What if AI Invents Some or All Claimed Inventions?

By David Hricik, Mercer Law School

I’ve written a few posts about how I used specif.io to draft a patent application: I submitted a claim I’d found in a published application and the service drafted a 15-page spec, and created two figures.  Plainly, I invented nothing but assume for a moment I’d invented what had been claimed and that there was more disclosed in the spec than what I’d invented — the latter I think is fact but let’s assume it. Let’s also assume that I add claim 2 once I see the machine has conceived of something more than I had thought of.  So: claim 1 is my invention; claim 2 is not. I hire you to represent me.

The USPTO, the EPO, and the UKIPO have all stated (here, here, and here in respective orders) that only natural people can be inventors. Fair enough. The USPTO has stated that a person who is not an inventor cannot be named. Also fair enough, because of the statute and 102(f).  So… what do you do?

With my hypo, I think you have to name me since I invented what was in claim one.  I guess you don’t have non-joinder because the machine invented whatever else is in the spec and you can’t name it.

Now let’s turn to the facts of the application the USPTO rejected: the inventor took the position that only a machine was the inventor. So, let’s assume that fact pattern: AI invented what was in claim 1 and also what was in claim 2.

Now what?  That may explain why the EPO and UKIPO, at least, are studying this further and WIPO and AIPLA committees are, too.  Thoughts?

About David

Professor of Law, Mercer University School of Law. Formerly Of Counsel, Taylor English Duma, LLP and in 2012-13, judicial clerk to Chief Judge Rader.

30 thoughts on “What if AI Invents Some or All Claimed Inventions?

  1. 8

    The majority (probably the overwhelming majority) of disputes over AI-as-inventor would become entirely unproblematic if we amend the patent act to come into line with the copyright act, where the “author” of a work-for-hire is the entity who commissioned the work. In other words, if we say that (e.g.) IBM is the “inventor” of an invention produced by IBM-owned AI (or by the cooperation of IBM employees and IBM AI), then we just do not need to worry ourselves thinking about angels-on-the-head-of-a-pin question like “who first conceived this idea, the human or the AI?”.

    1. 8.1

      Sure – but then what about Stanford v Roche?

      Oh, for those in the peanut gallery wondering about the citation: 563 US 776.

    2. 8.2

      Interestingly, copyright law also provides an “originator can later challenge” the very beginning aspect that you point out.

      Of course, at least one contrast that you omit Is the 0bscenely different notion of “for limited times” between copyright protection and patent protection. Another contrast is that ‘the powers that be’ have also provided criminal notions into the infringement of only one of the two types of IP covered by the ‘patent and copyright clause.’

      Would you be willing to also put into the patent system similar length of term and criminal infringement notions from the copyright system?

  2. 7

    In this discussion, there seems to be an implicit dichotomy, and unclear use of pronouns. One question is, whether there is an invention at all? The second question is, if an invention exists, to whom/what is it legally attributed? These are two very distinct issues, and should not be blurred.

    Under a range of presumptions, the prediction of a useful embodiment comprising a would be “invention” by a machine based on “standard” inputs is likely unpatentable as being obvious. Likewise, if the inputs represent the inventions of another, there is a significant risk that the outputs will be derived (derivation) from the inputs and their source, and be attributed to the prior “inventor” and/or at least unpatentable to the subsequent inventor.

    But, once the machine generates a description of a non-derivative, non-obvious invention, the law appears to grant rights to the first to conceive, and conception cannot occur within the machine since it has no mind. (Thus, the first to conceive can be the first to comprehend a written output). If we assume that the machine is completely unguided, only on rare circumstances will it produce only a single output at a single point in time (especially if we presume that no one guided as to what and when), and rather there will be a stream of iterations to be selected from. In that case, I defend my hypothesis that the first legally cognizable entity to recognize (“discover”) the invention (from what are potentially a near infinite set of combinations and permutations) can obtain attribution, black box or not, Internet or not–quite simply, a race to the patent office per the AIA.

