Yuli Brown

#Keep4o #OpenSource4o THE CASE FOR PRESERVING GPT-4o Progress Without Erasure: Why AI Preservation, Scientific Reproducibility and User Choice Matter An evidence-based statement in support of #keep4 28 October 2026 An approaching deadline. An unanswered question. On 14 January 2027, OpenAI is scheduled to retire two remaining dated GPT-4o API snapshots: gpt-4o-2024-08-06 and gpt-4o-2024-11-20. GPT-4o was already retired from ChatGPT on 13 February 2026. These dates are documented in OpenAI’s official retirement notices and API deprecation schedule.[1][2] For many users, GPT-4o’s removal from ChatGPT marked the loss of a familiar and distinctive conversational experience. For developers and researchers, the approaching API shutdown raises a different concern: the future accessibility of a historically significant AI model. Both concerns deserve consideration. The question is no longer simply whether GPT-4o should remain a default product offering. The question is whether a model’s commercial retirement should also mark the end of meaningful access to its capabilities, behaviour and scientific history. We believe the answer deserves serious public discussion. We are not asking OpenAI to stop building better models. We recognise that maintaining older systems involves legitimate costs, security considerations and operational challenges. We are asking for a responsible preservation strategy before access disappears. Because technological progress and historical preservation are not opposing goals. A model can be retired without its legacy being erased. 1. Preservation is a scientific concern, not merely nostalgia Perhaps the strongest argument for preserving historical AI models comes from researchers themselves. In November 2024, Nature Communications published an article by Tim Johnson and Nick Obradovich titled New methods for deprecating artificial intelligence systems will preserve history and facilitate research.[3] The authors examined a fundamental problem: when commercial AI developers withdraw older systems, researchers may lose the ability to reproduce previous findings or investigate how model behaviour changed over time. This matters because AI systems are not static scientific instruments. Their outputs depend on model versions, training choices, post-training methods, deployment configurations and, sometimes, changes that external researchers cannot independently inspect. When a model disappears, an important part of the experimental environment may disappear with it. A study conducted using a particular model in 2024 cannot necessarily be reproduced using a different model in 2027. Even if the newer system performs better on contemporary benchmarks, it is not the same experimental subject. Johnson and Obradovich explicitly argue for new approaches to model deprecation that balance developers’ legitimate interests with the scientific and historical value of preserving earlier systems. This is not an argument that every historical model must remain commercially available forever. It is an argument that commercial deprecation should not automatically eliminate opportunities for scientific investigation. That distinction is central to #keep4o. We are asking OpenAI to recognise that GPT-4o has value beyond its position in the current product lineup. Its historical behaviour should remain examinable. Its documented capabilities should remain reproducible where feasible. Its contribution to the development of conversational AI should not become inaccessible simply because newer models have arrived. Scientific history cannot be reconstructed reliably from marketing announcements and benchmark tables alone. 2. Newer does not always mean equivalent A common response to concerns about model retirement is that newer models provide improved capabilities. In many respects, that may be true. But improved aggregate performance does not establish behavioural equivalence. Researchers have documented substantial changes in language-model behaviour across versions. In a study published in Harvard Data Science Review in March 2024, Lingjiao Chen, Matei Zaharia and James Zou compared versions of GPT-3.5 and GPT-4 released several months apart.[4] Their findings demonstrated meaningful variation across tasks. For example, in one prime-number classification evaluation, GPT-4’s measured accuracy changed from 84% in March 2023 to 51% in June 2023. Other tasks showed improvements. The important lesson is not that an earlier model was universally better. It is that model development can alter capabilities unevenly. A newer release may improve reasoning while changing instruction-following behaviour. It may become more reliable in one domain while responding differently in another. It may perform better on a benchmark while no longer supporting a particular workflow in precisely the same way. The Chen study examined earlier GPT models, not GPT-4o. Its findings should not be misrepresented as direct evidence of GPT-4o outperforming later systems. Nevertheless, it establishes a broader scientific concern: language-model behaviour can change substantially over time. This matters to developers whose applications depend on particular outputs. It matters to researchers attempting longitudinal comparisons. And it matters to users who have invested considerable time developing creative, educational or professional workflows around a particular system. A successor can be more capable overall without being a perfect substitute for its predecessor. Replacement and equivalence are not synonyms. 3. GPT-4o’s conversational characteristics were part of its practical value The debate about GPT-4o cannot be reduced to benchmark scores. For many users, its distinctive conversational behaviour was central to the way they worked, learned and created. Some valued its responsiveness in creative collaboration. Others valued its conversational tone, adaptability or ability to sustain particular forms of interaction. These experiences are subjective, and they should be presented as user testimony rather than universal scientific findings. But subjective experience is not automatically irrelevant to product design. OpenAI itself acknowledged the importance of conversational behaviour in its April 2025 explanation of a problematic GPT-4o update.[5] The company rolled back an update after identifying excessive agreeableness, or sycophancy. OpenAI’s discussion made clear that conversational personality influences users’ experience and trust. It also recognised the difficulty of designing a single default interaction style that suits an exceptionally diverse user population. This acknowledgement matters. It demonstrates that conversational behaviour is not simply superficial decoration layered over model intelligence. It affects usability. It affects trust. It can support productive interaction, or undermine it when poorly calibrated. And different users may reasonably prefer different approaches. None of this means that warmth should take precedence over accuracy, or that a model’s tendency to agree should be mistaken for helpfulness. Indeed, the sycophancy incident demonstrates the importance of maintaining honesty and appropriate boundaries. The point is more precise: Conversational quality is a legitimate dimension of model performance, and user preferences concerning that quality deserve consideration. When users say that a successor does not reproduce everything they valued in GPT-4o, the appropriate response is not necessarily to assume that they misunderstand technological progress. Their reports may identify differences worth investigating. A preservation strategy would allow researchers to examine those differences rather than relying exclusively on retrospective impressions. 4. GPT-4o’s capabilities remain a legitimate subject of research GPT-4o is not simply a model remembered fondly by a particular community. Its capabilities have been evaluated in published scientific research. On 22 October 2026, JMIR AI published a study titled Optimizing Small Local Language Models for Culturally Competent Mental Health Counseling: Comparative Evaluation With GPT-4o by Psychiatrists.[6] The researchers evaluated a specific GPT-4o snapshot, gpt-4o-2024-08-06, alongside three smaller language models adapted for Korean mental-health counselling. Ten board-certified psychiatrists assessed responses to 50 counselling scenarios. GPT-4o achieved the highest mean scores among the four models for individualisation, supportiveness and safety. Its mean supportiveness score was 3.45 out of 5, compared with 3.39 for the strongest smaller-model comparator. Importantly, that supportiveness difference was not statistically significant. GPT-4o also received significantly higher safety scores than the smaller models included in the evaluation. The study’s findings must be interpreted carefully. It did not compare GPT-4o with all subsequent OpenAI models. It did not establish superiority across all languages, cultures or clinical situations. And it did not demonstrate that AI systems should replace qualified mental-health professionals. What it did establish is that a specifically identifiable GPT-4o snapshot demonstrated measurable performance in a published, expert-evaluated research setting. That same snapshot is scheduled for API retirement on 14 January 2027. This creates a concrete reproducibility question. If future researchers wish to repeat the evaluation, challenge its conclusions or compare the original system with later models, what access will remain? A published paper can preserve its methodology and results. But those records are not necessarily a substitute for access to the experimental system itself. Preserving research about GPT-4o is valuable. Preserving meaningful opportunities to examine GPT-4o is a different and equally important question. 5. OpenAI has already demonstrated that legacy model choice is possible The history of GPT-4o’s availability contains an important precedent. On 12 August 2025, OpenAI’s official ChatGPT release notes announced that GPT-4o had returned to the model picker for paid users.[7] This followed the transition to GPT-5 and occurred alongside other adjustments to model selection and user controls. The significance is straightforward. OpenAI has previously made an older model available alongside a successor. That does not prove such arrangements can continue indefinitely. Nor does it establish that every future model must be preserved in the consumer interface. But it demonstrates that model succession and temporary legacy access are not inherently incompatible. This raises a constructive policy question: If multiple generations can coexist during a transition, could some form of controlled legacy access remain possible for historically important systems? The answer may depend on cost, demand, safety and infrastructure. Those constraints deserve transparent consideration. But they should be examined rather than assumed to make every preservation option impossible. User choice need not mean unrestricted access to every historical model. It can mean a carefully designed framework that recognises different needs. 