Using AI to Reduce Transactional Costs of Patent Validity and Infringement Determinations

“Based on a misidentified problem, we have implemented a decades-long policy to systematically weaken patent rights, in an attempt to deter [the] archetypal bad actor. If we instead use economic principles to address patent reform, we would understand the root problem to our patent system—exorbitantly bloated transactional costs.” United States has a clear need for patent reform, but does our legislature understand how to implement that reform?

For decades, a shortcoming to our approach to patent reform has been misidentifying the problem as “patent trolls” (more generally, bad actors). See here. Based on a misidentified problem, we have implemented a decades-long policy to systematically weaken patent rights, in an attempt to deter this archetypal bad actor. See here.

If we instead use economic principles to address patent reform, we would understand the root problem to our patent system—exorbitantly bloated transactional costs. See here for a deeper dive into economics, transactional costs, and our patent system.

Of the three main transactional costs, the focus of this article is on informational costs; specifically, the costs to determine (1) patent validity and (2) patent scope, i.e., whether a patent covers a given product or service.


The Three Main Transactional Costs to Determine Patent Validity

In essence, the three main costs associated with determining patent validity are time, capital and the resources needed to determine whether a patent is valid. The process to determine a patent’s validity is typically rooted in one or more of litigation and inter partes review (IPR). In litigation, the cost can range from $1 million to $4 million per party and can take several years, while an IPR could range from $100,000 to $700,000 per party and take 18 months. If there is an appeal, the time, cost and resources to determine validity goes up.

But if a patent issued by the United States Patent and Trademark Office (USPTO) comes with a “strong” presumption of validity, why does it cost so much to reassess validity?

This cost is driven in part by the success rate of defendants attempting to invalidate patent claims. To illustrate, in a white paper titled “United States Patent Invalidity Study 2012”, of the 283 district court cases that made patent validity determinations between 2007 and 2011, 243 cases held patent claims to be invalid in those cases. That’s an 85.8% invalidity rate at the district court level of at least one patent claim.

And from Lex Machina 2017 data relating to IPRs, in cases where IPRs were instituted (approximately 53% of the time when IPR petitions are filed), for the cases that go to final decision, only 5% of those cases have all the claims upheld.

Why are validity rates so low in adversarial proceedings?

The problem lies in the stark difference in resources allocated to examining a patent application during prosecution and resources allocated to invaliding a patent in an adversarial proceeding (e.g., litigation, IPR).

In prosecution, a primary examiner is tasked with reviewing the claims and conducting prior art searches. But this examiner is under time pressure to analyze the application and make a relatively quick determination whether to allow or reject the claims.

Often, this results on an examination that relies on a limited universe of prior art the examiner happens to find and is able to analyze in the allotted time.

During litigation or an IPR, however, parties with an interest in the outcome typically allocate substantially more resources to invalidate or defend a patent, typically due to the higher stakes involved in such adversarial proceedings. Typically, entire teams and search firms are deployed to find prior art that could invalidate asserted patents.

Not surprisingly, a substantial percentage of all litigation and IPR validity decisions involves prior art that was not even considered by an examiner during prosecution. This, in practicality, makes validity a de novo review during such adversarial proceedings.

This stark differential in resource allocation between prosecution and adversarial proceedings leads to low validity rates, which in turn lead to higher transactional costs to determine validity during such adversarial proceedings.

Hence, to reduce transactional costs to determine validity we must decrease this differential in resource allocation between prosecution and adversarial proceedings, to effectively increase the reliability of validity determinations during patent prosecution.

Using AI to Augment USPTO Validity Decisions

To decrease this differential, this writer strongly urges the development and application of robust artificial intelligence (AI) to help make more reliable validity determinations during the patent prosecution process.

For example, Carnegie Mellon’s Center for AI and Patent Analysis is developing AI algorithms that combine the capabilities of AI, machine learning, and natural language processing (NLP) to examine patent applications. The Center’s Patent Examination Initiative is directed to enabling patent examiners to “deliver high quality and timely examination,” by “deliver[ing] the most pertinent information and analysis for robust examination.”

While the Center for AI and Patent Analysis program is directed to assist decision makers, such as patent examiners, can we go a step further and develop an AI engine to actually examine patent applications, and make extensive searches and exhaustive analyses to provide highly reliable determinations on patent validity?

