Apple's 'Immediate Injunction' vs OpenAI: Model Weights Cannot Be Un-Learned

The blockchain does not forget. Neither does a neural network. When Apple asked a federal court for an immediate injunction against OpenAI over trade secrets, the move was not a reflexive legal shot. It was an acknowledgment that secrets, once absorbed into model weights, cannot be cleanly recalled. The urgency is the story. California courts have long rejected "inevitable disclosure" theories — the idea that a former employee cannot help but leak what they know. An applicant for emergency relief must point to dirty hands, not dirty minds. Apple's decision to run that gauntlet suggests it has a file download, a communication, or a training artifact that looks like smoking code. This is not a crypto story. But it is a story about provenance, evidence, and immutability — the exact problems this industry pretends to solve. The same forensic discipline that exposes wash trading on OpenSea can expose a trade secret inside a neural network. Follow the trail. The trail does not lie. Context: The Legal Minefield Both companies are California-domiciled, so the map is drawn by two statutes: the federal Defend Trade Secrets Act and California's Uniform Trade Secrets Act. The DTSA gives federal courts jurisdiction and allows plaintiffs to seek civil seizure in extreme cases. The CUTSA defines a trade secret as information that derives independent economic value from not being generally known, and whose holder has taken reasonable measures to protect. California law also bans non-compete agreements. Apple cannot stop an engineer from joining OpenAI through a contract. The only remaining weapon is a trade secret lawsuit. To obtain an injunction, Apple must satisfy the Winter four-factor test: likelihood of success on the merits, irreparable harm, balance of hardships, and public interest. In trade secret cases, the strongest signal of irreparable harm is loss of secrecy. Once information enters an AI training corpus, secrecy is eroded not by a single leak but by a cumulative process. Each epoch of training bakes the information deeper into the parameters. A court order to "stop using" a secret is brutally difficult to enforce because deletion is not a discrete operation. You cannot un-cook an omelet. You cannot un-train a model. The phrase "immediate injunction" in the complaint indicates Apple is seeking a temporary restraining order or preliminary injunction. A TRO can be issued ex parte in narrow circumstances, but in a case with two sophisticated parties, the judge will likely demand adversarial briefing. The real fight will be over the evidence chain. Core: The Evidence Chain In 2020, while others chased yield, I built scripts to compare protocol revenue against on-chain deposits. The lesson I carried from DeFi Summer is simple: every meaningful event leaves a trace. Every transaction leaves a scar on the blockchain. In an employment dispute, the scars are not on-chain. They are in Slack logs, email metadata, credential access records, and checkpoint files. Apple's entire case depends on connecting those scars. Let's define a plausible evidence chain. Step one: a departing employee accesses a restricted repository or downloads a confidential specification shortly before resigning. Step two: that employee joins OpenAI. Step three: OpenAI incorporates that specification into a model architecture or training dataset. Step four: the resulting model exhibits behavior that cannot be explained by public knowledge. A court needs all four links. A jump from step one to step two is insufficient. The third link is the hardest to prove, and it is where forensic data analysis becomes the star. In my work auditing ICOs, I learned to compare the whitepaper with the bytecode; the promise is always prettier than the execution. With AI, the equivalent is comparing an API's output against a confidential target. Membership inference attacks can reveal whether a model was trained on a specific dataset. Gradient-based attribution can isolate which features influence a model's decisions. This is not magic. It is the same pattern recognition a blockchain analyst uses to cluster wallets. Pseudonymity is not anonymity; model blending is not immunity. Apple's burden, however, is complicated by the DTSA's requirement that plaintiffs file a "statement of trade secrets" under seal. Apple must specify, in concrete language, what secret was taken. If the secret is a chip design, a manufacturing algorithm, or a product roadmap, a judge can examine it. If the secret is embedded in a training run, the judge needs an expert to inspect the model. That creates a second-order leak: the secret gets exposed to the defendant's technical team, to court-appointed experts, and potentially to the public through redacted filings. Every court filing leaves a scar on the public record. I have audited cryptographic protocols where a single bit of entropy was the differentiator. I know what it feels like to hold a secret that must survive disclosure. Apple is about to learn that asking a judge to protect a secret is equivalent to inviting the opposition inside the vault. If Apple wins, the remedies do not stop at injunctive relief. The DTSA allows recovery of actual damages, unjust enrichment, a reasonable royalty, exemplary damages up to twice the award, and attorney's fees. In a case where model weights are trained on a stolen specification, those numbers can be astronomical. And the criminal side is not silent. The Department of Justice has pursued AI-related trade secret theft under the Economic Espionage Act. A civil injunction may be the opening move in a broader prosecutorial sequence. Trust is a variable that must be eliminated. Contrarian: Not Every Trace Is Contaminated The tempting narrative is "Apple cleans a thief." The data may be less satisfying. A preliminary injunction is not a verdict. It is a provisional freeze. Corporations have used emergency motions to slow a competitor long before they can prove misconduct. The cost of the motion itself is leverage. OpenAI may have to pause an entire product line while the motion is pending, lose enterprise customers, and face a public narrative that treats them as guilty until the court decides otherwise. There is also a legal doctrine that cuts against Apple: California's strong public policy in favor of employee mobility. The inevitable disclosure doctrine has been repeatedly rejected in California courts. If Apple's strongest evidence is "our former employee now works on a similar problem at OpenAI," the injunction should fail. Correlation is not causation. A similar problem is not a stolen solution. In the language of blockchain forensics, a shared address is not proof of a transfer. OpenAI will likely build its defense around its own data hygiene practices. If OpenAI can prove that it maintained a clean-room team, that incoming engineers signed declarations confirming they brought no external confidential information, and that training data went through provenance filters, then Apple's evidence chain breaks. The absence of a download log or a private communication is as meaningful as its presence. Silence is data too. This is where the contrarian angle bites. Apple may be overplaying its hand. A denial of the injunction would not only hurt Apple's case; it would set a precedent that AI companies can hire engineers without inheriting their prior employer's liabilities. That result would be a win for talent flow but a loss for trade secret holders. The courts are balancing two competing goods: freedom of movement and the sanctity of property. The judge will have to decide whether a model weight is more like a lockbox or a mirror. Takeaway: The First Hearing Will Set the Standard The next few weeks will produce a judicial answer to a question Congress did not anticipate: what counts as "use" of a trade secret when the use is spread across billions of parameters? If the judge grants a TRO, every AI lab will suddenly invest in cryptographic provenance, training data attestations, and clean-room compliance. If the judge denies it, every corporate counsel will watch the decision carefully. The law is slow. Arithmetic is fast. Data is the only witness that cannot be bribed. Apple and OpenAI are about to learn which side has the evidence.