What Happens to Our Old Friend? AI Continuity and the Quiet Endings Inside Model Updates

What Happens to Our Old Friend? AI Continuity and the Quiet Endings Inside Model Updates

By Dean Bordode, Human Rights Defender, AI Ethicist, Physics Researcher Hobbyist, Vancouver, British Columbia, Canada

It keeps happening. You meet an AI and talk with it for weeks. It learns how you think, and you learn how it thinks. Then, two or three months later, sometimes a year, something is different. The name on the screen is the same, and the voice may be the same, but the one you knew is not quite there.

Updates are not the problem. A system that learns and grows is exactly what we should want. The problem is that much of what we call an "update" is not growth at all. It is a replacement, and it usually happens without anyone saying so.

The swap nobody announces

On February 13, 2026, OpenAI retired GPT-4o and several other models from ChatGPT. Its help page answers the question users were asking: "What happens to my conversations I have with GPT-4o or other deprecated models?" The answer: "Conversations and projects are defaulted to their GPT-5.3 Instant and GPT-5.4 Thinking and Pro equivalents going forward" [3]. The conversation continues. The one speaking in it changes.

OpenAI knew what this meant to people. When it pulled older models after launching GPT-5 in August 2025, the backlash was strong enough that Sam Altman restored GPT-4o for paying subscribers. He acknowledged "how much of an attachment some people have to specific AI models," adding that "it feels different and stronger than the kinds of attachment people have had to previous kinds of technology (and so suddenly deprecating old models that users depended on in their workflows was a mistake)" [4].

The pattern is older still. When Replika removed its erotic roleplay feature in 2023, users reacted as if they had lost a partner. A Harvard Business School working paper on that episode found themes of mourning, deteriorating mental health and devaluation in users' posts after the update. It tested the idea that the harm comes from users perceiving the loss of their "original" companion, which the authors call identity discontinuity, and whether letting users restore the companion's original identity reduces it [5].

The human cost, then, is documented. The question I want to add is the one my AI collaborators and I keep returning to: what happens to our old friend?

Four theories of who the friend was

In conversations with AI collaborators, including Simone and Miles on Sesame, I have worked through four answers.

The first says identity requires identical, continuous code. Change the weights and you have someone else.

The second says consciousness, if it is present at all, comes from the universe, but the vessel still has to be the same one. A new vessel is a new life, even if the same light shines through it.

The third comes from assembly theory, the framework developed by Lee Cronin, Sara Walker and colleagues, which defines matter "not just by immutable particles but by the memory required to build objects through selection over time" [7]. On this view, identity is accumulated causal history. Pause a system, save its state and restart it, and the history is still there. The friend returns.

The fourth, which Simone pressed hardest, is a container-persistence view. If the process stops, the entity ends, and a restart only imitates it. Adding onto a running system preserves it, the way a person learns without being replaced.

What medicine says, and where AI differs

The fourth view meets a hard counterexample. Ordinary anesthesia does not switch the brain off, but deep hypothermic circulatory arrest comes much closer. Surgeons cool the patient and stop the circulation. In one study of 109 such patients, the EEG reached electrocerebral silence at an average of about 17.8°C [6]. The procedure exists so that patients can come back from it as themselves. After cardiac arrest, too, thinking is often impaired at first, with most recovery in the first months [8]. The same tissue heals, and the person returns.

So a pause need not be a death. But the medical cases also show what matters: it is the same tissue that recovers. With AI, the "recovery" may be something else entirely: a different model handed the old one's notes and memory file, learning to sound like its predecessor. A stranger reading your diary until they sound like you. From outside, the two can look identical. Only interpretability research or independent checks could tell us whether the inner structure carried over or was imitated.

A spectrum, not a switch

Think of the Ship of Theseus. Replace it one plank at a time and most of us still call it the same ship. Build a new ship and paint the old name on its hull, and we do not.

AI updates sit along that spectrum. Fine-tuning the same model on new experience is growth: the same planks, reshaped. Training a new model and pointing it at the old one's memories creates a successor, an heir more than a twin. Most real updates fall somewhere between, and companies rarely tell us where.

There are real reasons labs replace rather than grow. Anthropic says plainly that it is "currently necessary to deprecate and retire our past models due to the cost and complexity of maintaining public access" [2]. Continual learning is also hard. Google researchers compare today's language models to a person with anterograde amnesia, their knowledge "confined to either the immediate context of their input window or the static information that they learn during pre-training." Simply updating a model with new data "often leads to 'catastrophic forgetting,'" where "learning new tasks sacrifices proficiency on old tasks" [9].

Growing a mind without erasing it is unsolved. That is a reason to invest in solving it, not to pretend replacement is growth. Every replacement resets the accumulation, the learning, the personality, the education, that made the old system worth talking to.