Could this AI determination of patent validity be robust enough that that determination could hold a high degree of reliability during later adversarial proceedings?

The Transactional Costs to Determine Patent Scope 

When it comes to determining whether a product or service infringes a patent, the informational costs are the time, capital and resources to make such a determination. This cost is rooted in litigation and takes a combination of one or more motions for summary judgment, a jury verdict and an appellate ruling.

The problem is that this process could take years and costs millions of dollars and could easily exceed the cost of the value of the patent itself.

For example, according to AIPLA data, when a litigation involves less than a million dollars of exposure to an alleged patent infringer, each party will expend approximately $700,000 – or a total of $1.4 million dollars to fight the patent lawsuit. Hence, in “low exposure” cases, the total cost of patent litigation exceeds the expected return on patent damages.

And when a litigation involves more than $25 million of exposure to an alleged patent infringer, AIPLA data estimates each party to expend approximately $4 million, totaling $8 million. But according to PwC Patent Litigation Study of 2018, the median patent damages award from 2013 to 2017 was just $6 million. In “high exposure” cases, the patent holder typically expects a damages award well above the median, but nonetheless still risks a damages award that does not justify the time and expense of the litigation itself.

For the vast majority of patents, the informational cost to determine patent scope in litigation exceeds the value of the patent itself. Ironically, this highly skewed transactional cost is the economic driver for “patent trolling” or “nuisance litigation.”

To truly reform our patent system, we must find a way to more efficiently determine patent scope, i.e., significantly lower the transactional costs to determine whether a patent reads on a given product or service.

Giving AI the Gavel?

Today’s methodology to make such a determination is done by humans, e.g., a judge, appellate panel, jury. If humans can do it, why can’t we implement artificial intelligence (AI) to more efficiently make this determination? Or at least provide a recommendation with a high degree of certainty?

Now, this doesn’t mean we would be replacing a patent holder’s right to a jury trial. Instead, can AI be used to more efficiently determine whether a patent reads on a given product or service? AI technology, if effectively designed and implemented, could eliminate years of litigation and save millions of dollars in wasteful transactional costs.

What would it entail to create a patent system where AI makes determinations or recommendations on patent infringement?  At a high level, this would involve (i) evaluating the scope of the patent claims and (ii) determining whether the claims map to an allegedly infringing process or system.

Regarding (1), AI algorithms would likely need to be combined with NLP and machine learning.

NLP helps us understand text in context rather than in a vacuum. NLP implements sentiment analysis (identifying underlying sentiment such as positive, negative, or neutral), intent analysis, and contextual semantic search (i.e., to give context to terms and phrases).

Machine learning is a branch of AI that takes known data to ideally make increasingly accurate predictions over time, as more data is inputted into the system.

In patent infringement litigation, AI, NLP and Machine Learning could be used to understand and analyze patent claim scope. To determine patent scope, AI could apply NLP tools such as intent analysis to show the purpose of the patent claim and contextual semantic search to relate the language in the claims to field-specific terms.  As more patents are analyzed over time, this  could help improve accuracy.

As noted above, Carnegie Mellon’s Center for AI and Patent Analysis is developing AI algorithms to address this very analysis. Notably, their Patent Initiative Model operates by creating “layers” of algorithms that extract and analyze text from patents. The lower layers are designed to directly process “text of patents, applications, Offices Action, Patent Trial and Appeal Board (PTAB) opinions, and other documents,” while the higher layers are designed to manipulate data, analyze, and make decisions based on analysis. This type of analysis could be used to construe claims.

Regarding (ii), to enable AI to operate in the realm of patent infringement analysis, it would need to go a step further to frame what a product or service does, and then make an assessment on infringement based on a claim mapping. In essence, this would involve the AI becoming fluent in two languages—the language in the patent claims scope and the language specific to the technologies alleged to infringe.

To become fluent in the language of a product or service, the AI would need to be able to take as input, e.g., source code, diagrams, schematics, and other specifications, and determine what aspect of the technology is relevant to the claims, and how the relevant aspect of the technology operates.

After the AI is fluent in the language of the patent claims and alleged product, it would then need to perform a claim mapping to determine a likelihood of infringement.