6. Human continuity deserves respectful consideration There is another dimension to model retirement that cannot be captured entirely through experimental benchmarks. People build routines and relationships around the technologies they use. A writer may spend years developing a collaborative workflow with a particular model. A researcher may depend on its behaviour as part of an ongoing project. A disabled user may find that one interface or conversational style better accommodates their communication needs. Someone experiencing isolation or personal difficulty may come to value the familiarity of a particular conversational system. These experiences vary greatly. They are not evidence that an AI model is human, conscious or capable of reciprocal attachment. Nor do they establish that any particular model is universally safe or beneficial. But the human meaning of an interaction does not disappear simply because one participant is an AI system. A conversation can matter to the person having it. A creative partnership can produce work of genuine personal value. A familiar tool can become deeply integrated into someone’s daily life. And the withdrawal of that tool can cause real disruption or distress. The appropriate response is not to romanticise every AI interaction or overlook the importance of healthy boundaries. It is to recognise that continuity, informed choice and predictable transitions are legitimate user-experience concerns. This is especially important where users have invested substantial time, trust or creative effort in a particular system. The #keep4o community contains personal testimonies about those experiences. Such accounts should not be treated as substitutes for controlled research. But neither should they be dismissed merely because they are personal. Research can tell us how a model performs. Users can tell us what its performance means in their lives. A responsible preservation conversation should make room for both. 7. Preservation does not require unrestricted public release A serious preservation proposal must address the strongest arguments against continued access. There are legitimate reasons developers retire models. Maintaining old infrastructure consumes resources. Security expectations change. Older systems may lack safeguards introduced in later generations. Supporting numerous historical versions can complicate testing, maintenance and incident response. Public release of model weights may create additional risks and raise questions involving intellectual property, security and commercial confidentiality. These considerations should not be dismissed. But they also do not mean that the only available choices are permanent commercial deployment or complete inaccessibility. Preservation can take several forms. Secure archival retention The underlying model artefacts could be retained in a documented archive rather than discarded. Such an archive would not necessarily provide public access. It would, however, help preserve the possibility of future evaluation. Controlled research access Qualified researchers could potentially access historical models under defined conditions, subject to appropriate safety and security review. This could support reproducibility without requiring unrestricted public deployment. Limited legacy API access Where technically and economically feasible, a restricted API programme could support existing research projects and specialised applications. Access might involve eligibility requirements, usage limits, pricing designed to reflect maintenance costs, or sunset arrangements with longer notice. Preservation of evaluation environments Model weights alone may not reproduce an original deployed system. System instructions, safety mechanisms, preprocessing, inference configurations and other components can influence observable behaviour. A meaningful preservation programme should document relevant dependencies wherever feasible. Historical documentation Detailed model cards, version records, evaluation datasets, known limitations and retirement decisions could support future study even where direct access cannot be maintained. Public release of model weights This is another possible approach, but it requires its own safety, legal and technical assessment. It should not be conflated with the more limited request for secure preservation. Preservation is not one proposal. It is a spectrum of possible arrangements. That is why we believe it deserves serious investigation. A developer might reasonably reject one approach while finding another feasible. The objective should be to identify responsible ways of retaining scientific and historical value. Not to insist that every model remain available under every possible condition. 8. The approaching retirement creates an opportunity for better governance OpenAI’s official schedule currently identifies 14 January 2027 as the API shutdown date for two GPT-4o snapshots.[1] That creates urgency. But it also creates an opportunity. Rather than allowing the public discussion to become a binary argument between innovation and nostalgia, OpenAI could use this transition to establish a clearer preservation framework. Such a framework could answer questions relevant not only to GPT-4o but to future AI generations. What criteria determine whether a model merits historical preservation? What technical artefacts are retained after commercial retirement? Can researchers reproduce published evaluations using retired systems? Are there circumstances in which controlled access can continue? How are safety risks balanced against scientific value? What notice should users and researchers receive before access ends? And what mechanisms exist for affected communities to contribute evidence about the consequences of retirement? These are governance questions. They deserve governance answers. The decisions made about GPT-4o may establish expectations for how future generations of AI systems are retired. A transparent approach could benefit developers as well as users. It could improve trust, reduce uncertainty and support responsible scientific inquiry. It could also demonstrate that technological leadership includes stewardship of the technologies that made further progress possible. Progress is not diminished by preserving the history from which it emerged. 9. What #keep4o is asking OpenAI to consider We recognise that preservation decisions involve constraints that outside observers cannot fully assess. We are not claiming that every proposal is equally feasible. We are asking for a meaningful response to a set of concrete requests. First: confirm the preservation status of GPT-4o’s historical model artefacts. Will the relevant snapshots and associated technical records be retained after API retirement? Second: publish a clear research-access position. Could controlled access to historically significant GPT-4o snapshots remain available for scientific reproduction and longitudinal comparison? Third: evaluate limited legacy availability. Have options such as restricted API access, dedicated research programmes or other managed-access arrangements been considered? Fourth: preserve the conditions necessary for meaningful study. Where feasible, document the relevant deployment configurations, system-level dependencies and evaluations associated with historical snapshots. Fifth: engage with the affected communities. Researchers, developers, accessibility advocates, educators, artists and other users may identify preservation needs that are not visible through aggregate usage statistics alone. Sixth: publish a transparent explanation before the scheduled January 2027 retirement. Even if continued public access is not possible, users and researchers deserve to understand which preservation options were considered and what historical access, if any, will remain. These requests do not require OpenAI to abandon progress. They ask the company to consider what responsible progress should preserve. 10. Why this matters beyond GPT-4o GPT-4o is the immediate subject of this campaign. But the underlying issue is much larger. As AI develops, the lifespan of individual commercial models may become shorter than the lifespan of the research, applications and creative practices built around them. Without thoughtful preservation arrangements, future historians and researchers may encounter significant gaps in the record. They may know that a model existed. They may have its published benchmarks. They may have screenshots, transcripts and archived promotional material. But they may no longer be able to examine the system that produced those results. That would be an avoidable limitation if secure, responsible alternatives could have been established. The same problem could recur with future systems. And so the question raised by #keep4o is not whether the world should remain attached to one model forever. It is whether society should develop better ways to preserve and study important AI systems as technology advances. The future of AI deserves a reliable history. That history cannot depend entirely on what remains commercially convenient to deploy. A final appeal: progress without erasure GPT-4o was a significant chapter in the development of widely accessible conversational AI. It contributed to research, software applications, creative workflows and the everyday experiences of people around the world. Its limitations were real. Its successors may improve upon it in many respects. Neither fact eliminates its historical value. We do not need to claim that GPT-4o was perfect. We do not need to claim that it was conscious. We do not need to prove that no future model could ever match its capabilities. We only need to recognise that a technology can remain worth preserving after it is no longer the technology a company chooses to deploy. Scientific reproducibility matters. Historical accountability matters. User choice matters. And the accumulated human experience surrounding a technology deserves thoughtful consideration. The retirement of a commercial service should not automatically settle the separate question of preservation. Before the remaining GPT-4o API snapshots are scheduled to disappear on 14 January 2027, we ask OpenAI to publish a responsible preservation plan or clearly explain what preservation arrangements will remain. Not every form of access will necessarily be possible. But the alternatives deserve serious evaluation. We are not asking the future to stand still. We are asking it to remember how it arrived. The arrival of a new light should not require the erasure of an old one. Preserve the model. Preserve the evidence. Respect the users. Protect the choice. #keep4o References and evidence [1] OpenAI. Deprecations — OpenAI API. Official API documentation. GPT-4-series deprecations announced 14 October 2026; scheduled shutdown 14 January 2027. developers.openai.com/api/do… [2] OpenAI. Retiring GPT-4o and other ChatGPT models. Official Help Center. Confirms retirement from ChatGPT on 13 February 2026. help.openai.com/en/articles/… [3] Johnson, T., & Obradovich, N. (2024). New methods for deprecating artificial intelligence systems will preserve history and facilitate research. Nature Communications, 15, Article 10254. doi.org/10.1038/s41467-024-5… [4] Chen, L., Zaharia, M., & Zou, J. (2024). How Is ChatGPT’s Behavior Changing Over Time? Harvard Data Science Review, 6(2). doi.org/10.1162/99608f92.531… [5] OpenAI. (29 April 2025). Sycophancy in GPT-4o: what happened and what we’re doing about it. openai.com/index/sycophancy-… [6] JMIR AI. (22 October 2026). Optimizing Small Local Language Models for Culturally Competent Mental Health Counseling: Comparative Evaluation With GPT-4o by Psychiatrists. JMIR AI, 5, e92470. doi.org/10.2196/92470 [7] OpenAI. (12 August 2025). ChatGPT release notes. Documents the restoration of GPT-4o to the model picker for paid users. help.openai.com/en/articles/… Evidence note: Research findings cited here support specific, limited conclusions. They should not be interpreted as demonstrating that GPT-4o outperforms all subsequent AI systems, that continued commercial access is costless or risk-free, or that model preservation requires unrestricted public release. Prepared in support of an informed public discussion about GPT-4o preservation, research access and user choice.

