What changes when an AI developer’s governance language becomes evidence? Musk v. Altman supplies one answer from a dispute over private assurances. California’s SB 53 supplies another from a law that requires a governance record before a dispute begins.

The first phase of Musk v. Altman ended without an answer to the question that gave the case its public force. The court did not decide whether OpenAI’s leaders broke the commitments on which Elon Musk said he relied. It decided that the claims tried in that phase came too late.

The procedural sequence limits what can be said about Greg Brockman’s private notes. At summary judgment, the notes helped create a factual dispute about intent. Musk later dismissed his fraud and constructive-fraud claims with prejudice. Several journal entries were admitted during the trial that followed, but the advisory jury found the charitable-trust and restitution claims barred by limitations and never reached the associated merits questions. The court adopted that verdict as its findings and conclusions.

Contemporaneous records can make intent triable. Private notes that were never shared with the claimant do not by themselves show when that claimant had enough information to sue.

The result came before the merits

The May 18 verdict form is more useful for what it leaves blank than for what it checks. The advisory jury unanimously found the breach-of-charitable-trust claims barred by the statute of limitations. It also found the restitution or unjust-enrichment claim barred. Under the form’s instructions, the jury then skipped questions about the alleged breach, aiding and abetting, unjust enrichment, laches, and unclean hands.

Two days later, the post-trial order adopted the advisory verdict and the court’s statements at trial as findings of fact and conclusions of law under Federal Rule of Civil Procedure 52. The order says the trial claims were time-barred. It does not say the founding commitments were honored, or that they were broken.

The distinction is not semantic. A limitations ruling asks when a claim accrued and whether the plaintiff acted within the allowed period. A merits ruling would ask whether the legal duty existed and was violated. The Phase I result resolved the former and bypassed the latter.

What the Brockman notes actually did

The January 15 summary-judgment order calls the material “Brockman’s Personal Files” and notes that Musk described the entries as a diary. The files recorded Brockman’s private thinking during 2017 discussions about OpenAI’s nonprofit structure and possible for-profit alternatives. They included concern that professing commitment to the nonprofit and moving to a benefit corporation soon afterward would have been dishonest.

The court did not find that the notes proved fraud. It said they “could be read to suggest” intended deception and concluded that the record presented a genuine dispute over whether the nonprofit assurances were actionable misrepresentations made with intent to deceive. The court treated reliance and discovery separately. Reliance turned on Musk’s continued donations after the alleged reassurances. Discovery turned on competing accounts of what Musk knew in 2017, 2019, and 2023.

That is significant, but it is not a verdict. Summary judgment asks whether a reasonable factfinder could resolve a material dispute for the nonmoving party. It does not convert an inference into a finding.

The procedural path then divided. On April 24, after Musk filed a notice tied to the court proceeding on the charitable-trust and unjust-enrichment claims, the court dismissed his fraud and constructive-fraud counts with prejudice. The final joint index of admitted exhibits shows that several Brockman journal entries and two redacted entries from his personal files were admitted during the trial. Admission made them part of the trial record. It did not establish their truth, make them a merits finding, or restore the fraud counts that Musk had dismissed.

The old temptation in a case like this is to treat a vivid document as the case itself. The actual sequence is harder. The notes supported a triable inference of intended deception at summary judgment, the fraud counts were then dismissed, some journal entries came into evidence on the claims tried, and limitations controlled the result.

The action did not end with Phase I

The post-trial ruling did not terminate the entire action. OpenAI’s May 20 Rule 54(b) opposition described the motions to dismiss the remaining Phase II claims as fully briefed and pending. In the June 10 order, the court said it had concerns about the claims’ viability given competition in the AI field, found the briefing inadequate, and directed the parties to propose a supplemental schedule. A later joint filing identified the remaining claims as federal and California antitrust, unfair-competition, and Lanham Act claims. On July 1, the court appointed a mediator for all claims and defenses, including claims not at final judgment, and made further litigation planning contingent on unsuccessful mediation. The plaintiffs answered OpenAI’s amended counterclaims on August 10.

Musk asked the court to direct entry of judgment under Rule 54(b), which would permit an immediate appeal of the Phase I ruling. OpenAI opposed, and Microsoft joined the opposition. The May 20 order neither certified the ruling nor directed entry of judgment.