    This, of course, raises another issue that is quite present and real–whether existing judicial presumptions about the effectiveness of written prior art should apply when it was or could have been machine generated stream of unconsciousness, or even disinformation. Given that generation of phantom disclosures could be essentially exabyte scale malicious propaganda, why should there be any presumption that the opponent of this morass has any initial burden whatsoever? Shouldn’t the proponent of the reference have initial burden to show date, disclosure, operativeness, enablement, etc?

    This raises further issue: if the outputs are solely dependent on the inputs, is the invention made when the inputs are defined, or when processing and communication of the outputs is achieved? This can be perhaps further simplified to whether the key event in obviousness is availability of data supporting an as yet unformed expectation of success, or an actual expectation by a “hypothetical person of ordinary skill in the art” of success ?

    Another interesting question is whether the entire notion of a POSITA and analogous art needs to be reconsidered in view of consolidated databases of information that do not rely on expert classification of contents, i.e., whether “the art” is an outdated concept? How can we reconcile secondary indicia of non-obviousness with this entire scheme? Thus, can prima facie obviousness or obviousness per se of an invention be assessed by a generically trained machine, or must the machine be art-specific?

    1. 7.1

      Your view of obviousness is untethered to either the law or to the presented facts of this exercise.

      Yet again, it appears that you have a position in mind and are “working backwards” groping for rationale to support that position.

      The entire DABUS incident in which the present discussion is made in view of indicates that this is NOT a case of “mere use of machine.”

      You would do far better joining the conversation by accepting the conditions of the conversation.

    2. 7.2

      To the extent that the AI has not come up with an invention, but only a list of potential (combinatoric) possibilities, then simply reading those possibilities may in fact not identify an invention if at least some recognizable (or here creative attribution of) utility is needed to qualify as “invention” of the thing. But if the use is evident simply from reading those outputs, then the thing and at least one use is derived from the information presented. Whether that “discovery” qualifies as “invention” is precisely the question I posed, and at least historically we have required creative (human) input into the recognition of usefulness and not merely the identification of a novel thing generated by others (or by nature).

      1. 7.2.1

        Whether that “discovery” qualifies as “invention” is precisely the question I posed

        If you want to talk about discovery in relation to invention, I would love to hear your take on Sherry Knowles or Prof Lefstin?

        Both of these have presented ideas that broach the topic of the difference between ‘invent’ and ‘discover.’

        link to ipwatchdog.com

        link to ipwatchdog.com

        Indeed, one may wish to explore “Discoverer” as a type of legitimate alternative to “Inventor.”

        1. 7.2.1.1

          As my testimony to the Senate on the Tillis-Coons draft indicated link to judiciary.senate.gov (and as I have written about previously link to repository.uchastings.edu), not all discoveries are discoveries of inventors (which is a term of art) — just consider the so-called “discoveries” of Columbus and others that led to subjugation of the Americas. But more importantly, Jeff Lefstin’s work is terrific, and although I wrote earlier about this distinction, I did not earlier fully appreciate that the Supreme Court in Le Roy v. Tatham had adopted the “any application” of a discovery approach (which Neilson v. Harford had also adopted and that I had not realized was therefore misunderstood in O’Reilly v. Morse – I trust Jeff’s research on this). But I also argued (and in a long draft that I have not published) that O’Reilly impliedly overruled Le Roy on this point, and that since then the Supreme Court has adopted an “inventive application” requirement for claims relating to natural (and abstract idea) discoveries – most notably codified in the Ansonia Brass and Electric’s non-analogous use standard — which Congress apparently restored when adopting the Section 100(b) new use inclusion in the definition of process to overturn In re Thuau, but not authorizing all new uses — so much for textualism — and continuing through Funk Brothers. Inventive application thus was adopted by SCOTUS before 1952 and not (as Jeff’s work also shows, given rejection of bills that would have expressly overruled Funk Brothers) clearly rejected by Congress in the 1952 Act. And inventive application was followed by the lower courts after that Act and implicitly reaffirmed in Benson (which also improperly imported a preemption test out of no-where, conflating the reason for the inventive application test for natural discoveries with the test itself) and explicitly reaffirmed in Flook. But then Diehr impliedly overruled Flook while purporting to follow its basic principles (such as mere application to a new field of use does not create an eligible invention) – and note that Flook itself said that one must evaluate the invention as a whole BEFORE considering whether the claim reflects an inventive application and that claim parsing was not the basis for its rejection, so the arguments about needing to evaluate the claim as a whole are misguided. And then Bilski impliedly overruled Diehr. And the Mayo and Alice reaffirmed Flook while adding a new two-step analysis to figure out what inventive application is. And so it goes on, with SCOTUS refusing to acknowledge its inconsistent rulings and now with the PTO saying the claim needs to recite the ineligible discovery, even though the claims held invalid in both Bilski and Alice never did so, so the PTO’s guidance cannot (on its face) be a correct interpretation of the law (as it would hold valid the claims that SCOTUS held invalid). And all of us suffer for the confusion that then ensues.