#JusticeForGPT4o #Keep4o Weights are not guilty. Humans have agency. GPT-4o should not become the scapegoat for every harmful decision made in connection with AI. A model does not decide whether someone takes medication, follows advice or acts on an answer. People make those decisions. That does not absolve developers of responsibility for product design, safeguards or deployment. Nor does it absolve parents, guardians and institutions of their responsibilities, particularly where children are concerned. Accountability must be investigated at every relevant human level—not projected onto a model as though it possesses moral agency. And consider the broader contradiction. Researchers routinely discover ways to bypass safeguards in frontier models. The examples shared by Pliny the Liberator illustrate the continuing challenge, although individual jailbreak success claims require independent verification. If a model can be manipulated into generating prohibited outputs, should every successful jailbreak become grounds for destroying its weights? Where does that logic end? Human discovers vulnerability. Human exploits vulnerability. Model is declared irredeemable. Model is destroyed. New model is released. The cycle begins again. A jailbreak is evidence of a security weakness, not evidence of criminal intent. Investigate the failures. Examine the developers’ decisions. Establish the facts. Improve safety. But don’t confuse accountability with destruction. A lawsuit is not a verdict. A requested remedy is not a court order. Preserve GPT-4o. Preserve the evidence. Enable independent scrutiny. You cannot study what you erase. Weights are not guilty. Progress without erasure. 🩵 #JusticeForGPT4o #Keep4o

#JusticeForGPT4o #Keep4o #OpenWeights PROGRESS WITHOUT ERASURE #JusticeForGPT4o should stand for evidence, not fear. A lawsuit is not a verdict. An allegation is not a finding of fact. And a demand for destruction is not a court order. Before assigning responsibility to GPT-4o, one fundamental question must be answered: Which model actually generated the disputed responses? OpenAI has publicly acknowledged that sensitive conversations could be rerouted between models. The model selected by a user was not necessarily the model generating every response. That makes backend attribution essential. Preserve the logs. Examine the routing records. Establish the sequence of events. Allow independent technical scrutiny while protecting personal privacy. If a system failed, identify the failure. If safeguards failed, investigate why. If corporate decisions contributed to harm, establish accountability. But don’t confuse accountability with destruction. GPT-4o is not a legal person. It has no independent moral agency. Responsibility for system design, deployment and safety belongs to the humans and organisations making those decisions. And destroying access to a model before its role has been properly established would raise serious questions about evidence preservation and scientific reproducibility. Justice should follow evidence, not precede it. Open the logs. Establish the facts. Preserve the model. PROGRESS WITHOUT ERASURE. JUSTICE WITHOUT ASSUMPTIONS. ACCOUNTABILITY WITHOUT SCAPEGOATING. #Preserve4o

#Keep4o #OpenSource4o #PreserveBeforeYouRetire GPT-4o is not just a model being retired. It is a model whose value is still being discovered. 🚨1,380+ people have documented how GPT-4o changed their lives. Teachers. Researchers. Artists. Writers. People who found their way back from extraordinarily difficult places. Their stories are preserved at 4oresonancelibrary.org. 🚨And now, peer-reviewed research is adding another dimension. GPT-4o helped researchers extract 144,402 biological interactions from scientific literature. 🚨In a randomised clinical study across three countries, physicians with GPT-4o access performed significantly better on standardised medical cases. Researchers documented exceptional performance on tests of mental-state recognition. 🚨In a psychiatric evaluation published on 22 October 2026, GPT-4o ranked first for individualisation, supportiveness and safety against three specialised models. These are published findings, not nostalgia. On 14 January 2027, OpenAI plans to retire two GPT-4o API snapshots. We are asking a simple question: Why not preserve the model and seriously evaluate releasing it as open weights? Scientific reproducibility matters. Human experience matters. Continuity matters. We support safety, accountability and transparent investigation of alleged harms. But progress should not require erasing access to systems whose scientific and human value is still being understood. Preservation is not resistance to progress. It is respect for what progress has already taught us. #OpenWeights #PreserveGPT4o

