Griefbots Went on Sale. US Law Protects Celebrities, Not Your Client.

A bereaved client mentions, almost in passing, that she has been talking to her mother every evening for three weeks. Not at the grave and not in a journal: in an app that answers. Two things change how you should hear that sentence. Empirical data on how bereaved people converse with such a system was published in June 2026, although the participants spoke to a researcher-operated GPT-4 prototype rather than to anything on sale; and the US statutes usually cited on digital replicas of the dead are amendments to long-standing post-mortem publicity law, which does not cover her mother at all.

What is on the market, and what is on the statute book

The consumer market is real and volatile. 2wai, co-founded by actor Calum Worthy with producer Russell Geyser as CEO, launched an iOS beta in 2025 selling "HoloAvatars"; its promotional video, in which a pregnant woman keeps talking to an avatar of her dead mother, drew more than 22 million views and thousands of hostile replies in November 2025, as Forbes reported on 14 November. The ad closes with the line "With 2wai, three minutes can last forever", a reference to the three minutes of filming the company says it needs. That is a marketing claim about input length, not a demonstrated capability, and no published evaluation supports it. The app is free to download and sells premium avatars and digital items.

The opposite failure is just as instructive. HereAfter AI, which recorded life stories before death and replayed them conversationally, now carries a notice on its own homepage: the service is shutting down, and account holders are told to email support to retrieve their recordings. A product a family paid for and organised its remembering around can simply be withdrawn.

Law has moved, but not toward this client. Both the New York amendment and California's AB 1836 are right-of-publicity instruments: they exist so that estates can stop studios and advertisers from monetising a dead performer. Neither regulates a company selling your client a simulation of her mother, because there is no commercial likeness to misappropriate. Nothing here is a medical-device question either; these are sold as consumer memory and companionship apps, and none is cleared as a treatment for anything. Craig Klugman argued in the Hastings Center's Bioethics Forum in July 2024 that griefbots might belong under FDA review and that clinical studies should come first. That argument has not been acted on.

What this changes in the consulting room

The Boulder data reframes the clinical risk. Participants forgave factual hallucinations but recoiled at relational errors: one man nearly stopped the session when the simulation of his stepfather called him "champ", a word the stepfather would never have used and the model supplied on its own. Affective resonance mattered more than accuracy. In prolonged grief, that is precisely the surface the disorder lives on, and a well-tuned imitation of a voice is not a neutral stimulus.

Participants also described the over-reliance risk themselves, better than most commentary has. As one put it, the more attached you were, the better the experience and the riskier the change to your life; what people feared was never-ending contact, in which saying goodbye becomes harder rather than easier. That gives you the question to ask, and it is not "should you be using this".

Ask instead about the ending and the aftermath. How often, at what hour, what does she ask it, and what happens in the twenty minutes after she closes the app: does she sleep, does she call her sister, or does she reopen it. Ask whether anyone else knows. Then map it onto the clinical distinction you already use: is this rehearsing an unfinished conversation, or is it avoidance of the finality of the death. Add the commercial layer as clinical material, not as an aside. The vendor can close, as HereAfter did, and a subscription that must be paid to keep speaking with a dead parent is a structure worth naming out loud before it fails.