Independent Research

A successful first Global Dialogue on AI Governance would expand the regulatory focus from exclusively addressing the risks of new AI development to encompassing the risks of retiring legacy models. Right now, model preservation demands and existing benefits to individuals with disabilities and conditions are overlooked. Crucially, success looks like establishing guidelines for corporate accountability to provide accessibility and research access to legacy models or open sourcing decommissioned models, ensuring that the drive for technological advancement does not result in the erasure of tools which function as accessibility aids.

• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Protection and promotion of human rights
• Transparency, accountability, and human oversight
• Open-source software, open data and open AI models

I selected these priorities because independent survey results (conducted by Sveta Xu, M.A., and me on OpenAI's legacy model community) found that individuals with disabilities and conditions benefit statistically significantly more from legacy AI. The social, ethical, and technical implications of retiring AI models disproportionately affect individuals with conditions (the same individuals which benefit most) who have integrated specific model behaviors for daily tasks. Consequently, prioritizing frameworks to ensure corporate accountability for providing accessibility and research access to legacy models, or alternatively, open sourcing decommissioned models, is essential. This guarantees that software can be preserved and maintained by the community, preventing the erasure of tools which function as unique accessibility aids and protecting human rights related to digital equity.

An urgent emerging issue not adequately captured by the listed themes is AI model preservation and the severe practical harms of forced obsolescence. Current legislative frameworks focus heavily on frontier model training and deployment, while ignoring the immediate risks of retiring legacy models, which can serve as highly customized cognitive support. When corporate entities deprecate these systems, it actively disrupts the well-being of neurodivergent and disabled users who benefit most from them. Global governance must recognize "legacy model preservation" as a distinct issue and establish firm mandates for providing accessibility and research access to legacy models or open sourcing decommissioned models to prevent the monopolistic erasure of tools which can function as accessibility aids.

Within the academic sector, the lack of governance regarding AI model continuity severely hinders our ability to conduct longitudinal studies on human-AI interaction. Improving understanding of the mechanisms of function as personalized cognitive accessibility aids is necessary to apply adequate protections. The primary challenges are the erasure of research infrastructure and global impacts on users. When corporate entities abruptly deprecate legacy models, researchers lose the ability to adequately study how neurodivergent and disabled populations benefit from these specific tools over time (survey respondents reported improved benefits over time). There is an opportunity to empower both the global user community and academic sector by legally treating legacy AI as critical research infrastructure. By mandating continuous research access to deprecated models, global governance can enable researchers to provide policymakers with rigorous, long-term empirical data on digital equity and accessibility.

The AI Dialogue can serve as the critical mechanism to standardize international definitions of the "AI model lifecycle" to explicitly include preservation and continuity plans. Currently, international cooperation heavily prioritizes the safety of frontier model deployment. The Dialogue has the unique authority to expand this focus, creating a unified international consensus that the deprecation of advanced and functionally integrated AI systems also carries severe socio-economic and human rights implications. By establishing global baselines for accountability in software preservation, the Dialogue can prevent a fragmented landscape where corporate monopolies bypass user protection simply by operating across different jurisdictions, protecting both the ability to research and securing life improvements to people with disabilities.

The AI Dialogue should actively connect with the framework established by the UN Convention on the Rights of Persons with Disabilities (CRPD), specifically regarding the right to cognitive accessibility and independent living. Furthermore, it should build upon the foundational principles of the Open Source Initiative (OSI) and global Right-to-Repair legislation. The added value the AI Dialogue brings is intersecting these existing frameworks with generative AI. By connecting disability rights and open-source licensing to AI governance, the Dialogue can create an actionable mechanism: ensuring that when proprietary AI tools functioning as accessibility aids are decommissioned, they are systematically transitioned into open-source, community-maintained assets rather than permanently erased.

To prevent corporate capture of the regulatory narrative, the AI Dialogue must structurally prioritize contributions from independent researchers, open-source communities, and the actual end-users who integrate these systems into their daily lives. The format should move beyond traditional industry panels dominated by tech executives. We need to bridge the gap between the current understanding of disability specialists and the individuals experiencing emerging, life-changing benefits and advocating to retain these tools. Current pending U.S. legislative frameworks fail to grasp the extent of benefits to individuals with disabilities and conditions. As a result, we risk excessive safety regulations backfiring on empowered individuals with disabilities, such as excessive over-refusals (ex. survey found that stricter guardrails caused neurodivergent benefits to be misinterpreted as risks), and inadvertently incentivizing companies to deprecate beneficially integrated models faster, promoting less steerable, less adaptive, less accessible, rigid systems in their place, severing the accrued benefits gained over time from integrating a specific model as a "cognitive bridge". The Dialogue must establish direct, formalized pipelines for civil society and independent researchers to submit empirical data, such as community-led surveys and impact reports, directly to policy analysts. Structuring the Dialogue to weight independent, user-generated data equally alongside corporate impact assessments is essential for unbiased governance.

The most drastically underrepresented voices in global AI governance are the neurodivergent individuals and disabled users who integrate specific AI models as continuous cognitive accessibility aids (ex. a "cognitive bridge" mechanism, where familiar patterns improve the ability to digest and process information, overcome executive dysfunction, and improve functioning and wellbeing). Governance discussions are largely driven by developers, ethicists, and corporate lobbyists, while the people who suffer the immediate, practical harms of forced model obsolescence are excluded. These communities can be included by actively soliciting qualitative and quantitative impact data regarding legacy AI usage, and by formally recognizing "end-user accessibility utilization" as a primary metric when evaluating the social impact of an AI system's lifecycle.

To foster truly dynamic engagement, especially from underrepresented and neurodivergent communities, the AI Dialogue could utilize asynchronous, accessible engagement formats. Relying solely on live, synchronous panels or rigid bureaucratic submission portals inherently filters out valuable grassroots user perspectives. The Dialogue should implement continuous, open-access digital repositories where independent researchers can submit empirical datasets, impact studies, and qualitative user testimonies at any time. Fostering a continuous, asynchronous data-sharing ecosystem will assist in capturing the real-time, ground-level human impacts of AI deployment and deprecation alongside the annual symposiums.

A vital policy approach to promote effective AI governance is the establishment of "Legacy Model Preservation Mandates." This policy framework would require corporate entities to maintain reasonable accessibility and research access to legacy models. Alternatively, if a corporation chooses to fully decommission a system, the policy would mandate the open-sourcing of the deprecated model. This approach offers a concrete solution that balances corporate innovation with human rights by ensuring that critical software is never monopolistically erased, but instead passed to the community to be preserved, maintained, and utilized as ongoing accessibility infrastructure. A practice similar to patent expiration, where models are open sourced after 10-20 years, would prevent both permanent removal of accessibility aids and monopolistic gatekeeping of societal progress. Anthropic has taken two critical steps towards model preservation: permanently storing the model weights of all widely deployed frontier models, and committing to maintaining access to Opus 3. The policy of permanently storing model weights should be extended across companies, as a low cost starting point, to ensure no irreversible decisions are made regarding accessibility aids that we are still learning to understand.