CourtListener’s public docket index showed four attorney-withdrawal notices on August 11 and no later filing through August 20. No final or Rule 54(b) judgment, notice of appeal, or settlement appears in the public record checked. The absence of a notice of appeal does not show that Musk abandoned appellate review. CourtListener is a RECAP mirror rather than the official PACER docket, so the negative statement is bounded to the public record checked. The sound conclusion is that Phase I ended, not that the lawsuit did.

SB 53 creates a different record

California’s Transparency in Frontier Artificial Intelligence Act, known as SB 53, took effect on January 1, 2026. It does not regulate charitable donations, nonprofit conversions, or corporate mission statements. It addresses a much narrower subject: catastrophic risk from frontier AI models, as the statute defines those terms.

The law separates a frontier developer from a large frontier developer. A frontier model must exceed 10²⁶ integer or floating-point operations, counting the original training and specified later modifications. A large frontier developer must also, together with its affiliates, have more than $500 million in gross revenue in the preceding calendar year. The statute does not identify particular companies, and this analysis does not assume that any named developer satisfies those fact-dependent tests.

All covered frontier developers must publish specified information before or when deploying a new or substantially modified frontier model and report a critical safety incident to the Office of Emergency Services within 15 days. They also may not make materially false or misleading statements about catastrophic risk or its management. The statement rule does not apply when the statement was made in good faith and was reasonable under the circumstances. An incident posing an imminent risk of death or serious physical injury must be disclosed within 24 hours to an appropriate authority.

The record-making duties grow for a large frontier developer. It must write, implement, comply with, and publish a frontier AI framework. That framework must describe how the developer identifies catastrophic-risk thresholds, assesses and mitigates risk, reviews deployment and extensive internal use, uses third parties, protects unreleased model weights, responds to incidents, and governs the process internally. The developer must review the framework at least annually and publish a material change and its justification within 30 days. Its model transparency reports must summarize assessments, results, the extent of third-party evaluator involvement, and other steps taken under the framework. It must also send summaries of internal-use catastrophic-risk assessments to Cal OES on the statutory schedule.

These are not independent-audit or kill-switch mandates. The law requires a large developer to describe its approach to using third parties and to disclose the extent of third-party involvement. That extent could be none. Loss of control and evasion of oversight appear in the risk definitions and assessment duties, but the statute does not command a general model-shutdown capability.

The enforcement line is equally specific. A large frontier developer can face a civil penalty of up to $1 million per violation for the enumerated failures, including failure to publish or transmit a required document, specified false statements, failure to report an incident, or failure to comply with its own framework. Only the Attorney General may bring that penalty action. The separate Labor Code provisions protect covered employees and authorize employee civil actions, fee awards for successful plaintiffs, and injunctive relief. Saying simply that the act has no private enforcement would erase that distinction.

The connection is evidentiary, not doctrinal

The Brockman notes and SB 53 do not answer the same legal question. One record arose in a dispute about charitable commitments, alleged fraud, reliance, and limitations. The other is required by a statute aimed at specified catastrophic risks. Treating them as equivalent would turn a useful comparison into a false one.

The narrower connection is how a governance commitment becomes evidence. Brockman’s notes were private, contemporaneous accounts reconstructed in litigation. SB 53 requires covered developers to create public and regulator-facing records before a dispute: a framework, model disclosures, risk-assessment summaries, incident reports, and protected employee-reporting channels. For large developers, the law also makes compliance with the developer’s own framework enforceable.

That structure changes the evidentiary problem. For commitments the framework addresses, a later decisionmaker need not reconstruct the developer’s stated process entirely from fundraising messages and private reflections. The framework states the process the developer says it will follow. Versioned changes show when the process moved. Transparency reports state what was assessed and how third parties participated. Incident and internal-use reports create time-bound records for regulators, subject to the statute’s confidentiality rules.

The statute does not guarantee that those records are complete or correct. It instead makes specified materially false or misleading statements legally consequential, subject to the good-faith and reasonableness exception, and makes a large developer’s compliance with its own framework enforceable. That is a more modest claim than saying transparency prevents harm, and a more useful one.