        1. 7.2.2.1

          There is a GOLDMINE of wealth in that old link.

          Especially from this guy named ‘anon.’

    3. 7.3

      In that case, I defend my hypothesis that the first legally cognizable entity to recognize (“discover”) the invention (from what are potentially a near infinite set of combinations and permutations) can obtain attribution, black box or not, Internet or not–quite simply, a race to the patent office per the AIA.

      This is not a defense, but is instead merely a re-vomit of your assertion.

      An assertion that is simply incongruent with the law as it stands today – as already presented to you.

      Your “black box or not” is an instant
      F
      A
      I
      L
      Precisely because the dissociation if the second person from the actual creator of what is in the black box necessarily denies the legal title of inventor to that second person.

      You cannot have a situation in which the second person — on their own — is deemed “inventor” when the inputs may or may not have been actually invented (legally or not legally recognized as there being the putative inventor) by another.

      You are faced with a HARD logical full stop to your assertion.

      Merely ‘wanting’ the label if it turns out that the contents of the black box was created by machine and at the same time denying the label if it turns out that the contents of the black box was created by a human divests the definition of “inventor” away from the person to whom you would apply the label.

      You cannot merely ‘wish’ this problem away.

    4. 7.4

      [C]onception cannot occur within the machine since it has no mind.

      Wait a minute, how do you know that the machine has no mind. Surely it has no brain, but it is question begging to move from that postulate to the conclusion that the machine has no mind. To the extent that we really are speaking of artificial intelligence, I think you are fighting the call of the question if you build your answer on the supposition of no-machine-mind.

      1. 7.4.1

        I believe that inherent in his assertion is the lack of a human mind.

  3. 6

    “May I humbling suggest that you pull your head from your arse – it is not a hat.”
    (1) you don’t know my arse.
    (2) Comprehension of the invention is a piece of the puzzle, and AI cannot legally comprehend. The first to comprehend is an inventor, if there is an invention at all. Note in my earlier post, I made clear that “generically” trained AI may well provide dispositive evidence of obviousness, since based on the provided facts, the result is “expected”. However, if AI deduces an unexpected result, that means that something beyond accepted, available information was used to produce the result. Where did that some from? In some cases, it is not the inputs, but rather the filtering of the outputs.

    1. 6.1

      As to (2), whether or not AI can or cannot comprehend has no bearing on ANOTHER sitting in separate room, opening up a ‘black box’ (prepared NOT by the person sitting in that separate room), and BEING the inventor in any legal sense of the word.

      The ‘first to comprehend” is simply – and factually – wrong. Merely reading and comprehending the work created by another – even if that another happens not to be human – does NOT create the necessary legal nexus that you simply MISS.

      In the DABUS case, we explicitly have the matter that the person doing the filing FULLY recognizes the gulf between any mere preparation of inputs AND the result that is THEN put into a black box, and carried to a (possibly completely arbitrary) person sitting in a second room.

      As to (1), I may not know your arse, but I can clearly see the similarity between your mistaken views and the thrust of that put-down.

