Oracle's interim policy on generative AI bars model-generated content from OpenJDK source code, pull requests, and wiki pages. It also bars it from e-mail messages and Java Bug System issues. A bug report has never introduced a segfault. Neither has a mailing-list post. The policy page is dated April 9; it became news on Monday, when The Register set it against what Oracle's own executives keep saying, and it reached the top of Hacker News this week.
Read the FAQ and the reason isn't the one the headline implies. This is not a judgment about whether models write good code. It's a judgment about whether a contributor can promise they own what they're handing over.
The Oracle Contributor Agreement (OCA) requires that a contributor own the intellectual property rights in each contribution and be able to grant those rights to Oracle, without restriction. Most generative AI tools, however, are trained on copyrighted and licensed content, and their output can include content that infringes those copyrights and licenses, so contributing such content would violate the OCA. Whether a user of a generative AI tool has IP rights in content generated by the tool is the subject of active litigation.OpenJDK
That paragraph is the entire policy; the rest is elaboration. The OCA is a warranty, and a warranty a contributor cannot honestly make is worse than no warranty at all — it's a defect in the chain of title on every line it touches. Which is also why the ban reaches e-mail and JBS. The record of how a change was proposed, argued, and justified is evidence of where it came from. Oracle wants that record clean, not just the tree.
The Register found the contrast with Oracle's sales pitch puzzling. Larry Ellison told Oracle AI World that "the code that Oracle is writing, Oracle isn't writing," and Oracle cited AI deployment when it cut 21,000 jobs in June. It isn't puzzling. Oracle owns Oracle's code outright — nobody has to assign it, and there's no counterparty to warrant anything to. OpenJDK runs on contributions from people Oracle doesn't employ, under an agreement Oracle would have to defend. Same code, different chain of title.
The obvious objection is that the rule is unenforceable theater. Oracle concedes it first. The FAQ states plainly that reliably distinguishing human-generated content from AI-generated content "is impossible," then offers reviewers a list of tell-tale clues: emoji, unnecessary comments, gratuitously defensive programming, prose that seems "uncannily cheerful or meticulous." That isn't a control. That's a vibe. And the one crisp signal on the list is a Co-Authored-By trailer naming the tool — meaning the contributors it reliably catches are the ones who disclosed.
So Oracle shipped the mechanism that actually matters: a checkbox. Skara is being reconfigured to add a box to the body of every pull request, which the contributor must tick to affirm the contribution complies. Undetectable-but-attested is not a broken filter. It's the architecture of the Developer Certificate of Origin, which has never been verifiable either and has carried open-source provenance for twenty years.
The checkbox stops no one. It moves who is holding the liability when the question gets asked under oath.
Oracle will borrow to spend $70 billion on datacenters this year, up from $55.7 billion, and took an S&P downgrade to one notch above junk for it. It will bet the company on the model. It won't let the model near the one asset whose value depends entirely on who wrote it.