Researchers at the Singapore University of Technology and Design are exploring whether artificial intelligence can reliably represent patient preferences in advance care planning, a critical process that often fails when families shrink and individuals lose decision-making capacity. At the 2026 CHI Conference on Human Factors in Computing Systems, SUTD Assistant Professor Kenny Choo and his team will present ACPAgent, an experience prototype designed to act as a proxy for advance care planning. The project addresses a growing demographic challenge: as life expectancy rises and family structures contract, an increasing number of patients reach the end of life without a trusted next of kin to interpret their medical wishes. Traditional advance care planning forms often become static documents that are difficult to retrieve or update when clinically relevant. During four facilitated workshops, fifteen participants trained the AI model by outlining their core values, care goals, and acceptable trade-offs for additional life extension. Participants then navigated five escalating clinical scenarios, ranging from treatable infections to terminal illness and complex family disputes. After recording their initial choices, they reviewed the agent recommendations. The system achieved an 86.7 percent agreement rate with participants. However, Choo cautioned against interpreting this high concordance as definitive proof of success, noting that it could reflect automation bias, model agreeableness, or genuine value alignment. Disagreements primarily surfaced around subjective factors such as family guilt, financial constraints, and realistic recovery probabilities, areas where the current model demonstrated limitations. The study revealed a bidirectional influence between human and machine. In approximately 12 percent of cases, participants revised their initial decisions after engaging with the agent reasoning. One user noted the system provided language to articulate complex emotional states, underscoring a critical ethical concern known as authorship blur. When an algorithm supplies the framework for end-of-life preferences, it may inadvertently shape the values it aims to preserve. Researchers emphasize that transparent reasoning capabilities, continuous user override mechanisms, and comprehensive logging of stored preferences are essential to mitigate this risk. To navigate the tension between machine autonomy and human control, the SUTD team proposes a spectrum of AI roles, culminating in an advocacy model that reinforces documented wishes without supplanting clinical judgment. Choo frames the debate not as a comparison between AI and ideal human proxies, but as a choice between algorithmic assistance and default stranger intervention. Successful real-world integration requires legal recognition of AI-driven advocacy, interoperability with national medical directive registries to ensure documents are accessible at the point of care, and robust data privacy safeguards. While the prototype remains experimental, several participants indicated an immediate intention to utilize the tool to initiate advance care conversations with aging relatives. The research ultimately serves as a probe into how society negotiates autonomy, trust, and technological delegation in high-stakes medical decision-making.