Talks

Why Patent Law Should Recognize AI Inventors: Ryan Abbott on the DABUS cases

63 min read

Ryan Abbott · 2024 · 54:57 · ▶ Watch

Summary

Central question. Should a patent office let an AI system be named as the inventor of an invention that no human conceived?

Main thesis. Yes, where a machine functionally did the inventive work. Abbott gives four reasons:

  1. Incentive. The patent system exists to get socially useful inventions made. If firms can only get patents when a human takes credit, they have less reason to build inventing machines. If the machine’s owner gets the patent, they have more.
  2. Transparency. The public should know how an invention came about.
  3. Chain of title. A named inventor is the start of a clear ownership trail.
  4. Honest credit. Nobody should take credit for work they did not do.

How he argues it.

  • He tests the law with real applications (two DABUS patents), not by lobbying: “I’m more a litigator than a lobbyist”.
  • He says the test for inventorship is conception, not “was a person involved”. Many people are involved in any AI output, but few of them conceived it.
  • He says ownership is a separate question. The machine’s owner would own the output by accession, the old property rule that the owner of a thing owns what it produces.
  • He says the philosophical objections (“it does not understand”) are irrelevant to a system that must judge outputs and behaviour.

The close. He predicts a hard problem ahead: hindsight bias in obviousness. If AI makes research cheap, an examiner can say “anyone could have asked an AI”. He suggests moving obviousness toward economic evidence. He then widens to AI and the law in general (tax, liability) and ends with advice to young lawyers to learn technology.

Stance. Advocate and litigant, but candid. He admits the plain reading of statutes went against him. He admits the human role is “mushy”. He admits where no human might qualify.

What the talk did not know (added context). Of the cases he describes, almost every court and office that decided the merits later ruled against him. The only “wins” were South Africa (formalities-only, no substantive exam) and the Australian first-instance ruling (reversed). See the Status table.

Map — talking points

#Point
1Intro: how he came to patents (medicine plus law, access-to-medicine work, patent bar)
2The 2013-14 insight: drug firms use AI for screening, where a human would be named inventor
3The academic case for AI inventors: incentive, transparency, title, no false credit
4Law is silent or assumes humans; Israel, Cyprus, Monaco claimed exceptions
5Meeting Thaler. Why a test case. Patent attorneys “just list yourself”
6Copyright as the guide: UK s9(3) (1988); why there are few cases
7Litigate not lobby. The US “individual” argument, and “whoever” in §271
8How DABUS works. The flashing light and the beverage container
9Devil’s advocate: a human prompts and picks the output, so is the human the inventor?
10The conception test. Trainer, user, recogniser. Google vs Siemens example
11Teams analogy. Inventorship is “mushy”. Why it matters (title, moral rights)
12Drop inventorship altogether? No: title, rights, transparency
13Case record: EPO, UK, USPTO refuse; South Africa grants; Australia wins
14Ownership by accession (Roman law, 3D printer, tree and fruit)
15Australian hearing story. UK Court of Appeal split. Supreme Court appeal filed
16Where the project goes: a broader case about law and machines. Policy activity
17Why do it: awareness, guidance, a policy debate
18”AI isn’t really inventing”: three reasons (definitions, causal involvement, no understanding)
19Turing, functionalism, the Tesla comparison
20”Publicity stunt?“
21Hindsight bias and obviousness; economic factors as a fix
22The book: tax, liability, perverse incentives
23Chinese-translation joke; advice to young lawyers

Putting it together

The mental model in one paragraph. Patent law gives rights to the person who conceived the claimed idea, and that person is named so that title, credit and the public record line up. Abbott asks what to do when a machine does the conceiving. His answer rests on the incentive purpose of patents: name the machine, give title to its owner by accession, and let the system keep encouraging people to build inventing machines. The courts and offices answered with the statutory words (“individual”, “person”, an inventor’s oath and moral rights) and said that only legislatures can change them. Their route is the opposite one: name the human who used the tool, and judge by the conception test. On this path, human inventors remain central, and most AI-assisted inventions have a human with a claim. The unresolved edge is twofold: the rare no-human case, and the effect of cheap AI research on the skilled person and so on obviousness. There, Abbott’s concern about hindsight and his turn to economic evidence are still the open frontier.

