We may be entering a world with dramatically stronger patents. The rate of innovation and filings may stay the same or accelerate, but the bar for non-obviousness could rise as models get stronger.

Much of what we call innovation is really recombination. Someone at the edge of an industry notices things that don’t usually touch and pushes them together. AI is great at this. To test, I typed into Claude: “[Product X] is solving [Problem Y]. Research the latest trends in AI, crypto, and other emerging industries, and come up with 10 patentable ideas.” It produced genuinely great ideas. It suggested merging three of the ten because the composite would make a stronger filing. Pre-AI, I would have approved some for filing as in-house counsel.

And yet, under current doctrine, none of this is prior art.

Which creates a strange gap. We have “prior art”—what was publicly accessible or documented. But now we also have what you might call “latent art”: ideas that were always reachable by a person of ordinary skill in the art (POSITA) armed with a capable model, even if no one bothered to write them down.

It’s not hard to imagine litigators arguing that as smarter models proliferate, what a POSITA “knows” effectively expands. The POSITA now can ask better questions and traverse adjacent domains instantly. If an invention falls out of a generic prompt, was it ever truly non-obvious?

The answer may depend on the prompt.

At Level 0— Open ended. “Give me novel ideas in X”—if your invention pops out, it’s probably obvious.

At Level 1— Problem-framed. ”Here’s problem X. Solve it.”—it’s not much better. The problem itself isn’t inventive.

But at Level 2—Constrained. “Combine technology X with fields A through G to solve Y”— the creative act may shift into the question itself. The human who knew which disparate domains to combine may have supplied the real inventive step.

By Level 3 — Solution-specified, where the prompt essentially sketches the solution and the model fills in details, the inventive insight largely resides in the human framing, not the model’s completion.If your invention requires a Level 2.5+ prompt, you may have something defensible. If it emerges from Level 0 to Level 1, you probably don’t.

Today, the USPTO and filers run prior art searches. Tomorrow, they may run "latent art" searches: prompt the best available models with the problem your invention solves and see what emerges. If the solution appears in a few tries, you narrow your claims.

Paradoxically, that could make the patent system more valuable.

Because what AI is doing, at least for now, is predicting. But as most should know, unexpected results surface only after building, testing, and refining.

The patents that survive may be extraordinarily valuable. They will have withstood both traditional prior art & the fire of generative possibility. A win for the humans!