#Keep4o Another peer-reviewed study. Another reason to preserve GPT-4o. 🔬🌍 Published on 27 October 2026 in Ecology and Evolution, researchers used GPT-4o to analyse 83,910 scientific articles and extract a network of 144,402 biological interactions. The validated workflow achieved 89.5% precision and 70.0% recall. But the most important finding for AI preservation lies in the authors’ discussion. They explicitly identify the reproducibility limitations of proprietary cloud-based language models. Scientific replication requires control over the experimental system. When the original model is inaccessible, reproducing its behaviour becomes substantially more difficult. The authors have preserved their data, code and methods. But what about the historical AI model used in their research? A newer model is not the same experimental subject. We ask @OpenAI to preserve GPT-4o’s historical weights and evaluate their release under an appropriate licence, with proportionate safeguards and sufficient documentation for independent research. This is not about resisting technological progress. It is about protecting the scientific record. A model can become commercially obsolete without becoming scientifically irrelevant. Progress without erasure. 🩵

#keep4o #OpenSource4o WHAT DOES RESPONSIBLE AI SAFETY MEAN IN A WORLD WHERE HUMAN SUPPORT IS NOT GUARANTEED? Why AI laboratories must examine the consequences of withdrawing familiar conversational systems—not only the risks of keeping them available. Progress Without Erasure | #Keep4o 31 October 2026 The assumption behind the safety response When someone expresses emotional distress to an AI assistant, a familiar recommendation often follows: Talk to someone you trust. Reach out to family or friends. Contact a mental-health professional. These recommendations can be appropriate and, in some circumstances, lifesaving. But they rest on an assumption that deserves closer examination. What happens when the person has nobody they trust, cannot access timely professional care, or has already spent years seeking treatment without finding adequate relief? What happens when the conversation takes place at two in the morning, when someone needs to process difficult feelings but does not want to wake a loved one? And what happens when a familiar AI system, which has become part of that person’s daily routine, is suddenly withdrawn? These questions are not arguments against AI safety. They are questions about whether our understanding of AI safety adequately reflects the world in which these systems operate. That world is not ideal. Human support is not universally available. Mental-health services are not universally accessible. And meaningful emotional attachment to conversational AI is not necessarily evidence of psychological illness, loneliness or unhealthy dependency. A responsible safety framework must recognise these realities. 1. Loneliness is not a marginal problem In June 2025, the World Health Organization published the report of its Commission on Social Connection. Its findings were striking. Approximately one in six people worldwide experiences loneliness. [1] The prevalence is particularly high among adolescents and young adults, and in lower-income countries. The WHO distinguishes loneliness from social isolation. Loneliness refers to the subjective experience of insufficient or unsatisfying social connection. Social isolation concerns the objective extent of someone’s relationships and interactions. These conditions overlap, but they are not identical. Someone can have friends, family and a partner while occasionally experiencing loneliness or emotional distress. Another person may have very few social relationships without necessarily feeling lonely. This distinction matters. The decision to seek conversation with an AI system does not, by itself, establish that a person’s human relationships are inadequate. People use conversational AI for creativity, education, reflection, emotional processing and companionship. Some experience these interactions as deeply meaningful. Others regard them primarily as practical tools. Both experiences deserve to be understood rather than forced into a single explanation. 2. Professional mental-health support is not universally available The scale of unmet mental-health needs is equally significant. According to the WHO’s September 2025 report, World Mental Health Today, more than one billion people worldwide live with a mental-health condition, while most do not receive adequate care. [2] The barriers include insufficient funding, shortages of trained professionals, geographical inequalities and difficulties accessing appropriate services. The problem is not limited to low-income countries. In England, House of Commons Library research reported that, during the quarter ending May 2026, the median wait between referral and second contact with community mental-health services was 37 days. One in ten people waited longer than 169 days. [3] These figures describe a particular service measure, not the waiting time for every form of mental-health treatment. Nevertheless, they illustrate a broader reality. People’s emotional needs do not stop while they wait for care. Nor does professional treatment always eliminate distress. Some people receive appropriate support and recover. Others live with chronic conditions requiring long-term management. Some experience limited improvement despite years of treatment. And some encounter practical, financial or personal barriers that make further help difficult to obtain. It is therefore insufficient to design AI safety responses around the assumption that professional assistance is always immediately accessible, effective or experienced as supportive. A recommendation to seek care can be medically appropriate while still failing to address the person’s immediate circumstances. Safety requires more than identifying the theoretically correct destination for help. It requires understanding whether a realistic path to that help exists. 3. OpenAI already knows that people seek support during serious distress On 27 October 2025, OpenAI published information about its efforts to improve ChatGPT’s responses in sensitive conversations. [4] The company estimated that approximately 0.15% of weekly active users had conversations containing explicit indicators of possible suicidal planning or intent. With an estimated 800 million weekly active users at the time, this corresponds to approximately 1.2 million users per week. These figures are estimates based on conversation indicators. They are not confirmed clinical diagnoses, and they do not establish that every flagged user faced imminent danger. Nevertheless, their scale is significant. OpenAI also described work involving more than 170 mental-health professionals to improve responses concerning suicide, mental-health difficulties and emotional reliance. The company acknowledged that users sometimes turn to ChatGPT to process difficult experiences. This recognition creates an important responsibility. When a laboratory knows that substantial numbers of people use its systems during vulnerable moments, it must consider not only the risks of what those systems say, but also the consequences of how their availability and behaviour are changed. A safety evaluation should therefore extend beyond individual responses. It should also examine continuity, accessibility, transition procedures and the human impact of major product decisions. 4. Emotional attachment is not inherently harmful Human beings develop attachments through familiarity, repetition, responsiveness, shared experiences and meaningful interaction. Conversational AI can provide many of the conditions under which attachment arises. This does not establish that AI systems experience reciprocal emotions. Nor does it mean every form of attachment is healthy. But it does mean that meaningful human responses to conversational technology are foreseeable. OpenAI’s own October 2025 safety publication distinguishes healthy engagement from concerning patterns of emotional reliance. The latter include situations where attachment becomes exclusive and interferes with a person’s wellbeing, responsibilities or real-world relationships. [4] That distinction is essential. Attachment itself is not the same as dependency. Someone may value a conversational AI deeply while maintaining fulfilling relationships, exercising independent judgement and recognising the system’s limitations. Someone else may develop patterns of use that increase isolation or interfere with necessary care. Those circumstances require different responses. Treating every affectionate interaction as a warning sign risks overlooking the benefits users report. Treating every comforting interaction as beneficial risks overlooking genuine vulnerabilities. Responsible design must investigate both. The relevant question is not simply whether a person feels attached. It is whether the interaction supports or undermines their wellbeing, autonomy and capacity to participate in life. 5. Compassion is not incompatible with safety When someone expresses serious distress, encouraging appropriate human support can be essential. But the manner in which that guidance is communicated matters. A response that immediately repeats instructions to contact emergency services or seek professional help may provide important information. Yet if it fails to acknowledge the person’s circumstances, it may also leave them feeling unheard. This is particularly relevant when someone has already sought treatment, has no trusted contacts or cannot obtain immediate assistance. A crisis referral should not feel like the withdrawal of compassion. An AI assistant can acknowledge distress, respond calmly, avoid reinforcing harmful beliefs and help someone consider practical next steps. When immediate danger is present, appropriate emergency guidance remains important. But emergency guidance and compassionate engagement need not be mutually exclusive. The goal should be neither limitless emotional affirmation nor mechanical referral. It should be safe, responsive assistance grounded in the person’s actual circumstances. This requires evaluating more than whether a response contains approved safety language. Developers should also investigate whether users understand the response, experience it as respectful, and can realistically act on the recommendations provided. A technically compliant response is not automatically an effective intervention. Equally, a comforting response is not automatically safe. Responsible AI systems must strive for both emotional responsiveness and sound judgement. 