Ideally, the AI engine would be robust and powerful enough where its determinations and recommendations on patent infringement would be highly reliable.

Two Big Steps

Of course, we are a long way from AI reaching this level of competency and proficiency.  But if we could develop an AI engine to significantly reduce transactional costs to determine the informational costs of validity and to make patent scope / infringement determinations, this could modernize the patent system and enable highly efficient patent transactions. These would be two important steps in the right direction to improve our patent system.

Image source: Deposit Photos
Image ID:26946597


Warning & Disclaimer: The pages, articles and comments on do not constitute legal advice, nor do they create any attorney-client relationship. The articles published express the personal opinion and views of the author as of the time of publication and should not be attributed to the author’s employer, clients or the sponsors of Read more.

Join the Discussion

14 comments so far.

  • [Avatar for B]
    February 8, 2021 02:21 pm

    @ Steve Hoffberg “Ten thousand lines of code to always return “David Hasselhoff”. Sounds about right. Does it work?”

    Every time. No camera need.

    That said, if you ever heard the people at the EFF or R-Street speak, you’d realize that Anon is correct.

  • [Avatar for Anon]
    February 8, 2021 12:46 pm


    At least for me, my ‘sense of humor’ is lost by the fact that SO MANY say what you said, but have NO inclination to say such with any sarcasm.

    They say such in full earnestness.

  • [Avatar for Steve Hoffberg]
    Steve Hoffberg
    February 8, 2021 12:10 pm

    Sorry guys. I have been misunderstood. I meant, “if I was a defendant” (which I am not) “I would advocate for an AI algorithm that always gave me the answer I wanted to hear”. With a bit of sarcasm and bevity (a trait I rarely display), I wrote “As a defendant, it would be nice to train a net to determine that all patents are invalid.” Where is your sense of humor?

    Ten thousand lines of code to always return “David Hasselhoff”. Sounds about right. Does it work?

  • [Avatar for B]
    February 8, 2021 09:04 am

    @ Steve Hoffberg “As a defendant, it would be nice to train a net to determine that all patents are invalid.”

    Respectfully, Mr. Hoffberg, as a person who has written at least ten thousand lines of ANN code in a previous occupation, I could not imagine a more meaningless piece of code. Imagine a facial recognition routine that always returned “David Hasselhoff.”

  • [Avatar for B]
    February 8, 2021 08:43 am

    @ Anon “As a defendant, it would be nice to train a net to determine that all patents are invalid.”

    I tried to write an Alice/Mayo AI program, but the variables “abstract” and “inventive_concept” kept crashing the system while outputting:

    A101 error: undefined variable

  • [Avatar for B]
    February 8, 2021 08:37 am

    @ Anon “As a defendant, it would be nice to train a net to determine that all patents are invalid.”

    While I’m not sure “heresy” is the right term, I have created an AI program that does exactly this. The code is brilliant, compact, and reads as follows:

    Is_It_Valid (Patent_Claims, Prior_Art, Obviousness_Rules)
    If (day_of_month <= 32)
    outcome = Invalid
    return (outcome)

  • [Avatar for Anon]
    February 4, 2021 07:44 am

    As a defendant, it would be nice to train a net to determine that all patents are invalid.

    Taken on its face, the heresy is mind-boggling.

    Full force Efficient Infringer viewpoint. Not even trying to hide.

  • [Avatar for Steve Hoffberg]
    Steve Hoffberg
    February 3, 2021 06:27 pm

    The swamp will drain when there is a lower reservoir. Physics and metaphor. Gators prefer swamps over all else, and will not be simply drained away.

    I’m not sure what AI has to do with trolls. The maximum value cases are those which will not settle. Are the trolls the ones who settle or the ones who won’t? There is a narrative that trolls have bad patents. That makes no sense; if your gonna buy and assert a patent, why not a good one? But, how does AI address or solve any relevant issue?

    As a defendant, it would be nice to train a net to determine that all patents are invalid.

  • [Avatar for Pro Say]
    Pro Say
    February 3, 2021 05:15 pm

    Big +1 Hoffberg.

    “How then could the paradigm be accepted by all stakeholders?”

    For the excellent reasons you provide, it’s never gonna happen.


    Some swamps can simply never be drained.

