Your AI Chat History Could Decide Who Owns Your Invention. Experts Explain How to Protect It.

Years ago, Gary Shuster was on an airplane, squinting at one of the first iPods that could play video. He recalls that at some point he stared at the tray table in front of him and thought how convenient it would be to project the picture onto that.

That moment led him to design a portable device with a short-throw projector. But, it turns out, someone had beaten him to the idea by a few months. He then decided to stop the patent process. What stayed with him in the years that followed was how his patent lawyer handled the review. When the optics didn't quite add up, the lawyer simply asked how a device that small could throw an image from so close.

"He didn't become an inventor. He just told me your invention's flawed this way, and then I fixed it. So that's a perfectly legitimate use for AI, too. Tell me where I missed something, but don't tell me exactly how to put it in," said Shuster, co-founder of Inventiply and partner at DeepLaw, a firm focused on helping innovators manage their intellectual property strategy and legal issues.

Had his lawyer supplied the fix, he could have counted as an inventor. A chatbot can't be named as one. That distinction is easy to lose when today's collaborators are often AI tools built to hand over the answer. And more often than not, we find ourselves leaning on AI for more answers than guidance.

According to experts at DeepLaw, this could one day cost a founder a patent.

What counts as too much AI depends on what you're trying to protect. Today, it's important to know where those distinctions currently lie and how to best protect your idea.

Patents: AI can help with everything except the idea

Last November, the Patent Office reaffirmed that only a human can be named as an inventor, and that AI counts as a tool in the process, similar to lab equipment or a research database. Using it that way is fine. Shuster warns against letting it supply the idea you're trying to patent.

"There's no reason that AI can't help you organize the sections, or, if you give it a really bad handwritten drawing of something, generate the actual patent drawing. All of those things it can do," Shuster said. "But if you want patentable weight attached to something, it has to be human. For now."

The idea is where it stops. His DeepLaw partner, Cameron Powell, suggests creating a habit of putting that rule in the prompt.

"I'm going to tell you the problem I'm trying to solve ... All you should do, Claude, is give me a better understanding of the problem and adjacent problems, so I understand the problems better. Do not give me any ideas, don't give me any solutions," Powell said.

The reason, he argues, is protection. "You don't want to take ideas from AI, if you want to make them protectable. If you're looking for ideas for a book report or throwing a children's party, by all means, get ideas. But if you're trying to invent something that you want to protect, you've got to give Claude the opposite instruction."

As far as Shuster knows, no patent has been thrown out yet because of AI involvement. The timing may be why.

"Typically it's like three to five years, then you get your patents. So the ones coming through the pipeline now are largely pre-AI," he said.

Powell said the problem could very well surface in a lawsuit.

"If people are using AI now to assist in their patent applications, in drafting the actual idea, that information will not come to light for several years from now, because it'll be several years before the patent is granted, and then maybe more years after that before somebody challenges the patent, like a competitor," Powell said. "In discovery, if this information is still even available, either on the person's computer or at the LLM, then we would discover that the things that were claimed as inventions had been suggested by the LLM five or 10 years earlier."

An inventor then has two options available. Disclose the AI's role, and that part of the invention may not be protectable. Say nothing, and risk an accusation of inequitable conduct, essentially misleading the patent office, which can invalidate the patent.

"You want to be fully upfront and honest with the patent office about what part came out of a machine and what part came out of you," Shuster said. "It doesn't mean you can't use a machine."

Copyrights: the more the work is yours, the more you own

A copyright also requires a human. In March, the Supreme Court declined to hear a challenge to the human-authorship requirement, so the rule stands. Something an AI produces from a one-line request has no copyright. The harder question is how much you have to add before it's yours, and Shuster said nobody has drawn that line yet.

"If I did a thousand rounds with an AI that had written a short story for me, reshaping the story, I'm pretty sure I have a copyright interest in it. If I did zero rounds and just said, write a good short story, I don't have any copyright," he adds. "So somewhere in between those two points is where copyright arises."

He expects U.S. courts, not Congress, to decide where the boundary falls. Until then, Powell's advice for a business is to keep a paper trail instead of banning the tool.

"It's better to say, if you're going to use AI, we want to see your first draft along with your last draft, and please do us a favor of running a red line in between," Powell said.

The red lines are the changes a human makes to the machine's draft. As an easy rule of thumb, he said, you have copyright in the red lines as the employer. The untouched draft earns no protection. The gap matters most when you pay someone else: if a contractor's work was written by AI, Powell said, "there is no copyright to transfer."

Trademarks: one place where AI is welcome

A trademark identifies who makes a product, and here AI has more flexibility in the process.

"The trademark law does not say there's a creator of it. It doesn't say author, doesn't say inventor," Powell said. It only asks whether the mark is a unique designation of its source.

"That's protectable, so you can make your logo on Midjourney or ChatGPT and file it with the trademark office. Presuming it doesn't look too much like someone else's logo, but they don't care that you didn't write it yourself."

The limit is that the name has to be distinctive. Shuster notes that one that merely describes the product can't be protected.

For now, the rule across all three should be modeled around letting the tool point at the problem, polish the draft, and make the logo, while always keeping the idea and the shaping of the work in your own hands.