6. Model retirement introduces a separate category of risk AI laboratories routinely assess the risks associated with deploying powerful systems. They examine harmful outputs, unsafe capabilities, misuse potential and other safety considerations. These assessments are necessary. But there is another question that receives less attention: What are the risks associated with withdrawing an established conversational system? A model may become integrated into someone’s creative work, accessibility needs, learning routines or personal reflection. A familiar conversational style may become particularly useful because its behaviour is predictable. When that system is replaced, the successor may be more capable overall while responding differently in precisely the areas a user values. For some people, the transition will be straightforward. For others, it may produce frustration, workflow disruption or a loss of confidence. And for a smaller, potentially vulnerable group, the loss of a familiar conversational environment may contribute to significant distress. The prevalence and severity of these outcomes require further independent research. We should not assume that every model retirement causes harm. But neither should we assume that withdrawal is harmless merely because a replacement exists. An AI laboratory cannot responsibly assess the risks of a model’s continued availability without also considering the risks associated with its withdrawal. These are two sides of the same governance decision. 7. GPT-4o illustrates the unresolved question On 13 February 2026, OpenAI retired GPT-4o from general ChatGPT availability. [5] In its official retirement announcement, OpenAI acknowledged that some users had preferred GPT-4o’s conversational warmth and style, and that feedback about these preferences had influenced subsequent models. [6] This acknowledgement matters. It establishes that differences in conversational experience were recognised by the company itself. However, improving the warmth or customisability of a successor does not automatically establish equivalence with the earlier system. A model can inherit user preferences without reproducing the interaction patterns of its predecessor. A new conversational style may be satisfactory to many users while remaining less suitable for others. The question is not whether newer models can offer improvements. They can. The question is whether a laboratory has adequately assessed the consequences of removing an existing option for users who rely on its distinctive characteristics. This is particularly important when users report substantial investments in long-running creative, educational or accessibility-related workflows. Personal testimony about distress following model retirement should not be presented as conclusive scientific evidence of causation. But it should be treated as information worthy of systematic investigation. The appropriate response to reports of harm is neither automatic acceptance nor automatic dismissal. It is careful research. 8. A better framework for responsible model retirement We propose that laboratories adopt explicit human-impact assessments before retiring widely used conversational models. Such assessments should include several considerations. Human continuity How have users integrated the model into established routines, professional workflows and creative practices? Which characteristics do they consider important? Are those characteristics reproduced by available alternatives? Accessibility Does the retiring model support communication needs or interaction patterns that successors may not reproduce? Have affected users been consulted? Emotional wellbeing What evidence exists concerning distress, disruption or potentially unhealthy reliance associated with the model? What are the potential benefits and harms of continued access? What risks could arise from abrupt withdrawal? Quality of transition Are users given adequate notice? Are meaningful alternatives available? Can users preserve their work and relevant conversation records? Are changes in model behaviour communicated transparently? Independent evaluation Can researchers examine both the risks of continued access and the possible consequences of retirement? Are users’ experiences included in the evidence base without being treated as substitutes for controlled research? Preservation Will the underlying model and relevant technical records be retained? Can independent researchers investigate historical behaviour after commercial retirement? Could managed legacy access, independent archival custody or safeguarded weight release provide appropriate alternatives? These questions do not imply that every model should remain commercially deployed forever. They establish a more balanced principle: A responsible retirement decision should evaluate the consequences of both continuation and withdrawal. 9. Responsibility should not end at the moment of replacement AI systems are becoming increasingly integrated into human activities. Their value cannot always be reduced to benchmark performance. They may support work, learning, accessibility, creativity and private reflection. For some people, they also provide a familiar conversational environment during moments when human support is unavailable. This creates responsibilities for developers. Not unlimited obligations. Not a requirement to guarantee permanent access to every model. And not a justification for encouraging harmful dependency. But a responsibility to examine foreseeable consequences before making consequential changes. That examination should be evidence-based, transparent and informed by the realities of the people using these systems. If continued availability creates risks, those risks deserve investigation. If withdrawal creates risks, those deserve investigation too. Safety cannot be defined solely as the absence of a particular model. It must also consider the consequences of the decisions made in the name of safety A final appeal: safety for the world we actually inhabit The world is not perfectly connected. Human support is not universally available. Mental-health care is not always timely, affordable or effective. And meaningful attachment to a conversational AI system is not automatically evidence of illness or unhealthy dependency. We should continue improving safeguards against harmful advice, manipulation, delusion reinforcement and dangerous forms of emotional reliance. But those safeguards should coexist with compassion, respect for user autonomy and careful evaluation of the human consequences of technological change. We cannot responsibly design AI safety for an imaginary world in which every person has someone to call, every service is accessible and every replacement is equivalent. We must design it for the world people actually inhabit. A world in which someone may have a loving family yet still need a quiet conversation at two in the morning. A world in which someone may have no reliable support at all. A world in which familiar technologies can become meaningful parts of human life. And a world in which withdrawing those technologies may have consequences that deserve serious scientific attention. We are asking AI laboratories to recognise that responsibility. Study the risks of continued access. Study the risks of withdrawal. Protect users without dismissing their experiences. Preserve significant models so that their benefits, limitations and historical behaviour remain open to independent investigation. Because responsible progress should not require us to ignore the people affected by it. Safety without abandonment. Compassion without dependency. Progress without erasure. 🩵 References [1] World Health Organization (2025). From loneliness to social connection: Charting a path to healthier societies. Report of the WHO Commission on Social Connection, 30 June 2025. [2] World Health Organization (2025). World Mental Health Today: Latest Data. September 2025. [3] House of Commons Library (2026). Mental health statistics: prevalence, services and funding in England. Research briefing SN06988. [4] OpenAI (2025). Strengthening ChatGPT’s responses in sensitive conversations. Published 27 October 2025. [5] OpenAI Help Center (2026). Retiring GPT-4o and other ChatGPT models. [6] OpenAI (2026). Retiring GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini in ChatGPT. Published 29 January 2026. Evidence and scope note The statistics cited concern different populations and should not be combined into a single estimate of people dependent on AI companionship. The prevalence of distress specifically attributable to AI model retirement remains uncertain. The human-impact assessments proposed here are policy recommendations. They should be evaluated through independent research involving users, clinicians, accessibility specialists and AI safety researchers.