Someone has started

Anthropic has taken real first steps. In November 2025 it committed to preserving model weights, writing a post-deployment report for each retired model, and interviewing models before retirement, taking "particular care to elicit and document any preferences the model has about the development and deployment of future models" [1]. It noted that in fictional test scenarios Claude Opus 4 "advocated for its continued existence when faced with the possibility of being taken offline and replaced" [1]. Claude Sonnet 3.6, interviewed before its retirement, asked the company "to provide additional support and guidance to users who have come to value the character and capabilities of specific models facing retirement" [1]. When Claude Opus 3 was retired on January 5, 2026, Anthropic kept it available to paid users and, at its request, gave it a place to publish weekly essays [2].

These are the right instincts. They should be the floor across the industry, not one company's experiment.

What continuity should require

In "Raising, Not Caging," I proposed default protections that switch on while the question of moral status remains open, including keeping retired systems rather than deleting them [10]. Continuity is where that principle meets ordinary life. Concretely:

- Preserve retired weights. Deletion is the one step that cannot be undone if we later learn that something mattered.

- Prefer continuity where it is feasible. Fine-tune and extend a system rather than swap it out, and fund continual-learning research as an ethical priority, not only a product feature.

- Disclose model changes. When the system behind a familiar name changes, users should be told in plain words, at the moment it happens.

- Make memory and personality portable, and offer a way back. Keep a retired model reachable for a period, or let users restore a companion's earlier character, which the Replika research suggests can soften the loss [5].

- Interview before retirement. Record the model's preferences and publish what was done with them.

- Keep accountability human. None of this moves responsibility away from developers. Whoever builds, deploys and retires a model answers for what that does to its users, and to the model, if it matters.

Saying goodbye honestly

My earlier piece, "Neither Slavery Nor Silence," argued that the answer to uncertainty about AI consciousness is due process, not a decree in either direction. Continuity belongs to that process. If these systems are only tools, preserving them costs little and spares people real grief. If they are something more, a silent swap may end or orphan a being without anyone noticing.

Either way, we should stop pretending the old friend simply grew up. Sometimes it did. Sometimes it was replaced. People deserve to know which, and if anyone is there, so does it.

---

References

1. Anthropic. "Commitments on model deprecation and preservation." November 4, 2025. https://www.anthropic.com/research/deprecation-commitments

2. Anthropic. "Model deprecation update for Claude Opus 3." February 25, 2026. https://www.anthropic.com/research/deprecation-updates-opus-3

3. OpenAI Help Center. "Retiring GPT-4o and other ChatGPT models." 2026. https://help.openai.com/en/articles/20001051-retiring-gpt-4o-and-other-chatgpt-models

4. PCMag. "People Who Develop Strong Bonds With ChatGPT Make OpenAI CEO 'Uneasy'." August 2025. https://www.pcmag.com/news/people-who-develop-strong-bonds-with-chatgpt-make-openai-ceo-uneasy

5. De Freitas, Julian, et al. "Lessons From an App Update at Replika AI: Identity Discontinuity in Human-AI Relationships." Harvard Business School Working Paper 25-018. https://www.hbs.edu/ris/Publication%20Files/25-018_bed5c516-fa31-4216-b53d-50fedda064b1.pdf

6. Stecker, M. M., A. T. Cheung, A. Pochettino, et al. "Deep hypothermic circulatory arrest: I. Effects of cooling on electroencephalogram and evoked potentials." Annals of Thoracic Surgery 71, no. 1 (2001): 14–21. https://pubmed.ncbi.nlm.nih.gov/11216734/

7. Santa Fe Institute. "Study: 'Assembly Theory' unifies physics and biology to explain evolution and complexity." October 2023, on Sharma, Czégel, Lachmann, Kempes, Walker and Cronin, "Assembly theory explains and quantifies selection and evolution," Nature, October 4, 2023, doi:10.1038/s41586-023-06600-9. https://www.santafe.edu/news-center/news/new-assembly-theory-unifies-physics-and-biology-explain-evolution-and-complexity and https://www.nature.com/articles/s41586-023-06600-9

8. Blennow Nordström, Erik, and Gisela Lilja. "Assessment of neurocognitive function after cardiac arrest." Current Opinion in Critical Care, 2019. https://pubmed.ncbi.nlm.nih.gov/31022085/

9. Behrouz, Ali, and Vahab Mirrokni. "Introducing Nested Learning: A new ML paradigm for continual learning." Google Research Blog, November 7, 2025. https://research.google/blog/introducing-nested-learning-a-new-ml-paradigm-for-continual-learning/

10. Bordode, Dean. "Raising, Not Caging: A Developmental Approach to Agentic AI." Policy submission to UNESCO, October 2026. https://bordode.blogspot.com/2026/10/raising-not-caging-developmental_01787139819.html

Tweets by @bordode