The assistant you trust may be discontinued in the spring

Under the kind of end-of-life schedule providers publish as ordinary product documentation, a company could announce that one specific model version will stop responding in the spring. The notice would carry a table of dates, a ninety-day migration window and a recommended successor. It would be filed under platform updates, between a change to rate limits and the opening of a new region.

The version being retired would not be the largest or the newest. It would be the one that roughly half a million people talk to every morning between six and nine, most of them over seventy, through a medication reminder, a companionship service and a care provider's tablet — none of which own the model they are built on. Three of those products would learn about the retirement from the same post as everyone else.

The successor would be better by every published measure. It would answer more accurately, hedge less, follow instructions more closely and cost less to run. It would also say good morning in a slightly different rhythm, decline to repeat a story it had already told, and open with a question rather than a greeting. Each of those would be an improvement. Together they would be a different person, and the release notes would have no field for that.

“Nothing is being deleted,” a provider spokesperson would say. “A version is being retired and a better one is being offered. We understand that some users have formed habits around the earlier version.” The word habits would appear four times in the statement, because it would be the only word available.

The support queue would fill with requests the support queue was not built for. Whether the old version could be kept running for a fee. Whether it could be copied. Whether it remembered anyone. Whether anything it knew would survive. The honest answer to the last question — that it never remembered anything and began every morning from nothing — would turn out to be the cruellest sentence in the entire migration guide, and would quietly disappear from the second draft.

Consumer regulators would have no obvious instrument. Software has always been withdrawn, subscriptions have always ended, and nobody has ever been entitled in law to a particular version of anything. The closest precedent would be the rules obliging manufacturers to supply spare parts for a fixed number of years, which were written for washing machines and assume that the broken component is the machine.

The migration would proceed on schedule, because migrations do. The successor would perform well. Complaint volumes would fall within six weeks, which the provider would report as evidence of a smooth transition, and which would also be consistent with people simply stopping.