AI is Hallucinating: Why human expertise isn't optional in intellectual property research work?

Let's be direct: AI is an extraordinary tool. It can scan thousands of documents in seconds, surface patterns no human would catch in a lifetime, and draft a first-pass memo faster than you can pour your morning coffee. I use it. You probably use it. Most professionals do now.

But there's a conversation we're not having loudly enough, especially in fields where a single citation, a single case reference, a single mis-stated fact can cost you your client, your license, or your reputation. That field is intellectual property law and research. And the AI hallucination problem here isn't theoretical. It's documented, expensive, and accelerating.

The Mata v. Avianca Moment That Should Have Stopped Everyone

In 2023, a lawyer named Steven Schwartz submitted a legal brief in Mata v. Avianca that cited six cases. Compelling cases. Detailed citations with docket numbers, judges' names, the works. There was just one problem: none of them existed.

ChatGPT had fabricated them wholesale, and when the lawyer asked the AI to confirm whether the cases were real, it assured him they were. He never independently verified a single one. When the opposing counsel flagged the citations as untraceable, and the judge demanded verification, the lawyer admitted AI use, and the AI had even confirmed the fake cases were real when asked. The judge sanctioned both attorneys with a $5,000 fine and mandated further legal education.

That should have been a watershed moment. A loud, clear alarm. Instead, it became the opening act of an ongoing crisis.

This Isn't an Isolated Bug. It's a Pattern.

Since that case, the numbers have only grown worse. Legal researcher Damien Charlotin, a data analyst and consultant, created a public database tracking cases where courts found AI had hallucinated quotes, invented cases, or cited non-existent legal authorities. Courts have detected AI-generated hallucinations increasingly each year since 2023, with lawyers, not just self-represented litigants, increasingly found to be at fault.

The consequences have been severe and varied. U.S. courts have taken a range of actions: some lawyers received warnings, others faced bar referrals, continuing-education orders, or suspensions, and some cases were dismissed with prejudice, meaning the party cannot try again.

In Colorado, attorney Zachariah Crabill received a one-year and one-day suspension (with 90 days served) for filing ChatGPT-generated fake citations and then blaming a legal intern. The case was characterized as the first US attorney-discipline ruling implicating AI misuse.

Then there is Michael Cohen who used Google Bard to research supervised-release precedents and forwarded three fabricated Second Circuit citations to his then-lawyer, who filed them without verification. Judge Jesse Furman called the episode "embarrassing and certainly negligent."

And it's not slowing down. The April 2026 apology from Sullivan & Cromwell to Chief Judge Martin Glenn for an emergency motion containing roughly 28 erroneous citations in the Prince Global Holdings Chapter 15 bankruptcy is the highest profile of a documented 1,348 worldwide cases tracked by the AI Hallucination Cases Database.

Intellectual Property Is Especially Vulnerable

Now here's what concerns me most specifically about IP work. General litigation is bad enough. But intellectual property research operates in a domain of extraordinary precision where the difference between a valid patent claim and an invalid one can rest on a single prior art reference, where trademark clearance decisions can be worth millions, and where the entire foundation of a client's business might rest on the accuracy of your research.

In 2024, two trademark opposition matters at the Canadian Intellectual Property Office, Monster Energy Company v. Pacific Smoke International Inc. and Industria de Diseño Textil, S.A. v. Sara Ghassai, both involved fabricated AI-generated citations. These weren't fringe cases; These were proceedings at a national IP office, and the hallucinations crept right in.

A UK Intellectual Property Appeal ruling involved a self-represented party who admitted using ChatGPT to generate citations that proved erroneous. The adjudicator also criticized the trademark lawyer for the opposing party for citations that detracted from rather than supported his arguments, and the lawyer's inability to explain those citations fueled the adjudicator's suspicion that AI was involved.

The UK High Court's commentary on one related matter is perhaps the most damning line I've read in this entire debate. Dame Sharp, president of the King's Bench Division, commented that it was "extraordinary that the lawyer was relying on the client for the accuracy of their legal research, rather than the other way around."

Even Legal AI Tools Are Not Safe

You might be thinking: "Fine, I won't use generic ChatGPT. I will use purpose-built legal research AI such as Westlaw, LexisNexis, or Ask Practical Law." Reasonable instinct. Still not sufficient on its own.

A Stanford study examining leading legal AI tools found hallucinated responses across multiple platforms including Westlaw's AI-Assisted Research product making up a statement in the Federal Rules of Bankruptcy Procedure that does not exist, LexisNexis's Lexis+ AI providing an incorrect legal standard, Thomson Reuters's Ask Practical Law AI failing to correct a user's mistaken premise and instead providing additional false information, and GPT-4 generating a statutory provision that does not exist.

These are not obscure startups. These are the most trusted names in legal research, and their AI layers are still hallucinating.

Why IP Research Demands More, Not Less, Human Judgment

Here is the core issue: AI models generate probabilistic text. They are predicting the most statistically likely next word, not retrieving verified facts. In IP research, you need verified facts. Specifically:

Prior art searches are perhaps the most misunderstood area when it comes to AI's limitations. Yes, AI can get the basics wrong fabricating patent numbers, misattributing inventors, or citing documents outside the relevant jurisdictional window. But the deeper problem is one that even accurate AI often misses entirely: understanding the context and conceptual tone of an invention.

A sound prior art search is not a keyword exercise. Two inventions can share identical language and even overlapping technical purpose, and still not constitute prior art against each other because what matters is whether the earlier disclosure teaches the specific problem being solved, in the specific way the inventor solved it. The inventive concept, the technical gap being bridged, the problem-solution relationship, these are things an experienced patent professional reads between the lines of a document. AI reads the lines themselves and little else.