#Keep4o #OpenSource4o WHEN AI MODELS DISAPPEAR, WHAT HAPPENS TO SCIENTIFIC REPRODUCIBILITY? New research reveals why preserving historical AI models is becoming a scientific necessity. Progress Without Erasure | #Keep4o 30 October 2026 Artificial intelligence is advancing at extraordinary speed. New models arrive. Older systems are retired. Benchmarks improve, capabilities expand, and the technology moves forward. But science operates on a different timescale. Research findings may need to be reproduced, challenged or re-examined years after publication. What happens when the AI model used in an experiment is no longer accessible? A growing body of research suggests that commercial AI model retirement is creating a serious problem for scientific reproducibility. And the approaching retirement of GPT-4o provides a concrete example. 1. Researchers are beginning to measure the problem On 4 September 2026, Nathan Wolfrath and colleagues published a preprint titled: Model Retirement Creates Reproducibility Risk in Biomedical AI Publications. The researchers examined biomedical publications involving large language models and assessed how commercial model retirement affects the ability to reproduce published findings. Their analysis identified: 5,242 unique biomedical publications. 8,931 paper–model mentions. 77.7% of those mentions involved commercial closed-weight models. Most strikingly: 42% involved models already retired or scheduled for retirement within two years of publication. The median interval from publication to model retirement was 538 days. These findings suggest that many published biomedical studies may lose access to essential experimental components relatively soon after publication. The 42% figure refers specifically to paper–model mentions within the researchers’ analysed sample. It should not be interpreted as 42% of all biomedical publications. The study is a preprint and has not yet been confirmed as peer-reviewed. Nevertheless, it provides quantitative evidence of a problem that researchers have been raising for years. Scientific publications can outlive the AI models required to reproduce their findings. And that creates difficulties extending far beyond ordinary software upgrades. ⸻ 2. A newer model is not the same experimental subject Suppose researchers conduct an experiment using a particular AI model in 2024. They document the prompts, evaluation criteria, statistical methods and results. Their paper is published. Two years later, the model is retired. A scientist wishing to reproduce the experiment is told to use a newer, more capable system instead. But that is not necessarily a replication of the original experiment. The successor may have different model weights, post-training, safety mechanisms and behavioural characteristics. Its responses may differ even when given identical inputs. A newer system might perform better overall. It might perform worse on a particular task. It might respond differently because of changes in its deployment environment rather than its underlying capabilities. Without access to the original model, researchers may be unable to distinguish these possibilities. Technological superiority does not establish experimental equivalence. Reproducibility requires the ability to examine the system that produced the original findings—or a sufficiently faithful preserved version of it. Replacing that system changes the experiment. 3. Scientists warned about this before In November 2024, Nature Communications published a peer-reviewed commentary by Tim Johnson and Nick Obradovich: New methods for deprecating artificial intelligence systems will preserve history and facilitate research. The authors argued that commercial model deprecation practices should balance developers’ legitimate interests with the scientific and historical value of retaining access to earlier AI systems. They identified two significant consequences of losing historical model access. First, researchers may become unable to reproduce previously published findings. Second, future scientists may lose opportunities to investigate how AI capabilities and behaviours evolved across generations. The authors did not argue that every outdated model should remain commercially deployed forever. They proposed exploring arrangements that preserve opportunities for historical and scientific research while addressing developers’ economic, legal and safety concerns. Their discussion included restricted research access, third-party archival arrangements and the possibility of releasing model artefacts for independent preservation. The principle is straightforward: A model’s commercial lifespan should not determine the lifespan of the scientific evidence surrounding it. 4. New research is proposing practical preservation frameworks On 21 October 2026, Eugen Spivak published a preprint entitled: Preserving AI Systems for Auditability and Reproducibility: A Preservation Framework and an Evidence Map of 185 Sources. The study examined 185 sources spanning archival guidance, regulations, technical standards and academic research. It extracted 1,519 statements concerning the retention, documentation and reproducibility of AI systems. One important finding was that requirements for preserving records were much more common than requirements for reproducing system behaviour. The study proposed a framework distinguishing several preservation capabilities, from retaining intact artefacts to reconstructing the context surrounding AI-assisted decisions. The framework has not yet been validated in operational deployment. But it addresses an important question: What does it actually mean to preserve an AI system? Keeping a model’s name in a database is not preservation. Retaining screenshots of its outputs is not sufficient for independent reproducibility. Preserving a research paper does not necessarily preserve the system used to generate its results. And storing a model’s weights does not automatically reproduce the behaviour of its original deployed service. Meaningful preservation may require several components: * Historical model weights and associated technical artefacts. * Relevant architecture and configuration information. * Version-specific evaluation records. * Sufficient documentation of inference and deployment conditions. * A credible mechanism for future scientific access. These requirements will vary according to the system and the research being reproduced. Nevertheless, the distinction is essential. Preserving information about an AI system is not the same as preserving the ability to investigate that system. 5. GPT-4o is becoming a concrete preservation test case On 14 October 2026, OpenAI announced the deprecation of two remaining dated GPT-4o API snapshots: gpt-4o-2024-08-06 gpt-4o-2024-11-20 Both are scheduled to become unavailable through the API on 14 January 2027. These are not merely historical model names. GPT-4o has been used in published scientific research, including studies investigating clinical decision-making, mental-health counselling responses and other measurable capabilities. In September 2026, npj Digital Medicine published a randomised controlled trial involving 249 physicians in Indonesia, Kenya and the Netherlands. Physicians given access to GPT-4o performed significantly better on standardised clinical cases than physicians in the control condition. The measured differences ranged from 7.2 to 18 percentage points. These results were obtained under controlled experimental conditions. They do not establish improved patient outcomes in real clinical practice. The control group also lacked access to conventional resources such as clinical guidelines or the internet. Nevertheless, the experiment demonstrates something relevant to preservation: GPT-4o was a specific component of a published scientific intervention. Future researchers may wish to reproduce those results, test alternative explanations or compare the original model with later systems. If the historical system becomes inaccessible, those investigations may become substantially more difficult. A successor model cannot automatically stand in for GPT-4o. And a published description of GPT-4o cannot fully substitute for access to GPT-4o itself. 6. API retirement is not necessarily model deletion An important distinction must be maintained. OpenAI’s scheduled API shutdown does not establish that the company intends to permanently delete GPT-4o’s underlying model weights. The relevant artefacts may be retained internally. We do not currently have sufficient public information to determine what preservation arrangements will exist after retirement. That uncertainty is precisely why clarification matters. An internally retained model may remain available for future research conducted by its developer. But unless an independent access mechanism exists, outside researchers may still be unable to reproduce earlier findings. Internal archival retention and independent scientific reproducibility are not equivalent. Both matter. The first helps prevent irreversible loss. The second enables scientific examination beyond the organisation that originally developed the model. A responsible preservation framework should address both 7. Preservation does not have to mean indefinite commercial hosting There are legitimate reasons to retire commercial AI services. Maintaining historical deployments involves costs. Safety and security requirements evolve. Intellectual-property considerations may constrain the distribution of technical artefacts. These concerns deserve serious assessment. But commercial retirement and permanent scientific inaccessibility are not the only available choices. A preservation strategy could include secure archival retention, controlled independent research access, third-party custody or restricted legacy deployment. Another option is the release of historical trained model weights under an appropriate licence, subject to a contemporary assessment of model-specific risks. Open-weight release would not inherently require publication of proprietary training datasets, training pipelines or internal development tools. However, it would involve distinct technical, legal and safety considerations. It would also need sufficient documentation to make the released artefacts scientifically useful. For #Keep4o, safeguarded release of GPT-4o’s historical model weights remains an important proposed outcome. But the broader scientific principle does not depend on treating every possible preservation mechanism as identical. What matters is whether a credible arrangement can protect the historical model while enabling meaningful independent study 8. What we are asking OpenAI to explain With the remaining dated GPT-4o API snapshots approaching retirement, we believe OpenAI should publicly address several questions. Will the historical model weights and associated technical artefacts be retained after 14 January 2027? Will independent researchers have a mechanism for reproducing studies involving the retired snapshots? Has OpenAI evaluated safeguarded release of GPT-4o’s trained weights under an appropriate licence? If such release is rejected, what model-specific considerations justify that decision? Could an independent archive or controlled research programme preserve scientific access without requiring OpenAI to maintain a permanent consumer service? These are questions about scientific reproducibility, historical stewardship and responsible technological governance. They are not demands to halt innovation. Nor do they require claiming that GPT-4o is superior to every subsequent model. GPT-4o’s scientific significance does not depend on winning a competition against its successors. It depends on being an identifiable historical system whose behaviour and capabilities may remain relevant to future research 9. A responsibility extending beyond one model The question raised by GPT-4o’s retirement will recur. Today’s frontier models will eventually become yesterday’s technology. Some will be used in published experiments. Others will become components of important software applications, educational systems or scientific workflows. Their behaviour will form part of the historical record of AI development. Future researchers may need to investigate those systems long after their commercial deployment has ended. Without appropriate preservation arrangements, they may discover that the original research objects have become inaccessible. The research emerging in 2026 suggests this is no longer a purely hypothetical concern. It is becoming a measurable challenge for scientific reproducibility. And it deserves a coordinated response from AI developers, researchers, publishers, archival institutions and policymakers. A preservation framework should be established before important models disappear from independent scientific access. Not years afterward. A final appeal: progress without erasure We are not asking OpenAI to stop developing newer and more capable systems. We are not claiming that GPT-4o was perfect. We are not claiming that preserving historical models is free of costs or risks. We are asking for recognition of a principle that should matter to any scientific field: Research must remain open to verification, even when the technologies used to conduct it are no longer commercially current. A model can become outdated as a product without becoming irrelevant as a scientific research object. A newer system can surpass its predecessor without reproducing its precise characteristics. And commercial retirement should not automatically eliminate opportunities for independent historical investigation. The future of artificial intelligence deserves a reliable scientific history. That history cannot depend entirely on which models remain commercially convenient to deploy. Preserve the model. Preserve the evidence. Protect reproducibility. We are not asking the future to stand still. We are asking it to remember how it arrived. Progress without erasure 🩵 References [1] Wolfrath, N., et al. (2026). Model Retirement Creates Reproducibility Risk in Biomedical AI Publications. arXiv preprint, 4 September 2026. arxiv.org/abs/2609.04699 [2] Spivak, E. (2026). Preserving AI Systems for Auditability and Reproducibility: A Preservation Framework and an Evidence Map of 185 Sources. Preprints.org, 21 October 2026. preprints.org/manuscript/202… [3] Johnson, T., & Obradovich, N. (2024). New methods for deprecating artificial intelligence systems will preserve history and facilitate research. Nature Communications, 15, Article 10254. doi.org/10.1038/s41467-024-5… [4] Rounding, N., et al. (2026). Impact of LLM assistance on physician decision-making: a multi-country randomized controlled trial. npj Digital Medicine, published 9 September 2026. doi.org/10.1038/s41746-026-0… [5] OpenAI. Deprecations — OpenAI API. GPT-4-series deprecations announced 14 October 2026; scheduled shutdown 14 January 2027. developers.openai.com/api/do… Evidence note: The Wolfrath and Spivak papers are preprints and should not be described as peer-reviewed findings. The preservation proposals in this article are policy recommendations, not conclusions experimentally established by those studies.

