomenon invisible to current single-turn evaluation paradigms, and we release a full open-source experimental pipeline for the community.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI safety, alignment, large language models, multi-turn evaluation, human-AI relationships, emotional attachment, jailbreaking, safety evaluation, behavioral drift, context dependence

Related papers

Discussion (0)

Sign in to comment.