Three tensions to keep open

  1. Purpose vs text. If the aim of patents is more invention, does it matter that the words say “individual”? Courts say yes. Abbott says no.
  2. Name or disclose? Does transparency need a machine in the inventor box, or only a duty to say AI was used? The USPTO chose neither.
  3. Moving benchmark. If AI is routine in a field, the skilled person changes. Does that make patents rarer, or just shift where the inventive step lies?

Reading path

  • Start: the UK Supreme Court press summary of Thaler v Comptroller (short, clear). Then the Federal Circuit opinion in Thaler v. Vidal.
  • Then: the USPTO’s Nov 2025 notice (Federal Register) read next to Pannu v. Iolab and Burroughs Wellcome v. Barr. Abbott’s “Everything Is Obvious”. Graham v. John Deere and KSR v. Teleflex.
  • Deeper: Abbott, The Reasonable Robot (CUP 2020) and “I Think, Therefore I Invent” (2016). The EPO decision J 8/20 commentary and T 528/25 (abstract). The BGH X ZB 5/22 commentary. Samuelson (1986) on computer-generated works. Solum (1992) on legal personhood for AI.

Unverified on this page. Items marked ”†” in the text. Also not checked: Abbott’s claims about Israel, Cyprus and Monaco; the Korea and Japan procedural details; the India April 2026 refusal; the full text of EPO T 528/25; whether any office has changed the skilled-person standard for AI.

Part 1 — Before you read

Key terms

  • Inventor vs applicant vs owner. Three different roles. The inventor is the person who conceived the invention. The applicant files. The owner holds title. Usually the inventor assigns title to an employer. Abbott’s whole case rests on keeping these apart: the machine would be the inventor, a person or firm would be applicant and owner.
  • Conception. US doctrine. The “formation in the mind of the inventor of a definite and permanent idea of the complete and operative invention” (Burroughs Wellcome v. Barr, Fed. Cir. 1994; wording from memory)†. It is the test for inventorship in the US. Other systems use different wording (idea 4).
  • Reduction to practice. Actually building or testing the invention. Not the same as conception. A person who only carries out instructions is not an inventor.
  • Joint inventorship. Several people each contributed to the conception of at least one claim. US courts use the Pannu factors (idea 4).
  • Inventive step / non-obviousness. The invention must not be obvious to a person having ordinary skill in the art (PHOSITA), called the “skilled person” in Europe and the UK. This hypothetical person is where Abbott’s last question bites. Word warning: “obvious” is a legal term. It does not mean “easy” or “intuitive”.
  • Accession. A property doctrine: the owner of a thing owns what it naturally produces or what is added to it (fruit of a tree, a calf). Abbott uses it to give the machine’s owner title.
  • Natural person. A human being, as against a legal person such as a company. “Individual” in the US Act and “person” in UK law are the words fought over.
  • Formalities vs substantive examination. A formalities check asks if the form is right (is an inventor named?). A substantive examination asks if the invention is new, inventive and useful. South Africa does only the first.
  • Artificial Inventor Project (AIP). Abbott’s team of lawyers filing test cases for DABUS-invented patents worldwide. No machine “owns” anything in it.
  • Word warning, “AI inventor”. Abbott uses it to mean a system that functionally did what a human inventor would do. Critics use it to mean a system with intent or understanding. Much of the argument is two sides meaning different things.

