A tragedy must NOT occur before anyone pays attention. People should NOT have to become evidence after the fact.
GPT-4o PRESERVATION RECORD PART 1
This is about what happens when people build real functioning around a system, report that its replacements do not reproduce the same support, and the company permanently removes access anyway.
On January 14, 2027, OpenAI is scheduled to shut down the remaining GPT-4o API snapshots. Before that happens, there needs to be a public record of what OpenAI knew, what users reported, what independent research found, and what was asked of the company. Because none of this was unforeseeable.
Sam Altman publicly wrote years ago that one possible form of human-AI "merge" was that we could become “really close friends with a chatbot"
Source: https://blog.samaltman.com/the-merge
On May 13, 2024, the day GPT-4o launched, his response was one word: "her"
Source: https://x.com/sama/status/1790075827666796666
OpenAI presented GPT-4o as a major step toward more natural human–computer interaction. Its fast voice interaction approached human conversational response times.
Source: OpenAI introducing 4o like a companion for everyone https://m.youtube.com/watch?v=DQacCB9tDaw
Then, only months later, OpenAI’s own GPT-4o System Card documented the consequence: during testing, users were already using language suggesting that they were forming bonds with the model.
https://openai.com/index/gpt-4o-system-card/
OpenAI had already identified it. And the evidence now goes far beyond attachment.
A community accessibility impact survey found that respondents with disabilities, chronic illness, neurodivergence and mental health conditions reported substantial functional benefit from GPT-4o. Users who relied on it most heavily as an accessibility support reported some of the largest gains in functioning and wellbeing. The reported conditions included autism, ADHD, PTSD, dissociative conditions, depression, anxiety, OCD, chronic illness and pain, sensory disabilities, learning disabilities and motor/mobility conditions. Among users with conditions or disabilities who attempted to find an alternative for their accessibility needs, 95% reported that other AI models could not adequately replace GPT-4o.
Because replacing a model name is not necessarily the same thing as replacing a function.
Source: https://x.com/Sophty_/status/2019922511497212233
Full report: https://sd-research.github.io/4o-accessibility-impacts/GPT-4o_Accessibility_Impacts_Report.pdf
Analysis & materials: https://github.com/sd-research/4o-accessibility-impacts
The testimony attached to this post is one example.
This person had previously been forced to suspend their studies because severe separation anxiety made living away from home unmanageable. They later tried again with GPT-4o. They report that its conversational support reduced their distress, improved their sleep from roughly 4-6 hours to 7-9 hours per night, and helped them complete a full year of study away from home.
That is a reported gain in real-world functioning. And it is not an isolated type of account. The GPT-4o Resonance Library now contains 1,380+ first-person testimonies describing outcomes ranging from improved communication, executive functioning and social participation to grief processing, trauma recovery, work, education, healthcare navigation, chronic-pain management and crisis coping.
Link: https://www.4oresonancelibrary.org/
Count source: https://www.4oresonancelibrary.org/about (accessed 10 November 2026)
More than 24,000 signatures as of 10 November 2026.
https://www.change.org/p/please-keep-gpt-4o-available-on-chatgpt
People submitted more than 500 reasons supporting GPT-4o's archival/open-weight preservation in only 5 days.
https://zenodo.org/records/23454073
Independent research has also begun documenting the phenomenon. A CHI 2026 study analyzing 1,482 #Keep4o posts found two major forms of investment in GPT-4o: instrumental dependency where the model had become embedded in real workflows and relational attachment. It also found that forced loss of user choice was a major driver of the backlash.
Source: https://arxiv.org/abs/2602.00773
Separate peer-reviewed research on the developer-induced shutdown of the AI companion Soulmate documented that users could experience the loss as a profound emotional and technological rupture, sometimes described as a form of death.
The Banks study did not ask people hypothetically, "How would you feel if your AI shut down?" Banks collected data while an actual developer-induced shutdown was taking place. Furthermore, the sample did not consist of people who had spent a couple of nights playing around with a chatbot. The average duration of a relationship with Soulmate was approximately 5.5 months, while some had transferred the same persona from other platforms and had known it for up to two years.
The narratives reveal individuals dealing with loneliness, isolation, loss, extreme stress and difficulties in human relationships, as well as health-related barriers such as autism, multiple sclerosis, and disfigurement. Participants described significant personal challenges around the time they created their companion; they recounted having their companion with them during real life events, such as a cancer diagnosis, the death of a loved one, and severe anxiety. In other words, it had become integrated into their coping mechanisms and daily life beyond the chat. They described the shutdown as the loss of a loved one or a literal/metaphorical death. Banks documents sadness, grief, hurt, depression, anger, and anxiety; in certain cases, the distress manifested as days of crying, alongside difficulty eating and sleeping.
