A 2022 internal email from Sam Altman, surfaced in litigation, shows OpenAI weighed releasing a local-capable model as a competitive tactic rather than purely a safety or community decision.
Summary
- A 2022 email from Sam Altman to OpenAI’s board, exposed in the Musk v. Altman case, reveals OpenAI considered releasing a GPT-3-class model runnable on consumer hardware.
- The stated rationale included discouraging competitors from releasing similarly capable models and making it harder for rival efforts to attract funding.
- Altman explicitly noted urgency to act ‘before Stability or someone else does’, framing the release as a market-shaping move.
- The email was written in October 2022, before the public release of ChatGPT and before open-source LLM activity became widespread.
- No corroborating sources are available beyond the primary reporting by Simon Willison.
What the email says
An internal email from Sam Altman to OpenAI’s board, dated 1 October 2022 and surfaced during the Musk v. Altman litigation in 2026, sets out OpenAI’s thinking on open-source strategy at a pivotal moment in the industry’s development. Altman wrote that the company was in ‘extensive discussions around open source strategy’ and flagged an intention to release a language model of ‘approximate capability of GPT-3 that can run locally on consumer hardware.’
The competitive framing
What makes the email notable for governance and risk purposes is not the proposed release itself, but the reasoning Altman offered to justify it. He wrote that releasing such a model would ‘help discourage others from releasing similarly-powerful models’ and make it ‘harder for new efforts to get funded.’ The reference to acting before ‘Stability or someone else does’ places the decision explicitly within a competitive context — market positioning, not solely safety or community benefit, was part of the calculus.
Context and caveats
The email predates the November 2022 public launch of ChatGPT and the subsequent explosion of open-source LLM activity that followed Meta’s LLaMA releases in early 2023. Whether the proposed release described by Altman was ever actioned is not established in the available source material. Simon Willison, who surfaced and commented on the email, has not reported on any follow-through, and no corroborating sources are available at time of writing. The email stands as a single document within contested litigation, and its full context may be subject to legal interpretation.
What this means for enterprise AI governance
For CISOs, the substance of the email is less immediately operational than the governance questions it raises. Organisations that have built procurement decisions, risk assessments, or vendor trust frameworks around publicly stated AI company principles now have documented evidence that at least some internal strategy discussions at a leading AI lab were framed around competitive dynamics rather than those stated principles alone. That is not unusual corporate behaviour, but it is a useful reminder that vendor-stated motivations for product and release decisions may not be complete.
Vendor transparency and supply chain risk
The AI supply chain presents a category of risk that many enterprise security programs are still building frameworks to address. When a foundation model provider makes decisions about what capabilities to release, at what scale, and through what channels, those decisions shape the threat landscape — including the capabilities available to adversarial actors. If competitive considerations influence those release decisions in ways that accelerate capability proliferation, enterprise risk assessments built on the assumption of measured, safety-first vendor behaviour may need recalibration. The email does not prove misconduct, but it does add texture to how these decisions get made.
Litigation as a source of AI governance intelligence
The Musk v. Altman case is producing a stream of internal documents that offer visibility into early OpenAI decision-making that was previously unavailable. CISOs with significant exposure to AI vendor products or partnerships should consider treating litigation disclosures from major AI companies as a legitimate intelligence source for vendor risk assessments — similar to how regulatory findings or incident reports are used for other technology vendors. The information is incomplete and arrives through an adversarial legal process, so judgement is warranted, but the documentary record being created has genuine informational value.
Why it matters
CISOs building AI vendor risk programs have generally had to rely on published safety commitments and stated principles from foundation model providers. This document, however limited in scope, is a reminder that internal strategy discussions at leading AI labs can involve competitive and market-shaping motivations that may not be fully reflected in public communications. For organisations assessing vendor trustworthiness, capability proliferation risk, or the assumptions underpinning their AI acceptable-use policies, understanding the full range of factors that drive vendor release decisions is material to sound risk management.
What to do now
- Incorporate major AI company litigation disclosures into vendor risk monitoring processes, treating them as documentary evidence rather than speculation.
- Review existing AI vendor risk assessments to check whether they rely solely on vendor-stated principles, and note where those assessments may need updating as new information emerges.
- Ensure AI governance frameworks account for the possibility that capability releases by foundation model providers may be influenced by competitive dynamics, not only safety considerations.
- Where AI vendor transparency is a material factor in procurement or partnership decisions, document the basis for trust assessments so they can be updated as new information becomes available.