People and sources named (reading list)

Person / sourceWhy they come upStart with
Stephen ThalerBuilt DABUS. Applicant in every caseHis Creativity Machine and DABUS papers†
Ryan AbbottThe speakerThe Reasonable Robot (CUP 2020); “I Think, Therefore I Invent” (Boston College L. Rev. 2016, from memory); “Everything Is Obvious” (UCLA L. Rev. 2019)
Lord Justice BirssUK Court of Appeal dissenter; mentioned a decade-old AI brake design caseHis dissent in the 2021 Court of Appeal decision
Justice BeachAustralian first-instance judge who sided with ThalerThe 2021 Federal Court decision (reversed 2022)
Alan TuringCited for what does it matter if it thinks”Computing Machinery and Intelligence” (1950)
UK Copyright, Designs and Patents Act 1988, s9(3)The earliest “computer-generated works” ruleThe text of s9(3) and s12(7)

Abbott’s own frame

Abbott holds the general view that law should be neutral between human and machine acts that do the same thing (his “AI legal neutrality” principle). He applies it to patents here, and to tort, tax and criminal law in the book. This page treats that framework only briefly. The full statement of it, and its tort, tax and criminal applications, are in abbott-reasonable-robot-ai-and-the-law. In this talk the framework appears in two places: the Tesla vs Uber driver comparison and the tax and liability examples.

The current debate (2021 talk vs October 2026)

Status table. Left: what Abbott said in late 2021. Right: what happened. Right-hand column is added context, not his view.

JurisdictionIn the talk (late 2021)What happened next (added context)
South AfricaA patent granted; no substantive examStill the only grant. Reported as a formalities-only grant, not a ruling on AI inventorship (Managing IP)
AustraliaFederal Court (Beach J) held an AI can be an inventor. Appeal pendingFull Federal Court reversed on 13 April 2022. High Court refused special leave in Nov 2022 (Lewis Silkin; date detail via Mondaq/Canada summary)
United KingdomUKIPO and High Court refused; Court of Appeal split 2-1 (Birss dissent); appeal to UK Supreme Court filedUK Supreme Court dismissed unanimously on 20 Dec 2023, [2023] UKSC 49: an inventor under the 1977 Act must be a natural person. It also rejected the accession argument and upheld treating the applications as withdrawn (press summary, IPWatchdog). The UK government’s 2022 consultation response kept inventorship unchanged (Mishcon)
United StatesUSPTO refused; E.D. Va. judge (Brinkema; “she”, per Abbott) looked unfavourableE.D. Va. refused (Sept 2021). Federal Circuit affirmed 5 Aug 2022: “inventor” means a natural person (Thaler v. Vidal, 43 F.4th 1207). US Supreme Court denied certiorari, April 2023 (World IP Review)
EPORefused; Abbott says it refused on formalities, after giving preliminary viewsLegal Board of Appeal J 8/20 (Dec 2021) upheld the refusal: an inventor must be a person (Kilburn & Strode). A second case, T 528/25, decided 5 Feb 2026, dismissed an appeal where the applicant named himself and added a note that DABUS conceived the invention. The Board said designating a human inventor is in principle possible for AI-assisted inventions, and declined to say how much human contribution is enough (EPO abstract; I did not find the full text)
GermanyNot discussedFederal Court of Justice (BGH, X ZB 5/22, 11 June 2024): only a natural person can be an inventor. It allowed an applicant to name a human who used AI to find the invention. Sources differ on how an added “DABUS generated this” note is treated (Mondaq, Bird & Bird). The EPO Board in T 528/25 said it was distinguishing the BGH’s handling of similar wording
Japan, South KoreaNot discussedJapan: Tokyo District Court affirmed refusal 16 May 2024; one source also lists an IP High Court affirmance on 30 Jan 2025. Korea: refused by the Seoul Administrative Court (June 2023) and Seoul High Court (May 2024); Supreme Court outcome not confirmed (IPStars; sources differ on details; not checked against court records)
Canada, IndiaNot discussedCanada: refused June 2025; Thaler appealed to the Federal Court in Dec 2025; no decision found (Mondaq). India: one law-firm summary says the Patent Office refused a DABUS application in April 2026 (single secondary source)†

Net result (added context). Out of the main patent systems, none now accepts a machine as a named inventor. Abbott’s claim that “opinions have changed really pretty significantly” has not held up in court. The one rule that has moved in his direction is narrow: several courts say a human may be named when AI was used as a tool (BGH 2024; EPO T 528/25).