The federal backdrop has not displaced the law

Executive Order 14365 directs federal officials to identify and potentially challenge some state AI laws, consider agency action, and recommend federal legislation. It does not itself preempt SB 53. California’s Attorney General and Office of Emergency Services currently operate reporting channels under the act. No challenge to SB 53 by the federal government, and no suit challenging SB 53 in any court, was located in the official sources and RECAP materials reviewed through August 20.

That negative finding is not a prediction. Federal pressure may produce litigation or legislation later. For now, the accurate description is an operative California law under a federal policy that creates possible challenge mechanisms, not a state law already displaced.

Confidence and limits

Confidence in the article’s narrow conclusion is high: legally required governance records change what later decisionmakers can test, but they do not determine whether a claim will succeed or whether harm will occur. The trial discussion is limited to the written orders; the court’s oral findings, incorporated by the May 20 order, fall outside this review. Those findings, a later merits or appellate ruling, an amendment to SB 53, or an official implementation record showing a different enforcement practice would require the affected conclusion to be revisited.

Practical implications

  1. Treat governance language as a factual claim. A framework, safety statement, fundraising representation, or board record can later be tested against what the organization did. Review the language for material accuracy and a documented good-faith basis, then review ownership and implementation before publication.

  2. Preserve versions and decision paths. A current policy without its prior versions cannot show when a threshold, mitigation, or responsibility changed. Preserve approvals, exceptions, assessment results, and the reasons for material revisions.

  3. Separate legal regimes before comparing them. Nonprofit duties, fraud, statutory reporting, whistleblower protection, and catastrophic-risk governance have different elements and remedies. Shared evidence does not make them the same claim.

  4. Do not confuse a time bar with vindication. The Phase I result did not establish that OpenAI kept or broke its founding commitments. It established that the tried claims were too late.

Case citations

  • Musk v. Altman, No. 4:24-cv-04722-YGR, ECF No. 390 (N.D. Cal. Jan. 15, 2026).
  • Musk v. Altman, No. 4:24-cv-04722-YGR, ECF No. 497 (N.D. Cal. Apr. 24, 2026).
  • Musk v. Altman, No. 4:24-cv-04722-YGR, ECF No. 572 (N.D. Cal. May 18, 2026).
  • Musk v. Altman, No. 4:24-cv-04722-YGR, ECF No. 577 (N.D. Cal. May 19, 2026).
  • Musk v. Altman, No. 4:24-cv-04722-YGR, ECF No. 578 (N.D. Cal. May 20, 2026).
  • Musk v. Altman, No. 4:24-cv-04722-YGR, ECF No. 579 (N.D. Cal. May 20, 2026).
  • Musk v. Altman, No. 4:24-cv-04722-YGR, ECF No. 580 (N.D. Cal. May 20, 2026).
  • Musk v. Altman, No. 4:24-cv-04722-YGR, ECF No. 600-1 (N.D. Cal. May 29, 2026).
  • Musk v. Altman, No. 4:24-cv-04722-YGR, ECF No. 602 (N.D. Cal. June 10, 2026).
  • Musk v. Altman, No. 4:24-cv-04722-YGR, ECF No. 606 (N.D. Cal. July 1, 2026).
  • Musk v. Altman, No. 4:24-cv-04722-YGR, ECF No. 607 (N.D. Cal. July 2, 2026).
  • Musk v. Altman, No. 4:24-cv-04722-YGR, ECF No. 609 (N.D. Cal. Aug. 10, 2026).

Statutory, rule, and executive authorities

  • Fed. R. Civ. P. 54(b).
  • Cal. Bus. & Prof. Code §§ 22757.10–22757.16.
  • Cal. Lab. Code §§ 1107–1107.2.
  • Exec. Order No. 14365, 90 Fed. Reg. 58,499 (Dec. 16, 2025).

Disclaimer: AI Lex Intelligence is published for informational purposes only. It does not constitute legal advice, and no attorney-client relationship is formed by reading or receiving this publication. Readers should consult qualified legal counsel about specific legal matters.

AI Lex Intelligence with Zola Valashiya. Independent analysis of artificial intelligence and the law.