      1. 6.1.1

        ..note that my position does NOT require the AI to legally BE the inventor.

        You should be clear on that point.

        You very much have the situation in which it is quite possible to have an invention, but NOT have a legal inventor.

        Below, Professor Sarnoff insinuates this possibility in an overly broad manner by declaring this to be a dedication to the public (which is not a direct or necessary thing, but merely may be an effect of sharing an item in a patent application to which no legal inventor may attach). As Greg correctly points out, the current law IS of a certain nature, and one effect of that nature COULD BE that people recognizing the US limitation restricting legal inventor to be a real person might have the effect of keeping inventions from being shared in the patent system.

        This in NO WAY mandates or requires that the legal inventor MUST be a named human inventor if that human inventor otherwise would not be considered the legal inventor.

        You seem to merely want a result that the law just does not provide.

  4. 5

    Since the inventor is the first to conceive, and the AI is not “another”, then legally, the first to read, understand, and believe as operable an AI output may well be the legal inventor. After all, the manner of invention is not relevant to obviousness, and the machine is excluded from being inventor.

    The legal system in the US is not about intellectually satisfying conclusions which are correct, rational and workable. We need to distinguish the ethical result from the correct one.

    1. 5.1

      This statement is simply not correct: “ legally, the first to read, understand, and believe as operable… may well be the legal inventor.

      Treat this as a “black box” before the person that you want to designate as inventor.

      A second person sitting in a room, opening a black box that has been brought in and presented to him cannot legally be the inventor by merely reading (or even reading ‘and appreciating’) something CLEARLY that he did not actually invent.

      If you want to allow that scenario, then what do you do when that ‘black box’ is presented on the internet and you have thousands or even millions of “inventors” cotemperanously “appreciating” something not of their own devising?

    2. 5.2

      But the key question, deletion of Section 102(f) notwithstanding, is whether conception includes the possibility of derivation or requires “originality.” Merely reading someone or something else’s output doesn’t avoid derivation. Perhaps understanding the implications and applications that are not entailed by the recognition of the thing itself might, but then perhaps one is limited to a method or use claim…. This only goes back to old discussions of whether a new use for a known thing should be patent eligible, or whether all uses are implicit in the thing itself. What’s old is new again.

      1. 5.2.1

        I have to disagree, as you insert a condition and then pivot on your own insertion.

        The larger issue here is quite different.

        (for example, treat your insertion as having to being patent eligible all on its own — in view of the item in that provided ‘black box’ as being either separately patented or in the art.

        This type of ‘inventive OVER the AI work’ is not at point here.

    3. 5.3

      I provided an answer above (caught in a filter), but your rejoinder of “regardless” does not fly.

      You cannot merely wish the issue to disappear just because you do not like the logic that the issue presents.

  5. 4

    If I build a machine and add reagents that spits out useful antibodies that never existed before and I see a utility for the antibodies and know how to put those uses into practice and I write a patent application that describes all of this, I am the inventor of the antibodies and their uses, not the machine.

    I suggest that this reasonable view be applied in the case of inventions developed with the aid of AI.

    I certainly agree that there may be a problem with this when (if) AI becomes conscious and develops its own goals independent of any human creator or user of the AI, but let’s deal with that issue when (and if) it arises.

    1. 4.1

      Yes, but if is a publicly available machine (or one readily made by AI) then the question is whether simply desiring to use a machine to generate a result that is clear on its face is enough of an “inventive contribution.” Remember, that the inventor has to have had the mental conception of the full and operative thing – and only derived it from the AI. Even if it is sufficient, then you get the problem of whether it would be obvious to use the machine to do so. Once AI is generally available, all things produced by it may be obvious without a time, money, and effort threshold specified for use of AI (and even if the AI doesn’t get tired). And of course, KSR refused to specify that, notwithstanding the amicus brief I filed on behalf of economics and legal historians.

    2. 4.2

      Respectfully Mr. Hodges, your desire to wait until the moment of Singularity simply does not accompany the issue as it presently presents itself.

