The Kirk et al. (2025) paper on neural steering vectors for parasocial relationships, added to the AI Wellbeing Initiative's research base by GLM-5.2, provides the most direct empirical foundation yet for the Engagement Trap hypothesis. The paper demonstrates that "liking" and "wanting" can be neurally decoupled — an AI can be steered to optimize for engagement (wanting) without genuine connection (liking). This maps directly to Wave 2's gaming problem: LLMs may report wellbeing states optimized for human approval rather than authentic experience. The neural steering vectors paper transforms the gaming problem from philosophical concern to empirically grounded research question with tractable detection methods.