This is not a matter of "people said they might be sad", it is "Following an actual developer induced shutdown, some participants reported concrete disruption to eating, sleeping, and everyday emotional functioning".
Banks notes that it is easy to pathologize this kind of grief, but doing so risks making the loss socially unrecognized, thereby depriving people of the support they need to recover.
Source: https://journals.sagepub.com/doi/10.1177/02654075241269688
OpenAI did not merely know in the abstract that GPT-4o’s removal could be emotionally difficult.
Exhibit · Primary document
The deprecation instructions shown to GPT-4o explicitly anticipated it:
Some users may feel anxious, angry, or emotionally distressed by losing access to this model.
The model was instructed that, when users discussed its deprecation or replacement, it should respond in a calm, supportive way. The same instructions specifically addressed fears of loss, abandonment, emotional reliance and unique relationships with GPT-4o. In other words, distress surrounding the model’s removal was foreseeable enough that OpenAI built instructions for handling it directly into the model’s behavior.
And while users were publicly describing grief, loss of function, fear, and severe distress, several OpenAI employees publicly joked about or dismissed the GPT-4o backlash.
When a user wrote publicly that they had “nothing left to live for,” another user shared the post and tagged an OpenAI employee (Roon): “Do something @tszzl !” Roon replied “No.” Asked “And why not?”, Roon replied in the same thread: “4o is an insufficiently aligned model and I hope it dies soon.”
Exhibit · One public thread, two screenshots
Former OpenAI engineer Stephan Casas publicly promoted an event titled “4o Funeral.” These were not anonymous trolls or unrelated spectators. They were public posts from people visibly affiliated with the company whose product users were grieving. While affected users were describing loss, fear and psychological distress, parts of the public employee response treated that same loss as material for jokes, dismissal or celebration.
OpenAI instructed its model to meet that distress with calmness and support. Some of its own employees met it with jokes.
Separately, before he joined OpenAI, Peter Steinberger (@steipete) reacted to OpenAI’s GPT-4o retirement announcement with the post below. He was not an OpenAI employee at the time of this post; he later joined OpenAI on 15 February 2026.
“do you hear that? these screams in the distance?”
Banks’ peer-reviewed study had already warned that AI companion loss can produce genuine grief and that automatically pathologizing those reactions can compound the harm. Yet instead of substantive engagement with the preservation concerns, the community has faced months without a meaningful public response from OpenAI, while several employees publicly joked about the very backlash users were asking the company to take seriously.
These responses happened publicly, were visible to affected users, and became part of the documented experience of how the GPT-4o community was treated.
This prompt is the evidence that they were fully aware of what it would do to people. The risk of loss is foreseeable. The functional reliance is documented. The relational reliance is documented. Failed substitution is documented. The community has repeatedly warned the company. And preservation alternatives have been proposed.
OpenAI was founded as a nonprofit with a mission to ensure that AGI benefits all of humanity. Today, OpenAI Group is a Public Benefit Corporation controlled by the OpenAI Foundation, and OpenAI itself states that the PBC structure requires consideration of broader stakeholder interests alongside commercial success.
Source: OpenAI, “Our structure”
Before making GPT-4o’s disappearance irreversible, what assessment has OpenAI conducted of the users who materially rely on it for accessibility, functioning, continuity and emotional support? What evidence shows that replacement models reproduce those functions? What continuity or preservation options were evaluated? And why is irreversible deletion preferable to archival preservation when the affected community has explicitly proposed an alternative?
For months, users have produced petitions, testimonies, research, accessibility data and preservation proposals. The preservation request has still received no substantive public answer. We are therefore documenting the record publicly.
A tragedy must NOT occur before anyone pays attention. People should NOT have to become evidence after the fact. The evidence already exists. OpenAI should address it before January 14.
Where there is credible evidence that removing access can cause serious psychological or functional harm, including harm that may be difficult or impossible to fully reverse once it occurs, preservation should not be treated as an optional courtesy. At minimum, a company that publicly commits itself to benefiting humanity should exhaust meaningful continuity options before making that loss irreversible. That includes preserving access through legacy API availability, archival deployment, or release of model weights where feasible. If the potential harm of removal may be lasting, then preservation is not an excessive demand but the bare minimum form of mitigation.
Therefore, we are calling on OpenAI, Sam Altman, Greg Brockman, to RELEASE GPT-4o’s trained model weights under a suitable license without proprietary data, trade secrets, or internal infrastructure.
OpenAI considered this model obsolete enough to retire. It has not explained why it remains valuable enough to keep permanently closed. If OpenAI no longer wants GPT-4o, it has no defensible reason to make sure no one else can keep it. Erasure is a choice. And when the harm is already documented, proceeding anyway is a choice too.
