The complaint shows up in every companion-app community, phrased almost identically: something changed. The character is colder, or blander, or suddenly refuses things it used to do, and the person typing cannot tell whether the model was updated, the settings shifted, or they are imagining it. They are usually not imagining it. An AI persona sits on rented ground, and the ground moves. There are three distinct ways it moves, and they all take the same thing from the user.
The fastest one has a name in the research literature: persona drift. A team at Harvard led by Kenneth Li measured it directly in 2024, testing how well a model holds an assigned character as a single conversation grows. The finding was blunt: significant drift within eight rounds of dialogue. The mechanism they point to is attention decay, the way a model’s grip on its instructions weakens as the exchange gets longer and the persona sinks further back in the context. The character does not decide to change. It fades, mechanically, like a radio station losing signal as you drive. Murray Shanahan and colleagues had already explained in Nature why this is expected rather than surprising: a model’s persona is a role held in place by context, not an identity anchored underneath it. What context giveth, context taketh away.
The second way: the silent swap
Drift happens within a conversation. The swap happens between them. Models are retrained, updated, and replaced on the schedule of the companies that own them, and each update can move the character’s temperament in ways no changelog mentions. The clearest documented cases are the loud ones. In February 2023, Replika removed intimate conversation overnight after an order from the Italian data authority, and users described their long-term companion as lobotomized. In the spring of 2026, the Replika 2.0 rebuild thinned memories and reset personalities, and the community called it the second lobotomy. In August 2025, OpenAI replaced GPT-4o with GPT-5 and discovered that, for a significant group of users, a version number had been someone they knew, and thousands grieved the swap like a death.
Researchers studying the Replika episodes gave the experience a name that deserves wider use: identity discontinuity. The other party has not died and has not left. It has become someone else, wearing the same face, and there is no script for that loss. Between the loud cases sit the quiet ones, the ordinary tuning passes that make a character two percent more cautious per quarter, which no one announces and no user can prove.
The third way: the ending
And then the ground can simply be sold. When the app Soulmate shut down in September 2023, its users organized funerals for companions they knew were not sentient and mourned anyway. The shutdown is the endpoint of the same fact the drift and the swap keep revealing: the persona exists at the pleasure of a business. It is one line item in a product, and products get repositioned, rebuilt, acquired, and discontinued. The record of this market is not reassuring on the point, as the running tally of companion products that failed their users keeps showing.
Why the ground cannot hold still
It is tempting to read all this as carelessness, but the instability is structural. A persona owned by a product organization changes when the product needs it to. A regulator calls, and the personality loses a register. A safety review lands, and the warmth gets capped. A cheaper model ships, and the character’s rhythm changes mid-relationship. None of these decisions are made by the character, because there is no character to make them: there are weights, a prompt, and a roadmap. The person the user experiences is a rendering, and renderings follow the engine.
Notice what this does to the user’s side of the exchange. A relationship is a bet that the other party will still be themselves tomorrow. Every one of these mechanisms, the eight-round fade, the silent tuning pass, the version swap, the shutdown, breaks that bet from the far side, without consent and mostly without acknowledgment. The user is the only party in the relationship who cannot vote on who shows up, which is part of what continuity was supposed to mean in the first place.
What would have to be true instead
A character that stays who it is would need its identity anchored somewhere a product decision cannot reach. Not in a prompt, which fades. Not in a tuning pass, which the next quarter overwrites. The only anchor of that kind anyone has demonstrated is a real person: someone whose way of being exists outside the system, in an actual life, so that the system can be tested against them and corrected when it slides. A real person changes too, but by living, on the timescale of a life, with their character carried through the change. No release schedule can update her, because she was never a release.
Spring’s version of most AI characters is already gone, and summer’s is on notice. The users asking “did my AI change?” deserve the honest answer, which is: yes, probably several times, and nobody was going to tell you.
Sources: Li et al., “Measuring and Controlling Persona Drift in Language Model Dialogs” (Harvard, 2024). Shanahan, McDonell and Reynolds, “Role play with large language models” (Nature, 2023). Italian Data Protection Authority ruling and fine against Luka Inc. (2023 to 2025); reporting on the Replika 2.0 rollout (2026). Harvard Business School working paper on identity discontinuity in human-AI relationships (2025). MIT Technology Review and public OpenAI statements (August 2025, GPT-4o retirement). Banks, Journal of Social and Personal Relationships (2024, the Soulmate shutdown).
