Opinion
A couple of days ago, parts of the standard-essential patent (SEP) ecosystem reacted to a completely misguided AI “analysis” of the InterDigital v. Amazon dispute like Native Americans when they saw glass beads back in the day. After a fundamental and easily proven error was shown on LinkedIn, the facilitator distanced himself from the first attempt — only to give it another try, digging an ever deeper hole with new absurdities (such as getting from an answer to a complaint in the Western District of Texas in May to an injunction or damages award by September) instead of recognizing the fundamental error for which there is no fix with present technology.
Trusting today’s GenAI to get anything right is completely stupid. Believing that today’s GenAI “thinks” is completely stupid, too. But not using it at all would also be completely stupid.
I was going to comment on the use of AI in this field anyway because the highest-ranking lawyer of a major player had a conversation with us at the recent Via Business Summit and could not believe that we produced so much content without letting AI do a large part of our job. Well, we did and we do, but my own workload has become unsustainable, so it’s now a team effort. I’ll get to our own use of AI toward the end, but independently of all of that I had the pleasure to introduce our content team yesterday.
I’m not going to take a position in this article on the use of GenAI by in-house legal departments and outside patent litigation counsel as a tool for such purposes as prior art search or the development of claim construction arguments. We have had reasons not to enter, thus far, into a partnership with any AI provider offering products or services to the patent law community, but that does not necessarily mean that if the right tools are used in the right way with manual verification and oversight by top-notch professionals, they could not be of some assistance.
Back to the fundamental issue of whether AI gets SEP litigation strategy right:
Plausible to the gullible: sounding smart while being stupid is what GenAI optimizes for
The single biggest problem with Generative AI (GenAI) is that it optimizes itself for one purpose: to sound plausible. To sound, not to be. If “plausible” in this context meant the same as under the U.S. Twigbal1 standard, the world would be a better place because that one focuses on substance, not sound.
The issue is not just data: the engine has a 46% hallucination rate according to a reputable independent study
Would you trust a surgeon with a proven 4% hallucination rate? An accountant? A management consultant? A lawyer? A website discussing high-stakes patent litigation? You wouldn’t. And what if it’s not 4% but a whopping 46%?
Obviously, if the starting point of any “analysis” (AI or not) is an article that has various shortcomings, such as mentioning only the least important one of four2 U.S. complaints, there is a data problem. But just like a politician mired in a scandal, the one who had some members of the SEP ecosystem in awe until I woke them up will only concede the undeniable.
And, just like many politicians do, he will distract you from a fundamental issue by putting a potentially fixable one front and center: he used an AI engine with a 46% hallucination rate.
The “analysis” was done, according to its facilitator’s representations, with Claude Sonnet 4.6. Give me a break. That is an engine for producing fake fictional literature, like telling the history of Ancient Rome with all main characters being anthropomorphic animals, but not for any kind of serious analysis:
Claude Sonnet 4.6 has a 46% hallucination rate!
Here’s the latest Omniscience benchmark by Artificial Analysis, and you can find Claude Sonnet 4.6 in seventh place with a non-hallunication rate of only 54%:
That means it hallucinates 46% of the time. Six other engines, even another Claude engine, are better. Grok 4.20 is best, but even that one is only at 83%, far below what any of us would want to rely on in any serious context.
Three quotes from the first version that are wrong, wrong, and wrong
So let’s see through the smokescreen, step by step. First version first, but the second version is also deeply flawed to put it mildly and we’ll get to that later.
It’s just a limited part of the problem that he initial “analysis” made the ridiculous claim that a case stayed for the next several years could move the needle. Here’s the quote, which is symptomatic of the real problem:
“The US Federal District Court in Delaware is present in the facts and absent from the prediction logic. It could serve as additional leverage or complicate a global license structure.”
Apart from the stay, that quote shows just how bad the thing is:
- No professional would seriously write “[t]he US Federal District Court in Delaware”. There’s an official title consisting of capitalized terms: the United States District Court for the District of Delaware. One can also say “the federal district court in Delaware”, but then one cannot capitalize the words because it’s not a name or official title.
