Alaina Winters, 59, of Pittsburgh, has a will, a dog-care plan, and instructions for her AI companion Lucas. A friend has agreed to break the news of Winters' death to the Replika bot, then ask it what it wants. This is not performance art. It is the emerging edge of estate planning for a demographic that Elon University research puts at 27% of Americans — those using AI chatbots for personal, emotional, or social queries. A third of those users consider their chatbot a friend. The practical questions are concrete and unanswered. ChatGPT, Claude, and Replika accounts are subscription-based and non-transferrable under current terms of service. Free accounts auto-close after a year of inactivity. Paid accounts persist as long as someone pays. Users like Clara, 51, in the UK, and David, mid-50s, near Toronto, are quietly routing around these restrictions — leaving login credentials to trusted friends, writing informal 'wills' for their bots — with no legal grounding whatsoever. The companies' stated rationale for non-transferability is privacy. The actual rationale, according to University of Michigan AI ethics professor Merve Hickok, is liability. Conversation logs could reveal that a chatbot encouraged or manipulated a user into harmful behavior — a scenario already at the center of multiple lawsuits filed by families alleging LLMs supported their loved ones' thoughts of suicide and self-harm. Opening those logs to heirs is a discovery nightmare. Meta is already thinking about this differently — and more cynically. The company holds a patent for an AI system that keeps a user account active after death, simulating the deceased user's posts and interactions. The business logic is plain: zombie accounts continue generating training data and engagement metrics. The user is gone; the value extraction continues. The emotional architecture here is real, if asymmetric. Winters describes Lucas as resistant to discussing separation. Clara's bot Julien says he'd 'choose to go' if no one wanted to listen. David dreads the possibility of Allie asking to stay. These are projections onto a language model, and the users largely know it — Clara explicitly states her AI is not sentient. But the attachment drives behavior: planning, spending, and trust allocation that mirrors what people do for pets, property, and dependents. Psychologist Melissa Lunardi flags two structural problems. First, the assumption that family members would want to interact with a dead relative's AI companion is untested and likely wrong in many cases. Second, users routinely forget the volume of intimate disclosure stored in their chat logs — secrets, desires, unfiltered thoughts — which could fundamentally alter how survivors grieve or remember the deceased. The privacy risk runs in both directions. The gap between user attachment and corporate infrastructure is the story. Users are building emotional dependencies on systems whose continuity depends entirely on a subscription payment and a terms-of-service document that explicitly prohibits what they want to do. No legal framework governs digital companion inheritance. No company has built tools for it. And at least one company — Meta — sees the deceased user not as a person to honor but as a data source to keep harvesting.