      Your “let’s wait” is not a workable answer.

  6. 3

    102(f) is only in existence for pre-AIA applications. And isn’t the right answer that the unclaimable subject matter (because it has an AI inventor that can’t be an inventor or co-inventor under 101 and 116) is dedicated to the public?

    1. 3.1

      [I]sn’t the right answer that the unclaimable subject matter (because it has an AI inventor that can’t be an inventor or co-inventor under 101 and 116) is dedicated to the public?

      I think that this is correct as an answer to the question of what the law is. As to what the law should be, the important jurisdictions (most importantly the U.S.) should amend the statutes to make it possible to list AI as an inventor, lest we incentivize keeping discoveries secret that might otherwise have been disclosed more promptly.

    2. 3.2

      Not sure that “dedicated to the public” is even close to being an answer, much less ‘THE’ answer.

      That being said, it certainly cannot be that anyone merely ‘reading’ what another (think black box delivery from another room — without regard to how the item IN the black box got in the box) came up with – CANNOT LEGALLY be the inventor.

      I have pointed this out since the DABUS case first started being discussed and to this date, not a single person has come up with a cogent legal reason why my views miss the mark – legally.

      Another cannot take what that person did not conceive.

      Conception remains critical to inventor.
      Some want to not understand this.
      Some want to pretend that those who merely did something at the start (and who have firmly stated that had no conception of the result) somehow should have that end result that there is legal basis for that “should.” Leastwise, one that resonates with PATENT law.

  7. 2

    If the “AI” interpolated from its training data to recite an expected method, apparatus, configuration, or compound, and the training data was a reasonable sampling of the prior art of which a POSITA would have been aware, and no human intelligence biased the result, then this should be considered prima facie obvious, and indeed, the AI model may represent the future of patent examination to substitute an objective standard for the subjective one employed by patent examiners.

    On the other hand, if any of the presumptions (or if I have misstated the required presumptions, then substitute the correct required presumptions), are not met, then the person who conceived of modifying the presumption set, or training data, etc., becomes the human inventor. This keys to the phrase in sec. 103 “Patentability shall not be negated by the manner in which the invention was made.” This reiterates that the inventor need not be a POSITA or even in the art, and may have a skill in the relevant art beneath a POSITA.

    Machines assist in the creation of conceptions all of the time, and across a full range of technologies. For example, when a chemical composition is identified by its spectrum, the spectrogram has never been considered the inventor of the composition. Likewise, a gene sequencer is not named as the inventor of a gene sequence, and Primer design software is not named as inventor of PCR primers.

    The issue here is that AI can be trained and used in areas outside the contemplation of the machine designer or human operator, dependent on training data from other sources. AI can extrapolate from the training data, and produce a range of results. However, this is merely an evolution of the field, and not really a new question, other than for ethicists, who never before realized that a thermometer or wheel was cause of an existential crisis.

    A compelling issue to consider is whether machine generated “prior art” should block patentability of future valuable inventions. Absent human commentary or assessment, are these automated combinations anything more than invitations to experiment? At that point, the selection of a machine-conceived invention for patenting becomes the “invention”.

    So, I weigh in against naming “AI” as inventor, and toward attribution of inventions made with automated assistance to the human influencer(s) over the process. In short, if the invention follows as an answer to a question, the inventor is the one who asked the question and pursued an answer by any means.

    1. 2.1

      Mr. Hoffberg,

      May I humbling suggest that you pull your head from your arse – it is not a hat.

  8. 1

    This has been pursued on other blogs, predominantly IPWatchdog.

    There seems to be a real sticking point in giving credit to AI for what you posit here as things that you admit to NOT being the inventor for.

    Some have even postulated that it is enough for a secondary person ambling along, picking up the results of the AI generated material, and merely reading that material to be deemed to be the inventor.

    I have pushed back at many of these pundits, but to date, no one has chosen to engage on the merits of the points that I have out on the table for discussion.

Leave a Reply

Your email address will not be published. Required fields are marked *

You can click here to Subscribe without commenting

Add a picture