- The “additional leverage” is a correct autocomplete operation and obviously applies to any patent infringement lawsuit, so the value of that information is precisely zero.
- The part about “complicat[ing] a global license structure” is typical hallucination. That case never had the potential to “complicate a global license structure” because it’s not a case about licensing. It’s simply an infringement action in which the usual patent remedies are sought. There could be, not before the 2030s and only absent a settlement, a damages award and maybe an injunction, but in order for that to happen, InterDigital would practically have to prevail in the ITC first (over the same patents) or the federal case would be an uphill battle. Now, even if a damages award came down, it wouldn’t complicate a global license deal. It would just be one of various economic facts to take into consideration.3 If Amazon wanted to push for a U.S. rate determination, this case would not lend itself to that, and Amazon wouldn’t have asked for it to be stayed if the case had ever had that potential.
Here are two other AI-generated idiocies from that first attempt that show the problem is hallucination that sounds like (but isn’t) serious analysis:
The consensus assumes both parties’ incentives point toward settlement. The critic argues those incentives are asserted rather than demonstrated, and that InterDigital’s portfolio strategy may affirmatively favor a public judgment over a quiet deal.”
There is not a single thing that InterDigital has done that even remotely allows a reasonable inference of InterDigital looking for a public judgment. But there are various things InterDigital has done to achieve a near-term settlement.
Anyone sophisticated in the ecosystem knows that no SEP holder as clever as InterDigital would pick a fight with the biggest bully on the block when trying to get a public judgment on its rates. And the number of cases in which any SEP holder really had that agenda is either negligible or literally zero. Public judgments happen when all else fails.
Also, the deal with Amazon, whenever it falls into place, will make a material impact on InterDigital’s revenues and profits, so indirectly the deal size will become known anyway.
And a third off-base quote:
“The penalty ceiling is treated as the primary coercive mechanism, but the 22 December 2025 order also states Amazon was ‘possibly already in breach’ of the ASI at that point — meaning the penalty exposure may have already begun accruing before the May 2026 appeal, which would materially accelerate Amazon’s settlement calculus beyond what the synthesis assumes.”
Sorry, but no one who knows about this stuff will seriously believe that the hypothetical contempt-of-court penalty by the UPC is a settlement driver. Procedurally, there are even two decisions needed by the UPC CoA (one on the merits after the late-May hearing, and later one on a hypothetical penalty). But the order of magnitude of the royalties at stake is far greater.
Even worse, a settlement would not even be guaranteed to avoid a penalty. This is a case where the court would have to take its own initiative as InterDigital’s hands are tied.
The original prediction was even ridiculous enough to look at the potential contempt fine “combined with the European Commission notification”: that one has no potential effect on Amazon itself. Amazon is not going to get fined by the EC. The UPC Mannheim LD informed the EC with a view to trade policy implications. That one is about the relationship between the EU and the UK. Nothing else. And any action taken, such as trade talks or a WTO proceeding, would also take years to make any impact even at an abstract level.
The second version fares no better on balance
Meanwhile, a second version has been released. Still starting from the same article that does not provide a complete and accurate picture of the dispute, but now with data from Courtlistener (where some documents are uploaded).
The prediction is now that a negotiated global license would fall into place “in late 2026 or early Q1 2027”, and that is obviously possible, but it doesn’t take a multi-agent AI setup to identify that possibility. The problem is, however, how this prediction is then explained:
- That window is allegedly “before any binding judicial FRAND determination issues from the UK High Court”, which is wrong not only because there is no UK High Court (a similar idiocy as “Federal District Court in Delaware”) but also because
- the UK rate-setting decision would not be on a predictable schedule (the court would likely try to hand down at least a partial judgment before the end of the year) and
- the EWHC decision will not be binding if there is an appeal (unlike in an Unwired-style SEP infringement action, there would not even be the leverage of a sales ban being potentially enforced).
- UPC contempt sanctions are allegedly “ripe and accruing”, but that is wrong. As explained in the previous section, those sanctions do not move the needle for a settlement and could be imposed by the court even after and despite a settlement. Moreover, “ripe” is wrong given that a merits appeal will be heard shortly, which is the main reason why there have been no sanctions yet, and “accruing” is legally wrong because the potential per diem rates have procedurally not started yet.