An AI might surface a document that mentions the same components, the same field, even the same outcome, and flag it as prior art. A human expert looks at the same document and understands that it was solving an entirely different problem, from an entirely different technical angle, and is therefore not relevant. That distinction isn't semantic. It can be the difference between a granted patent and a rejected one, between a valid IP asset and a worthless filing. Relying on AI to make that judgment call, without expert human review, is not a shortcut. It's a liability.

Freedom to operate analyses require understanding claim scope in ways that are highly technical, jurisdiction-specific, and legally interpreted, not pattern-matched.

Trademark clearance requires understanding nuanced confusion standards, industry context, and how examiners at specific offices have historically ruled things that shift with precedent and practice, not just text patterns.

AI-assisted design tools can generate technical solutions that unknowingly fall within the scope of existing patents. A company that relies heavily on AI-generated outputs may find itself accused of patent infringement without ever having intentionally copied a competitor's technology. And for small inventors and startups, defending infringement claims tied to AI use could be financially devastating.

IP Monetization: Where Hallucinations Become Financial Catastrophes

If AI errors in research are dangerous, AI errors in IP monetization are potentially ruinous because here, the stakes aren't just about winning or losing a legal argument. They are about the dollar value assigned to your entire innovation portfolio, the terms locked into a licensing deal, or the royalty rate you have agreed to for the next decade.

Let's walk through where this gets critical.

Patent Valuation is one of the most complex exercises in IP practice. Determining what a patent is worth depends on claim breadth, remaining life, litigation history, market relevance, forward citations, and the competitive landscape, all of which interact in non-obvious ways. AI tools are increasingly being used to generate valuation scores and reports. The problem? Portfolio pruning and valuation require explainable AI, not black-box scores. Corporate IP teams need transparent scoring models that link legal strength, market relevance, and enforceability to renewal and budgeting decisions. When an AI assigns a high value to a patent with weak claim language, or fails to flag that its core claims were narrowed during prosecution in ways that gut its enforceability, a company can enter a licensing negotiation wildly overestimating its hand. The deals struck from that position can be challenged, unwound, or litigated for years.

Licensing Negotiations and Royalty Setting require a human understanding of industry norms, FRAND obligations, comparable license analysis, and the specific commercial context of both parties. AI tools analyzing historical licensing agreements to "recommend" royalty rates are working from pattern recognition don't understand why a particular rate was set, what commercial pressures existed, or what technical claim scope actually justified it. A hallucinated or misread comparable license can anchor a negotiation at the wrong number entirely, costing the licensor millions, or exposing the licensee to claims of bad faith.

Standard-Essential Patent (SEP) Assessment is an area where AI's limitations are especially consequential. In standards-driven industries, determining whether patents are standard-essential is notoriously complex, requiring semantic matching of claim language to standards documents spanning thousands of pages and millions of declared patents. An AI that incorrectly identifies a patent as SEP, or misses one, can steer a company into a FRAND licensing dispute it was never prepared for, or cause it to walk away from royalty revenue it was legitimately entitled to.

Technology Transfer and IP Due Diligence in M&A contexts is perhaps the most financially explosive area of all. When a company is acquired or when a patent portfolio is sold, the valuation underpins the deal price. If an AI invents non-existent technical literature or patent numbers during due diligence, the resulting assessment could expose the firm to massive Errors & Omissions claims over lost IP value. Imagine completing an acquisition based partly on an AI-generated IP valuation that overstated the strength of the target's portfolio, only to discover post-close that several key patents had prior art that the AI missed, or that the claims were far narrower than the AI's summary suggested. That's not a hypothetical scenario. It's a foreseeable consequence of deploying unverified AI outputs in high-stakes transactional contexts.

IP Portfolio Pruning and Maintenance Decisions also carry hidden risk. Companies regularly face decisions about which patents to maintain, which to abandon, and which to enforce. High-performing portfolio intelligence is built around repeatable decisions: pruning, filing, enforcement, and monetization, not static analytics. When AI recommends abandoning a patent it has misread as weak or flags a competitor's patent as non-threatening based on a surface-level claim analysis, those decisions can haunt a company for the remaining life of those rights. Rights, once abandoned, are gone.

The thread running through all of this is the same: AI in IP monetization can produce outputs that look authoritative, sound reasonable, and are factually or contextually wrong in ways that only a trained human expert would catch, and only if they're actually looking.

What Good Looks Like

None of this means abandon AI. It means use it correctly as a first pass, a draft, a research accelerant, and then apply human expertise to verify every material claim before it goes anywhere consequential.

The attorneys who have never been sanctioned for AI hallucinations are those who built verification into their workflow from day one. They use AI to surface leads, not to close cases. They treat AI output the way a senior partner would treat a first-year associate's memo: interesting starting point, must be checked.

In IP specifically, that means a qualified IP professional, not just any lawyer, not just any researcher reviews every prior art reference, every cited decision, every claim interpretation before it informs a filing, an opinion letter, or a litigation strategy.

The Bottom Line

AI is not going away, and it shouldn't. But the belief that it can replace expert human judgment in intellectual property research isn't just naive, it's a documented path to sanctions, malpractice exposure, and damaged client relationships.

The technology is powerful. It is also, right now, unreliable in ways that matter enormously in IP work. Until that changes, and there's no evidence it's changing fast enough, the most expensive thing you can do is trust it blindly.

Human intervention in IP research isn't a limitation on efficiency. It's the only thing standing between your client and a hallucinated disaster.

Have you encountered AI hallucination issues in your IP or legal research work? We would be interested to hear what verification practices your team has put in place.

#IntellectualProperty #IPLaw #AIHallucination #PatentLaw #TrademarkLaw #LegalTech #AIinLaw #AIinIP #Ai4IP #LegalResearch #IPResearch #BooleanIPConsulting