Hybrid by design: my crystal ball for AI in IP

Five predictions on where AI in IP is heading over the next three to five years, and why the answer is hybrid by design.

On 15 September I hosted an IPWatchdog webinar with Gene Quinn, Gregory Kline of Thermo Fisher Scientific and Azatuhi Ayrikyan of Astraveus. The title was Reducing AI Hallucinations: Why Reliable IP Workflows Begin with Better Data, and the replay is on IPWatchdog if you missed it.

Hallucinations make a great webinar title because everyone has a story. A patent number that does not exist. A family member that belongs to a different applicant. A confident summary of a claim the model never actually read.

My argument on the panel was that hallucinations are only the part you can see. Underneath sits a custody problem. For every dataset in your AI stack: who owns it, who audits it, who checks it, and who ends up using it?

In February 2023, Cormac Creagh and I wrote a report at Cipher on the patent risks around content-generating AI. Back then, most people I spoke to were asking whether these models were any good. Nobody asks me that any more. They ask where the model runs, what it is allowed to see, and who checks its work.

So since the webinar I have been thinking less about hallucinations and more about where all of this lands in the next three to five years. Here is my crystal ball: evolution, not revolution. We will keep what works and stop doing what doesn’t.

The legal world has started counting the damage. Damien Charlotin’s public AI Hallucination Cases Database had logged 2,077 cases by 23 September. Entries from the days around our webinar include lawyers who had used purpose-built legal AI products, not just a general chatbot.

That matches the research. A Stanford-led study of retrieval-based legal research tools found they hallucinated less than GPT-4, but still between 17% and 33% of the time. When our whole field is built on precedent and chains of citation, that is not an error rate we can live with.

IP has its own version of the problem, and it is quieter. Sometimes it is a fake case. Sometimes it is a family member missing from the set, a subsidiary never rolled up to its parent, or a Chinese utility model counted as a second invention.

The model answers confidently on top of a hole in the corpus, and the wrong answer reads exactly as well as the right one. Most of it never reaches a judge either. Search reports, landscapes and freedom-to-operate opinions are not audited in public, so their errors never land in anyone’s database.

So the fix cannot live in the model alone. It has to live in the data and in the workflow around it. Which brings me to the crystal ball.

A lot of people still use a chatbot as if it were a database. They ask it for a priority date or an assignee, and it answers from memory. That memory is a statistical impression of its training data, frozen at a cutoff, with no idea whether the patent it remembers lapsed last Tuesday.

I think that habit dies out. Language models are stochastic by design, so ask twice and you can get two different answers. Databases are deterministic. A hybrid setup simply gives each job to the thing built for it.

The model stays, but its job changes. It becomes the interface and the reasoning layer. The facts come from tools it calls: a search index, a family and ownership graph, a legal events feed, a classification lookup. Keeping those facts current and relevant becomes a mundane job for a good database.

The plumbing already exists. The Model Context Protocol, open-sourced in November 2024, is now governed under the Linux Foundation and supported in ChatGPT, Gemini, Microsoft Copilot and Claude. A model can already query a patent database much the way a searcher would.

So the test for any AI output in IP becomes simple. Can every statement be traced to a document number and a source I can open? If yes, a professional can check it in minutes. If not, it is a very articulate guess.

Every few months someone declares Boolean search dead. Boolean is pure logic and an antidote to generative stochasticism. Humans can audit the retrieval because we know the criteria we used to fetch the documents. We can set a risk and uncertainty threshold and decide whether we are comfortable with the result.

Semantic search is brilliant at finding what you did not think to ask for: the same idea in different words, from a different industry, in a different language. Boolean is brilliant at something else. It is reproducible and explainable. You can show a client, an examiner or a court exactly what you searched, and what you did not.

The future runs both together. The model drafts the first query, semantic retrieval widens the net, classification narrows it, and a human reads the result and spots the gap. The searcher who can do that last step becomes more valuable, not less.

The offices are moving the same way. The USPTO’s AI pre-examination search pilot, ASAP!, fed its tool the application’s CPC classification as well as its text, and the results notice becomes part of the public file wrapper once the application publishes (USPTO). Director Squires called it the first of many planned AI pilots (Nixon Peabody).

Uptake was cautious, with 169 petitions filed by mid-April against an original target of about 1,600. My read is that practitioners want to understand what the machine found, and why, before it lands in the record. That is a hybrid instinct, and a healthy one.

I expect most IP teams to end up running three tiers, whether or not they call it that:

• Public-data work, such as landscaping or a first pass over published art, goes to the best frontier model on offer. The documents are public already, and the quality difference still matters.
• Client-confidential work goes to enterprise deployments with contractual zero retention, known hosting regions and audit rights.
• Pre-filing inventions, trade secrets, litigation strategy and export-controlled technology stay inside the perimeter, on models you host yourself, air-gapped where it matters.