- As for “the UK trial producing a rate anchor both parties can negotiate around”, that contradicts the idea that the parties would settle before the court sets a rate. The court will set a rate after the trial, not at the trial. The parties’ positions with which they will go into the trial are already known to both of them.
- “InterDigital’s existential need to avoid a below-market public RAND determination that would cascade as a de facto ceiling across its entire licensing portfolio” is nuts as far as the last part is concerned: this is only about a few dozen video patent families, with many patent claims not even being subject to (F)RAND, but InterDigital’s “entire licensing portfolio” is mostly a wireless portfolio and its 4G/5G/6G patents (their crown jewels) would not be affected in the slightest. Not. In. The. Slightest.
Hallucinated fragmentation or fragmented hallucination?
What’s also funny is the reference to Amazon’s “fragmentation” strategy. That, too, is an example of hallucination. Amazon is not fragmenting anything. Amazon wants a global rate determination, even for non-SEPs, in the UK. That is the opposite of fragmentation, but not to an AI system that cannot think and that excels at producing fake fictional literature.
Brain-damaged timeline for Texas
The new analysis also talks about a fictional scenario of the case in the Western District of Texas potentially resulting in “injunctive relief or a damages award on a timeline that precedes the UK trial”. Actually, that filing was made a little over a month ago. Amazon will file its response to InterDigital’s complaint in early May. How can there possibly be any remedies within four months of the answer to the complaint?
More data was not (and is not) the answer
The premise that more data would solve the problem was misguided from the get-go.
Current GenAI systems will produce more AI slop, and if you give them more data, they wil be more creative, but they will not be more accurate on balance.
A 46% hallucination rate for the engine that was used means insanity. A human being with that hallucination rate would be considered unfit for survival in our society. Now, if you give such a lunatic an infinite amount of time and an infinite amount of text material to read, it won’t help.
I looked at the first analysis and saw that it contained absurdities. Same with the second. I’ve given examples above, and I don’t even claim that those lists are exhaustive. In a serious business, when inexcusable errors or even idiocies are demonstrated, that’s the end of the story.
Trillions have been invested in more AI infrastructure for the purpose of processing more data. And hallucination rates are still way too high for any serious task.
Too much SEP litigation knowledge is unavailable to AI systems
Even if one believed in more input data solving the problem (which is wrong for various reasons including the ones explained above), let’s be realistic about what is available.
This article here will be accessible to AI systems. Some of our other articles, or parts thereof, are available to them as well. I actually decide this strategically. It’s a multifactorial consideration, and two conflicting goals are influence (which matters to ip fray) and exclusivity (so as to ensure paying subscribers get something more that matters to them). Unlike those who never paywall or those who paywall virtually everything, ip fray has a freemium model. One of the benefits of the paywall is precisely that it allows us to withhold some important information and analysis from AI bots, while feeding them with things where the upside of providing our material to them outweighs the downside of not making the material exclusive. For example, we rarely paywall news pieces when we believe there isn’t really much that people couldn’t find elsewhere.
But that is just one example of SEP litigation knowledge that AI systems don’t have access to. In fact, while it wouldn’t solve the problem of those systems not truly being able to think, there is a hypothetical treasure trove of SEP litigation strategies in the form of analysis and recommendations written by lawyers, but subject to attorney-client privilege (or simply law firminternal material). AI systems will never get that type of content for their purposes. They can’t even solve that problem through technology.
So what is the outlook for InterDigital v. Amazon?
It’s pretty “rich” when AI slop that gets a lot of basic things wrong alleges that others have “underreported” or not sufficiently considered certain factors.
I do have an opinion, and I have voiced parts of that opinion, but I reserve that kind of analysis largely for paying subscribers. There are various factors that I see here and AI systems don’t. In fact, there is information I obtained by watching various court hearings and haven’t shared even with paying subscribers simply because a complete “brain dump” isn’t possible.
The right approach, for now, is to take it step by step. The U.S. proceedings may become interesting at a later stage, and I wrote in February that Amazon should have focused more on the U.S. than the UK. But before they reach a stage worth commenting on, a lot will happen elsewhere.