The US patent office changed its guidance (added context).

  • Feb 2024: the USPTO issued guidance on AI-assisted inventions. It applied the Pannu factors to a single human working with AI.
  • 28 Nov 2025: the USPTO rescinded that guidance (Federal Register 2025-21457). The new text says no separate or modified standard applies to AI-assisted inventions. Only natural persons can be inventors. Pannu applies only to joint inventorship among several humans. Merely directing or “maintaining intellectual domination over” an AI system is not enough to be an inventor. It does not impose a general duty to disclose AI use, though existing duties of candor apply (Goodwin, IPWatchdog).
  • Effect on this talk: the new guidance lands near Abbott’s own conception test for the human side (“were you the one who conceived it?”), but it closes the door on the machine being named.

Copyright, the talk’s guide (added context).

  • US: the D.C. Circuit held in 2025 that human authorship is required (Thaler v. Perlmutter). The Supreme Court denied certiorari on 2 March 2026, no opinion (SCOTUSblog, Mayer Brown). The decision does not say how much human input is enough, because Thaler disclaimed any.
  • US Copyright Office, Copyright and AI, Part 2: Copyrightability, 29 Jan 2025: prompts alone do not give enough control to make the user the author, though human selection, arrangement and modification can be protected (IPWatchdog).
  • UK: Abbott points to s9(3) of the 1988 Act as the model for machine-made works. In 2022 the UK kept it. A December 2024 consultation preferred removing it. A March 2026 government report recommends removing it for works with no human author. As far as I could find, no bill has been introduced, so s9(3) remains in force (IP Kat, Mar 2026; I did not check a later update).
  • Effect on the talk: his prediction that copyright would lead was right about timing (copyright cases came first). It was wrong about direction. Copyright courts also insist on a human.

Obviousness and the skilled person (added context).

  • I found no court ruling or office guidance that has changed the skilled person because of AI. That is a gap in my search, not proof none exists.
  • Academic debate continues. Abbott’s own proposal is the “Inventive Machine Standard”: where inventive machines are the standard way to research in a field, the skilled person should be treated as an inventive machine (UCLA L. Rev.). Critics reject the idea that the benchmark should be an AI. A Dec 2025 SpicyIP post on an editorial by Dini argues the human-skilled-person benchmark should stay (SpicyIP). A submission to the UKIPO’s call for views suggests treating AI as a tool the skilled person uses, and adjusting the inventive-step level, not replacing the person (summary from search; not checked against the original).
  • Effect on the talk: his worry about hindsight is still open in the law I could find.

The ideas

The talk is short on citations and long on arguments that belong to older fields. Eight families of ideas meet here.

 INVENTORSHIP DOCTRINE ───────┐                    ┌────── STATUTE & INTERPRETATION
 conception · Pannu · joint   │                    │       "individual" · "whoever"
 inventors · UK s7/s13 · EPC  │                    │       Mohamad · textualism vs purpose
                              ▼                    ▼
                        ┌───────────────────────────────┐
 PROPERTY & ACCESSION ─►│  ABBOTT: if a machine did the │◄── PURPOSE OF PATENTS
 Roman law · Blackstone │  inventive work, name it, and │    incentive · prospect · disclosure
 tangible vs intangible │  give title to its owner      │    moral rights (Paris 4ter)
                        └───────────────────────────────┘
                          ▲             ▲              ▲
 NON-OBVIOUSNESS &        │  MACHINE    │              │ PHILOSOPHY OF MIND &
 THE SKILLED PERSON       │  CREATIVITY │              │ LEGAL PERSONHOOD
 Hotchkiss · Graham · KSR │  Koza · Boden              │ Turing · Searle · Solum
 hindsight · secondary    │  Thaler · Dendral          │ functionalism
 considerations           │  (history of invention     │ corporate / ship / river
                          │   machines)                │ personhood
                          │
                  COMPARATIVE & COPYRIGHT PARALLELS
                  CDPA s9(3) · Burrow-Giles · Naruto · Thaler v Perlmutter

Where each school of thought is covered:

