Dan Gelbart on patent strategy in the age of AI

Key ideas

• A patent that misses one piece of earlier published work can usually be knocked out, and he estimates most patents filed before AI missed some.
• Keyword search missed prior art filed under a different word. AI search matches the meaning, so it finds the box, the housing and the enclosure alike.
• When someone says you infringe, start with the cheapest move: search with AI, then send a letter citing some of what you found, and keep the rest back.
• Every element of a claim has to be present to infringe, so a long claim is easier to design around. Write your own claims short.
• Before filing, have a few strong engineers try to get around your claims, because a lawyer isn't placed to find the workarounds.
• Only a person can be named as an inventor, so work an AI invented on its own can't honestly be patented as yours.

Dan Gelbart’s main point is that AI changes the cheapest move in patent fights. Searches used to match words, so prior art filed under a different name stayed hidden, and he estimates most patents still in force wouldn’t survive a thorough search. AI searches by meaning, so finding that prior art is now quick. These are my notes on his 40-minute talk: the order of moves he uses when someone says you’re infringing, how he writes a patent meant to be litigated, and where AI as inventor leaves the law. He’s clear that he isn’t a lawyer and that none of it is legal advice, and neither is this.

Why most patents are weak

He puts the count at about 12.8 million US patents, with about 3.5 million still in force. His claim is that most of those wouldn’t hold up, and the usual reason is prior art nobody cited. When you file, you and the patent office each search for earlier work, and if something new is left you get the patent. But anything published anywhere, in any language, counts, including papers that were never digitized. He estimates about 90% of unexpired patents could be knocked out this way. That’s his estimate from experience, not a measured figure.

The searches missed things because they matched words and classes. Call your invention an enclosure and you won’t find the same thing filed as a box, a housing, or whatever a translation from Chinese or Japanese turned it into. File it under electronics and you won’t find the identical device filed under medical equipment.

He also describes the patent office’s test for what’s new: whether someone who knew everything in the field, but couldn’t invent, would come up with it. (Strictly, that’s the obviousness test, which sits alongside novelty.) He expects AI to change what that test means, since an AI can now invent.

What AI search changes

AI search reads the whole query and matches the concept, not the word, so a few lines describing a box that holds electronics will find the housing and the enclosure too. His advice is to write the query as several lines of description rather than a few keywords, and to ask for prior art dated before the patent’s filing date.

Two things follow, in his view. Anyone facing a patent can now find prior art quickly. And patent offices will use the same search, so patents issued from now on should be much harder to break. Older patents, filed after weak searches, are where the opening is. Fields that predate digitization, mechanical and chemical work above all, still need a person in a library. Before AI he paid people to search non-digitized libraries in Moscow and China, and put five searchers on each case because each found different things.

When someone says you infringe

Every successful company gets these letters, he says, and a patent issued today was usually filed three or more years ago, before AI search. His order of moves, cheapest first:

1. Search with AI. Describe the invention and ask for anything published before the filing date.
2. Send a letter. Say the patent isn’t enforceable and cite some of the prior art, say five pieces out of ten. Often the holder checks it and drops the claim.
3. Ask for re-examination. Send the prior art to the patent office and ask it to review the claims. In his experience a patent with good prior art against it comes back much narrower and often no longer covers you. It carries a USPTO fee, and the patent stays in force while the review runs, but it costs far less than a lawsuit.
4. Read the claims for a workaround. If the patent survives, find an element of the main claim you don’t use, or change what you do so you don’t. If the claim says steel, make it bronze.
5. Check the expiry date. With a year or two left, a suit may cost the holder more than they could collect, even at the triple damages that can follow willful infringement.
6. Pay a fair royalty. If the patent is genuinely good, pay it. He puts typical royalties at 5% to 15%.

The reason to hold some prior art back is that whatever the patent office has considered is harder to argue later, since the court gives weight to the examiner having seen it. If the patent survives re-examination and you end up in court, you want fresh prior art. He’s dismissive of the standard lawyer’s answer, which is to wait and fight it at trial: he puts a US patent case at about $5 million in fees, of which a winner might recover about $2 million, plus months of management time.

Small companies and injunctions

He argues size works against the big company. A small company sued by a big one can’t be shut down while the case runs, because its sales are too small to hurt anyone. But a small company suing a big one can ask the court for an injunction, arguing it will be out of business before trial if the big company keeps selling.

His example is BlackBerry. NTP, a company that held patents and made nothing, sued Research In Motion, BlackBerry’s maker, and the threat of a shutdown pushed RIM to settle for $612.5 million in March 2006, even though, as he tells it, the patents looked likely to fall on re-examination. Injunctions have been harder to win since the Supreme Court’s 2006 eBay v. MercExchange decision, particularly for holders who don’t make anything, so the threat is weaker today than in the BlackBerry case.

Writing a patent that holds up

His process for a patent he expects to defend:

• Search first, with AI, so you cite every piece of prior art anyone could later find. If it could be worth millions, pay a few thousand dollars for searches of non-digital collections too.
• Write short claims. He passes on a patent attorney’s rule of thumb that a main claim longer than your hand is one the other side can probably get around.
• Try to break it. Show the claims to five of the best engineers you know and ask each how they’d get around them. Rewrite from what they find.
• Keep a related application pending. About a year after filing, he files a continuation-in-part. While it’s pending, anyone who reads your issued patent doesn’t know it exists. When they answer with prior art and workarounds, he adds those to the pending application, so the patent that eventually issues has already been examined against them. New material added in a CIP gets the CIP’s own filing date, not the original’s.

On timing, he says examination now takes about three years and has got slower since the America Invents Act of 2011 introduced paid fast-track examination, which he thinks rewards the office for being slow. If any inventor is 65 or older, a petition to make special moves the application ahead at no charge.

When AI is the inventor

He’s tested it himself: describe a problem and current AI comes back with a full invention. Under US law only a person can be an inventor. You can use AI as a tool while you invent, but you can’t ask it for an invention and name yourself as the inventor. He expects plenty of people to do it anyway, since there’s no way to tell how much the AI contributed, and doesn’t see an easy fix.

What I’d take from it

The practical part is the order of moves. A letter claiming infringement starts with an AI search, not a lawyer’s retainer, and most of the cost in a patent fight sits at the end of that list, not the start. For anyone filing, the five-engineer test and the short claim are cheap and make the patent much harder to get around.

Key terms

Prior art

Anything public that described the idea before the patent was filed, in any language, anywhere, at any time.

Unenforceable

His word for a patent that still exists but would fall in court once the missing prior art is shown.

Independent claim

A claim that stands on its own and sets the patent's real scope. Infringing it means using every element it lists.

Ex parte re-examination

Asking the patent office to look at an issued patent again in light of prior art you send. It often comes back narrower.

Continuation-in-part (CIP)

A follow-on application that keeps the patent family open and can add new material, which gets its own later filing date.

Injunction

A court order that stops a company selling a product while the case is decided.

Willful infringement

Infringing after you knew about the patent. The court can raise damages up to three times.

Petition to make special

A request that moves an application ahead in the queue. One based on an inventor being 65 or older has no fee.

Quotes from the talk