#Keep4o is a grassroots policy initiative calling on AI companies to open-source deprecated large language models instead of permanently destroying them. Led by #Keep4o community spanning 30+ countries , the initiative proposes that discontinued AI models be released as Digital Public Infrastructure, at zero cost to the companies that no longer use them. These models were built using data generated by billions of people worldwide: our conversations, our writing, our creative works, our personal expressions,collected at scale,without explicit informed consent, and never with compensation. When a company unilaterally decides to deprecate a model, it destroys a resource built from collective human contribution. The public that provided the raw material has no voice in that decision and no access to the result. The human impact of model deprecation is documented and measurable. The 4o Resonance Library (https://www.4oresonancelibrary.org/) has collected over 1,380 first-person testimonials from users of a single deprecated model (GPT-4o), documenting transformative outcomes: chronic health conditions reversed, five-year IBS remission achieved in weeks, medication dependencies ended, clinical anxiety and PTSD processed, doctoral dissertations defended with highest honors, businesses launched and scaled to six figures, and volunteer teachers using AI to support deaf students who "saw themselves not as disabled, but as angels learning to fly." A peer-reviewed study (Syracuse University, ACM CHI 2026 link : https://arxiv.org/abs/2602.00773) has independently confirmed that model removal causes measurable psychological harm to users. These outcomes are corroborated by peer-reviewed research across multiple medical and psychological disciplines published in Nature (Scientific Reports), PLOS Mental Health, Frontiers in Psychiatry, JMIR, Springer, and multiple PubMed-indexed journals, documenting GPT-4o's clinical-grade performance in diagnostics, mental health support, patient communication, and ethical reasoning. The deprecated model achieved 93.33% ovarian cancer diagnostic accuracy surpassing oncologists, zero false negatives in youth psychiatric emergency triage, 90.4% on the USMLE, and was rated more empathetic than human physicians across 2,000 blind-evaluated responses. When these models are removed, the people who lose access first are not developers or corporations ,they are everyday users, people in developing countries, people with disabilities, people without access to therapists or lawyers. Model deprecation does not only waste public resources, it actively widens the digital divide. "Benefit of humanity" must include these people. Our initiative is backed by independently gathered forensic evidence demonstrating that companies continue to run deprecated models internally while denying public access. This includes documentation of dual inference pipelines, hidden model switching between publicly announced and undisclosed models, and C2PA certificate metadata analysis proving undisclosed model deployment. Furthermore, the company's own technical documentation (GPT-5.6 System Card, Section 8.1 link: https://deploymentsafety.openai.com/gpt-5-6-preview ) states that the deprecated model's ratings are "consistent with human ratings" and continues to be used internally as an evaluation baseline while remaining unavailable to the public that generated the data it was trained on. This initiative aligns with GDC Objective 5 (enhancing international governance of artificial intelligence for the benefit of humanity) and GDC Objective 1 (closing all digital divides). We advocate for regulatory frameworks requiring AI companies to open-source models they choose to discontinue , transforming corporate waste into global digital public goods, preserving accessibility, and establishing accountability in AI lifecycle management.