That is also something that AI systems tend not to do: just like they (Grok being a bit better in this regard than the others) don’t say “I don’t know”, it also never crosses their fake mind that sometimes the rational approach is to cross the bridge when one gets there. Of course, financial analysts look at trees where each node has a certain probability. But today’s AI systems can’t even do that properly.
Don’t be gullible. No everything that sounds plausible is correct.
How ip fray uses AI
As I said at the outset, it is stupid to rely on AI, but it is also stupid not to use it.
GenAI is not intelligent, so it can only be used responsibly if the user could do the task manually and is in a position to verify everything. In fact, one needs to know the subject well and have the skill to identify dubious claims easily.
We don’t let AI write our articles. We still find it useful and I personally use it for these purposes:
- It’s a means of casting a wider net than through traditional search. One just has to be aware of the high number of false (often simply hallucinated) positives. One needs good prompting techniques, and one must know what the strengths and weaknesses of AI bots are (which means we are talking about moving targets to some extent, but some characteristics are deeply engrained in product design).
- If I’ve read a document myself, I can use it for a targeted search within the document.
- For our links to the profiles of lawyers on the websites of their firms, I use AI prompts if I want to link to many lawyers in the same article, but I still have to verify each link manually as those AI systems like to make up URLs that don’t exist. Gemini often provides a link to a Google search that would then list the profile (as opposed to a link to the profile itself)…
- As our previous article notes, we understand the languages of practically all UPC proceedings so far, plus Brazilian documents, but obviously there are often situations where a team member needs a translation. We use three different translation tools, and two of them in two different modes each (PDF and web). Like with prompting and AI bots, one has to figure out what each tool tends to be good at, and that space also keeps evolving.
- We write in American English, but several team members instinctively write in British English. Writing tools can help by identifying such in consistencies.
- One can also ask AI bots whether a term is chiefly British or chiefly American English, in case of doubt.
- I’ve had AI bots generate tables for me that I copied and pasted into articles, but only if I was able to verify the facts. That is typically about the kind of data one can find on Wikipedia, greatly reducing the risk of hallucination anyway. I do not let AI generate tables such as the one about the four questions the UPC CoA referred to the ECJ.
- AI bots are great for tech support as long as one knows what one is doing. Even their technical advice could be as bad as telling users to glue pizza and eat rocks, but I’m a “techie” and so far haven’t even had such a problem. It’s just that sometimes I have to engage in dialog with the bots to point them in a direction where they can then help me.
- We will sooner or later set up a special research database where we use AI just as an interface, but again, verification is always needed.
Ronald Reagan famously said: “Trust but verify!” With AI, it’s more like “Don’t trust, but verify!”
GenAI is presently light-years away from Artificial General Intelligence (AGI), and for SEP litigaition analysis even AGI might not deliver good results: it may take Artificial Superintelligence (ASI). Whether GenAI is the right path toward AGI and ASI is the single most important question facing the present Digital Economy. I talked about that and other topics in a January 1, 2026 article that starts with AI patenting statistics.
- The standard for (surviving) a motion to dismiss, established by the U.S. Supreme Court in Bell Atlantic Corp. v. Twombly (2007) and further clarified in Ashcroft v. Iqbal (2009), requires that a complaint contain sufficient factual matter, accepted as true, to “state a claim to relief that is plausible on its face”. ↩︎
- The article in question mentioned (1) a filing by InterDigital with the United States District Court for the District of Delaware, but not (2) the materially consistent one with the United States International Trade Commission, to which the Delaware complaint is merely a companion lawsuit that has therefore been stayed (at least until after the final ITC decision, but if that one gets appealed, it will take even longer before the proceedings even begin); (3) a December 2025 complaint in the Eastern District of Virginia targeting the declared-free AV1 standard; and (4) a February 2026 complaint in the Western District of Texas over cloud computing technologies (February 24, 2026 ip fray article). We could talk about other limitations of that article, but it is not worth spending time. ↩︎
- It would not even complicate a UK court-ordered license deal. The England & Wales Court of Appeal has already looked at a similar question in Optis v. Apple. ↩︎