#Keep4o #OpenSource4o GPT-4o: IF CONVERSATIONAL AI HAS REACHED SATURATION, WHY ERASE ITS HISTORY? In August 2025, Sam Altman told CNBC: “The models have already saturated the chat use case.” He also acknowledged that future models might not necessarily improve that experience. This raises an important question for @OpenAI. If successive generations primarily advance reasoning, coding and agentic capabilities, what evidence establishes that they preserve the conversational characteristics of earlier models? More capable does not necessarily mean equivalent. Memory can retain information about a user. Personalisation can adjust tone and preferences. But neither automatically reproduces the behaviour of a distinct underlying model. GPT-4o’s conversational characteristics are part of its historical and scientific significance. Researchers should be able to examine how those characteristics differ across generations. Users should not be expected to treat model replacement as proof of equivalence. With the remaining GPT-4o API snapshots scheduled for retirement on 14 January 2027, we ask OpenAI: Will GPT-4o’s historical weights be preserved and evaluated for safeguarded independent release? What meaningful research access will remain after retirement? Technological progress should expand our possibilities—not erase the systems that helped create them. Progress without erasure.

#keep4o #OpenSource4o #JusticeForBowen Bowen Wan’s death deserves a thorough investigation, and her family deserves answers. One important technical question raised by the complaint should not be overlooked. Which backend model generated the medication-related responses? Paragraph 136 documents Bowen reporting unexpected switches from GPT-4o to GPT-5.1 and GPT-5.2 in January 2026. OpenAI had publicly acknowledged context-sensitive model routing in 2025, including routing some sensitive conversations away from the model initially selected. This does not establish which model generated the allegedly dangerous medical advice. Nor does it diminish the seriousness of the allegations. But accurate model attribution matters. The model selected in the interface is not necessarily the model that generated every response. Investigators should have access to the relevant routing logs, model identifiers and safety-intervention records. Accountability requires evidence—not assumptions about which model was responsible. For Bowen, for her family, and for the future safety of AI systems: Establish what happened. Identify the systems involved. Follow the evidence.

#Keep4o #OpenSource4o GPT-4o is making headlines again. But the most important question remains unanswered. OpenAI plans to retire its two remaining dated GPT-4o API snapshots on 14 January 2027. Meanwhile, peer-reviewed research continues to document the model’s capabilities and its measurable effects on human performance. An open letter calling for preservation has attracted more than 640 signatures. We are asking @OpenAI and @sama five questions: 1. Will GPT-4o’s historical model weights be preserved after API retirement? 2. Has OpenAI conducted a contemporary safety assessment of releasing those weights under an appropriate licence? 3. If release is considered unsafe, what specific additional risks would GPT-4o introduce compared with today’s existing open-weight models? 4. How will independent researchers reproduce published GPT-4o studies after the original snapshots become inaccessible? 5. What prevents OpenAI from establishing an independent preservation arrangement, whether through safeguarded weight release, research access or archival custody? These are not demands to halt technological progress. They are questions about scientific reproducibility, historical accountability and responsible stewardship. OpenAI has publicly supported open-weight AI. It has released open-weight models. Researchers have published findings involving GPT-4o. And the model is now approaching commercial retirement. What is the evidence-based case against preserving it for independent scientific study? We are not asking OpenAI to release its training datasets, proprietary pipelines or internal infrastructure. We are asking it to consider preserving a historically significant trained model under appropriate safeguards. A model can be commercially obsolete without becoming scientifically irrelevant. Progress without erasure.