The USPTO flagged the underlying risks back in 2024. Its guidance on AI tools warns that AI systems may retain what users enter, that confidential material used for training can surface in outputs given to others, and that tools on servers outside the US can raise export control problems. It also says that simply relying on the accuracy of an AI tool is not a reasonable inquiry.

Big technology buyers are now drawing the same lines in public. In mid-September, Reuters, citing The Information, reported that Palantir, Nvidia and Booz Allen could restrict or stop using frontier models from Anthropic and OpenAI unless the labs guarantee they will not misuse their intellectual property. Nvidia reportedly limits Anthropic’s models to less sensitive tasks and runs its own Nemotron models for internal work.

Both labs say they do not train on customer data by default, and Microsoft is pitching isolated cloud environments to buyers with exactly these concerns. The direction of travel is clear either way.

An invention disclosure before filing is a trade secret with a deadline. I would not want it anywhere I cannot point to on a network diagram.

If the model runs inside your walls, it cannot phone out to a search API every time it needs a document. The corpus has to be in there with it: indexed, current, and yours to audit.

Running your own model used to mean accepting a big step down in quality. That is getting easier to live with. Epoch AI estimates that since January 2026 the best open-weight models have trailed the frontier by about four months on its capability index. For grounded work, where the model mostly reads documents it has been handed, those four months matter far less than the documents do.

The economics shift as well. A model on your own hardware is a fixed cost. A per-token bill grows with every searcher who finds out the tool actually works.

This is where data provision quietly changes shape. A chat window or a hosted API is not enough for these buyers. They need primary-source data they can take inside the perimeter, in formats built for indexing: full text, clean metadata, family and ownership links and legal events, refreshed on a schedule they control.

The USPTO nudged in this direction too. Its 2024 guidance tells anyone who wants to mine its records at scale to use its bulk data products rather than hitting its websites with AI tools.

My shorthand for all of this: own the data plane, rent the model plane. And where the work is sensitive enough, own the model plane too.

Most of the value will not come from a better chat box. It will come from workflows where every step has a named tool, a named data source and a log: invention disclosure, prior art search, claim chart, draft, check, sign-off.

Verification stops being a heroic late-night reread and becomes a built-in step. Every patent number is resolved against the source before a human sees the draft. Every quoted claim is compared with the published text, and every family and assignee is checked against the record.

The professional still signs. But they review the flagged exceptions instead of rereading everything, which is a far better use of an expensive brain.

Courts are already pushing this way. In the same week as our webinar, a federal court in New York imposed a two-year filing-disclosure sanction on a lawyer after fabricated citations. My bet is that “show me your verification trail” becomes a normal question from clients and courts, and eventually from offices.

That changes what good looks like for vendors too. The question stops being “which model do you use?” and becomes “what does your system check, against what, and can I see the trail?”

The risk here is token creep. Left inside an AI harness, these checks can end up costing more in compute than the frontier model itself.

Get comfortable with risk and uncertainty, but safeguard yourself with probabilistic outcome modelling. Before any piece of work goes near a model, ask:

• “If this leaks, how bad would that be?”
• “I’ll save a few hours by using generative AI, but this invention might be worth millions if it’s granted.”
• “Would my competitors gain an edge from uncovering the memo I’m writing with AI?”

Read up on synthetic data, too: just because data is anonymised doesn’t mean it can’t compromise intellectual property.

None of this needs a five-year programme. Most of it starts with a few honest conversations.

1. Sort your work by sensitivity: public, confidential, and pre-filing or secret. Then decide which models are allowed to touch which pile.
2. Ask every AI vendor the custody questions: who owns the data, who audits it, who checks it, and who ends up using it. That includes your prompts and your outputs.
3. Build a small test set of questions where you already know the answer, and run every tool against it before you trust it with anything live.
4. Keep your Boolean skills sharp, and teach them. They are how you audit what the machine found. We have to know how to look under the hood.
5. If pre-filing work matters to you, cost out running it in-house now. A fixed infrastructure bill is easier to defend than a per-token one that grows with success.

Good searchers have always mixed Boolean, classification and citation chasing, and good firms have always handled some files differently from others. AI is simply catching up with how IP professionals already work.

The IP system works in the dysfunctional way that it does largely because we built it that way. The information asymmetry game will continue, but now it’s token-based.

Thanks again to Gene, Gregory and Azatuhi, and to everyone who joined us. I would genuinely like to hear where your team is drawing its lines. What does your palantír show? (The Lord of the Rings crystal ball, not the company.)

This article was first published on LinkedIn.