When AI Learns Your Trade Secrets: Apple’s OpenAI Suit

Last updated: September 30, 2026

A senior engineer resigns on Friday, starts a new job on Monday. Somewhere in their head—and now inside a competitor's AI tooling—sits intellectual property your company spent months building. You have no clean way to pull it back out.

That is the nightmare Apple sketched out in a recent court filing. The company is asking a judge to speed up evidence gathering in its ongoing legal fight with OpenAI, and the arguments raise questions that reach far beyond Silicon Valley.

The short version:

• Apple is suing OpenAI over alleged use of its confidential know-how by former staff.
• Apple's attorneys argued that proprietary material passed into an AI could create uses that are hard or impossible to undo.
• Legal experts say existing remedies still apply, but the technical fix may be far harder than deleting a file.
• A related xAI suit against OpenAI was dismissed in June, showing courts are still working out how to treat AI-linked IP disputes.

What Apple Actually Argued

The core of Apple's filing is a paragraph that has lawyers talking. Its attorneys wrote that when proprietary know-how is absorbed by an AI that learns from it, that learning can spawn uses of the protected information that keep multiplying and cannot be walked back. Once a model has digested your confidential material, you cannot simply click "delete."

The brief was filed in support of an early-evidence motion—Apple wants the court to move quickly so material does not disappear while the case is young. Apple has not specified whether it believes any secrets were used to shape a model or only for a one-off AI-assisted simulation.

The Legal Tools Still on the Shelf

If you skim past the AI angle, the underlying problem is old. Companies have always worried about ex-employees walking out the door with sensitive material in their heads. Camilla Hrdy, a Rutgers law professor who studies trade secrets and generative AI, told Business Insider that workers have long been hard to police, and plugging what they know into AI tools makes that loss of control far worse.

Standard remedies remain available: courts can order a rival firm to stop using the secrets, bar further disclosure, and require protective steps. Plaintiffs can also chase damages, either actual loss caused by the leak or, in some cases, a royalty. None of that is new.

Why "Machine Unlearning" Is Hard

Sijia Liu, who teaches computer science at Michigan State and has published work on machine unlearning, told Business Insider that the technical fix depends on how the secret was used.

The easier path: if sensitive information lives in a document or knowledge base that an AI retrieves on demand, scrubbing or removing that file can help. The harder path: if the data shaped how a model was trained, the team has to draw a precise line around which capability should be removed and then try to excise it. Liu noted that influence baked into a trained system is genuinely difficult to measure.

A practical middle ground she described is a detection layer that flags sensitive prompts or agent traffic and slams the brakes before the secret goes anywhere. That is not true unlearning, but it is something an engineering team can actually ship.

Our Analysis

Trade-secret law was built for filing cabinets and email inboxes. The next generation of disputes will center on weight matrices and retrieval indexes—which is a very different kind of evidence locker.

The key tension courts will have to resolve: when does an AI system "use" a trade secret? Apple argues that ingestion itself constitutes use, which would lower the bar for plaintiffs. If that view prevails, any company whose secrets ended up in a rival's model—regardless of intent—could claim damages. That benefits deep-pocketed plaintiffs with patience and lawyers on retainer. It hurts smaller competitors who lack the resources for a years-long forensic fight, and it may chill legitimate hiring across the industry.

Watch the xAI v. OpenAI case closely. A judge dismissed it in June, partly because the alleged conduct looked more like individual misconduct than a corporate pattern. Courts may split hairs between "the company directed the use" and "an employee plugged the secret into a tool the company happened to provide." That distinction will shape how broadly Apple's theory ultimately reaches.

What to Watch Next

Apple is not inventing a new legal theory—it is showing courts a fact pattern that did not exist five years ago. Watch for two developments. First, whether the judge grants the early-evidence motion and signals that AI ingestion counts as a real "use" of a trade secret. Second, whether other plaintiffs start filing similar briefs, because litigation copycats tend to follow the first credible blueprint. If you handle sensitive IP today, do not wait for that case law to settle before tightening your own AI guardrails.

Key takeaways

• Treat AI assistants as repositories of last resort for any proprietary information, and document that policy in writing.
• Build detection layers that flag sensitive prompts before they reach a model, since true machine unlearning is still experimental.
• Keep detailed logs of who connected which data sources to which AI tools, so investigation is fast if a leak happens.
• Watch the Apple v. OpenAI ruling on the early-evidence motion as an early signal of how courts will treat AI-learned secrets.
• Review exit procedures to include a checklist on AI accounts, connected drives, and personal ChatGPT-style tools.

Source: businessinsider.com