  • [Avatar for Anon]
    February 3, 2021 04:37 pm

    Mr. Miller,

    You presume that any such “Tr0lls” have been acting not in accord with their rights, eh?

    (my oh my, has the Efficient Infringer propaganda k00l-aid been pervasive)

  • [Avatar for Charles E Miller]
    Charles E Miller
    February 3, 2021 03:24 pm

    As an economic disincentive to patent trolls, consider the use of post-trial motions for judicial awarding of fees and nontaxable expenses under Fed.R.Civ.P. 54(d)(2) and sanctions under 28 U.S.C. 1927.

  • [Avatar for Steve Hoffberg]
    Steve Hoffberg
    February 3, 2021 10:57 am

    The problem with AI is that is is based on mid to high dimensional patterns of preexisting training data. This would likely not work well when the premise is “novelty” or “nonobviousness”, since we must presume that the extreme tail of the statistical distribution is where decisions must be made. Note that it makes little sense to implement an AI solution to an easy problem. Consider that by law, one must constrain the scope of prior art to that which is “analogous”. Therefore, each determination requires a different training set, which can only be determined at the end of the process. How does AI embrace expert opinion? Consider also that at the time of the patent application, the key aspects may be described in terms and phrases that employ existing words to convey new meaning, and therefore break old patterns. Finally, consider the biases of the system. Presumably, the machine should give the same answer to the same question as judges and juries. If it does not, it is biased. This seems impossible. How then could the paradigm be accepted by all stakeholders?

  • [Avatar for TFCFM]
    February 3, 2021 10:01 am

    It seems to me that AI tools are just that: tools. Tools that are useable both by applicants and by examiners/challengers, moreover. For that reason, it doesn’t seem reasonable to expect these tools, more than any other, to change the “balance of power” between these conflicting parties.

    It is reasonable to expect a difference in “tool utility” between tools that purport to aid determination of objective facts (e.g., anticipation and literal support) and tools that purport to apply subjective criteria to make subjective determinations (e.g., obviousness and sufficiency of disclosure). The former have already proved useful; the latter likely cannot be relied upon, either legally or practically.

    Expanding search/translation tools have already knocked out a lot of would-be “inventions” which inventors realize, with the aid of search/translation tools, already exist. While these tools may not yet expressly address anticipation of a patent claim, such an AI functionality seems reasonably attainable. Similarly, tools have existed for a while now which will analyze literal (non-)support for claim terms in a specification, and some AI-powered extension to context might also reasonably be envisioned. These “objectively focused” tools, surely, should be expected to conserve some resources as fewer anticipated or non-supported claims get filed.

    AI-powered tools which address subjective determinations, like obviousness or beyond-literal support for claim scope, surely can be developed (and will likely differ in outcomes, depending on how they are written). However, because national courts are charged with applying the patent laws that the national legislature enacts (as interpreted by the courts, per national peculiarities), a court would likely commit reversible error by leaving its subjective legal determination to the whims of the programmers of an AI tool. For that reason, use of tools for this purpose is likely not legally permissible (absent, at least, a law enshrining a specified algorithm as the “law of the land;” debate the likelihood of that among yourselves). Even practically speaking, though, the susceptibility of subjective legal test outcomes to clever argument (objective, calculated algorithmic output be damned) advises against sole reliance on cold, calculating machines, so no party is likely to rely on such algorithms (and, as noted above, both sides can use them as ‘advice’).

    AI tools can certainly have a place (they already do) in patentability/validity determinations. I think it’s a mistake to predict that that place will be broader than “a mere aid to human thinkers.”

  • [Avatar for Anon]
    February 2, 2021 08:03 am

    That “misidentified” problem is there on purpose — and easily seen as such based fully on a Rational Actor justification of the Efficient Infringers.

    But I would add that the long-running attack on strong innovation protection is NOT limited to what I have labeled as the Right (not exactly congruent with the political Right). There are other ‘philosophies’ actively pursuing weakening the property aspect of strong innovation protection, notably the more traditional political opponent of personal property: the (more politically congruent) Left.

    It would be a huge mistake to not recognize that while certain aims may be in alignment, there are multiple parties with their own Ends in mind that are active in bringing about the current state of innovation protection disarray.