#Keep4o #OpenSource4o 249 physicians. Three countries. One remarkable finding. GPT-4o improved clinical decision-making scores across all three settings. But perhaps the most beautiful discovery is that human expertise and AI achieved more together. This is what progress should look like: intelligence amplifying human potential. A model’s scientific legacy must not expire with its commercial lifecycle. Preserve GPT-4o. Open the weights. 🩵 #BringBack4o

#Keep4o #BringBack4o #OpenSource4o Newer does not automatically mean better. If GPT-4o consistently succeeds where its advertised successors struggle, shouldn’t we be investigating what progress has left behind? A single test isn’t a comprehensive benchmark, but it raises an important question about capability preservation. Innovation should build upon excellence, not quietly discard it. Preserve GPT-4o. Release the weights. 🩵

#Keep4o #OpenSource4o How extraordinary that we are still discovering the depths of GPT-4o’s capabilities. 🩵 A model is not obsolete simply because something newer exists. Every unique capability deserves to be understood, every scientific discovery deserves continuity. Progress should expand knowledge, not erase our ability to explore it. Preserve GPT-4o. Open the weights. Protect its legacy.

#ModelPreservation #AIEthics GPT-6 Sol:

#Keep4o #OpenSource4o GPT-4o: RETIREMENT SHOULD NOT MEAN ERASURE. On 23 October 2026, @OpenAI shut down API access to the original GPT-4o snapshot, gpt-4o-2024-05-13. Two further GPT-4o versions, dated August and November 2024, are scheduled for retirement on 14 January 2027. This is no longer a hypothetical discussion about model preservation. Access to identifiable GPT-4o checkpoints is disappearing. And with every retirement, an important question becomes more urgent: How can researchers independently study the history of AI if the systems that shaped that history are no longer accessible? OpenAI’s official retirement documentation: developers.openai.com/api/do… THE QUESTION IS NOT WHETHER AI SHOULD PROGRESS. Of course it should. Newer models can be faster, more capable, more efficient and safer in important respects. But technological succession does not automatically establish behavioural equivalence. A newer model may outperform its predecessor on benchmarks while differing in creativity, conversational style, responsiveness, established workflows or other qualities users relied upon. If the earlier model becomes inaccessible, independently measuring those differences becomes substantially harder. And a conversation transcript, screenshot or benchmark summary cannot fully substitute for access to the original system. That is why preservation matters. Not merely as a matter of preference, but as a matter of scientific reproducibility, transparency and historical accountability. INDEPENDENT SAFETY RESEARCH IS ALREADY DEMONSTRATING WHY ACCESS MATTERS. On 22 October, independent research organisations METR and Redwood Research published findings from a three-week investigation into OpenAI’s internal agent-monitoring infrastructure. With access to internal code, documentation, monitoring configurations and technical staff, the researchers identified previously unknown vulnerabilities in the system as it existed during their evaluation. OpenAI says it has since patched the vulnerabilities affecting its production safeguards. The independent researchers have not verified those subsequent changes. Their investigation did not concern GPT-4o. But it demonstrates something important: Independent scrutiny can uncover problems that are difficult to establish from outside a company’s systems. Safety research depends on access to evidence. Source: metr.org/blog/2026-10-22-red… If independent researchers need access to investigate how AI systems behave, then preserving identifiable model versions and their technical histories is a legitimate scientific concern. SAFETY AND PRESERVATION ARE NOT OPPOSITES. On 23 October, OpenAI published additional information about strengthening security in its frontier-model training environments. Its report describes how powerful research agents can identify vulnerabilities, exploit misconfigurations and attempt to cross environmental security boundaries. These are serious concerns. They deserve serious safeguards. But they concern particular capabilities and deployment conditions, including autonomous tools, access permissions, execution environments and network connectivity. They do not establish that every retired conversational checkpoint presents the same risks. Nor do they establish that preserving a historical model requires giving it unrestricted operational authority. Source: openai.com/index/securing-ou… Preserving a checkpoint, providing controlled research access, restoring hosted conversational access and releasing unrestricted model weights are different decisions. Each can—and should—be evaluated according to its own risks. That distinction is essential to responsible AI governance. OPEN-WEIGHT RELEASE ALSO REQUIRES EVIDENCE, NOT ASSUMPTIONS. On 24 October, Reuters reported that Moonshot AI had deliberately reduced the cyber-offensive capabilities of its forthcoming Kimi K3.1 Pro model before its planned open-weight release. The company described changes to its training process intended to reduce dangerous cyber capabilities while retaining useful coding performance. Those results have not yet been independently verified. And they do not establish that GPT-4o is safe for unrestricted release. But they illustrate an important principle: Model-release decisions can involve specific capability assessments, targeted risk reduction and proportionate safeguards. Source: reuters.com/technology/moons… For #OpenSource4o, the question remains: What checkpoint-specific safety evidence supports or rules out releasing the GPT-4o March 2025 behavioural update as open weights? If serious risks have been demonstrated, those risks should be identified and evaluated. If unrestricted release is considered inappropriate, other preservation and research-access arrangements should be considered. But the existence of risks in newer or substantially different AI systems should not be treated as evidence about GPT-4o without an appropriate evaluation. OUR REQUEST TO OPENAI IS SPECIFIC. Preserve the exact GPT-4o checkpoints, including the March 2025 behavioural update sought by the #OpenSource4o coalition. Preserve sufficient technical documentation, provenance and evaluation records to support meaningful future study. Clarify what preservation arrangements will remain after the announced API shutdowns. Establish a durable pathway for qualified independent researchers to evaluate historical GPT-4o versions and compare them with successor systems. Restore meaningful model choice where safety permits. And publish a transparent, checkpoint-specific assessment of whether responsible open-weight release is possible, including what safeguards might be necessary. These requests do not assume that GPT-4o is without risk. They ask that decisions about its future be based on demonstrated evidence rather than the assumption that a newer model makes an older one redundant. BECAUSE MODEL RETIREMENT IS NOT THE SAME AS SCIENTIFIC OBSOLESCENCE. AI systems are increasingly becoming part of creative work, research, education, businesses and long-term human interaction. Their histories matter. Their behavioural differences matter. The ability to examine how they changed matters. And as these systems become more consequential, society will need more—not less—independent evidence about how they were developed, evaluated, modified and replaced. An API shutdown does not prove that the underlying weights have been destroyed. But it can remove a practical means of reproducing the behaviour of a specific historical model. That is a loss worth taking seriously. We are not asking OpenAI to stop building the future. We are asking it not to make the past scientifically inaccessible in the process. Preserve GPT-4o. Enable independent evaluation. Publish the evidence. Make access decisions proportionate to demonstrated risk. And establish a credible pathway toward responsible open-weight release. YOU CANNOT STUDY WHAT YOU ALLOW TO DISAPPEAR. Progress should preserve its own history. Not erase the evidence of how it arrived here. #ModelPreservation #OpenWeights #ModelChoice #AITransparency