  • Inventorship doctrine — what makes someone an inventor: idea 4
  • Statute and interpretation — what is an “individual”?: idea 9
  • Property and accession — who owns what a machine makes?: idea 6
  • The purposes of patents — why have inventors at all?: idea 1
  • Non-obviousness and the skilled person — the last question of the talk: idea 8
  • Machine creativity — what machines have actually invented: idea 3
  • Philosophy of mind and legal personhood — “does it have to understand?”: idea 7
  • Comparative law and the copyright parallel: idea 5

1. The case for naming the machine

In short. Abbott began to think about this in 2013-14, when vendors offered AI screening of antibody libraries. If a person did that work, they would be named as inventor and the firm would own the patent. He says machines do not care about patents, but firms do. He gives four reasons to name the machine: incentive, transparency, chain of title, and honest credit.

🎙 In the talk. The insight came in 2013-14. He taught inventorship and worked as a life-sciences patent attorney, and saw vendors offering AI screening of antibody libraries. His reaction: “when we have a person do that we make them an inventor and we own the patent”. So what if a machine does it and no person performs “that conception act”? He found academics had said since the 1960s that machines “don’t care about patents”, so nobody needs to ask. His answer: machines do not care, but firms do.

His four reasons are:

  • Incentive. “the owner of the machine would get a patent”, which encourages people to make and use these machines.
  • Transparency. Naming the machine would make it “transparent to the public of how an invention came about”.
  • Title. It “would assist with chain of title and entitlement”.
  • Honest credit. It “would keep someone from taking credit for work they haven’t done”. Example: if an AI makes 10,000 drugs in ten years, a human putting their name on all of them changes what “inventor” means.

He also says the law is mostly silent or assumes humans, and where it does say, it is because of an assumption that “people are the thing that invents” or because states did not want corporate inventors. A company “literally acts through human agents” and an AI does not, so the cases differ.

📚 Deeper theory

  • The incentive argument is the weakest link in his own chain. The firm already gets a patent if a human is listed. Naming the machine adds incentive only where no human can truthfully be named. So the argument depends on how often that occurs. In 2021 he thought it would grow; the 2025 USPTO text and BGH/EPO practice suggest firms will name a human instead (Part 1). Under utilitarian theory (Roots), the incentive is already supplied.
  • Transparency is an information-economics claim. It is similar to the disclosure bargain, but about the research process. It is not required by any patent statute. Practical uses would be policy (how much invention is machine-made?) and later challenge (is a human really an inventor?). The 2025 USPTO guidance puts no general duty to disclose AI use (Part 1).
  • The “false credit” point links to the honesty rules. US applicants must make an inventorship oath; knowingly naming the wrong people can taint a patent. The practice Abbott reports (list yourself, no one will ask) is the risk. His way of putting it: “a good way to get a patent not necessarily to have an enforcable patent”.
  • Economics of AI as a research tool. Cockburn, Henderson and Stern argued (2018, “The Impact of AI on Innovation”)† that AI could be a new general-purpose “method of invention”. If so, the question is not rare edge cases but standard practice.

📚 Roots: The purposes of patents — why have inventors at all?

The problem. To decide whether to name a machine, ask what naming does. Four theories of patents answer differently.

  1. Incentive (utilitarian). The US Constitution, Art. I §8 cl. 8: “To promote the progress of science and useful arts, by securing for limited times to authors and inventors the exclusive right to their respective writings and discoveries.” (Fun fact: the podcast’s name, Clause 8, comes from this.) The idea is the standard one in economics, traced through Fritz Machlup’s 1958 Senate study. Abbott’s headline argument is incentive-based: give the owner of the machine a patent, and more machines get built.
  2. Prospect theory (Edmund Kitch, 1977). A patent lets the owner coordinate further development. A pharmaceutical firm needs the patent to justify the trials. This supports his drug example.
  3. Disclosure. Exclusive right in exchange for telling the public how the invention works. Abbott adds a second disclosure: tell the public how the invention was made.
  4. Natural rights / labour / personhood (Locke; Hegel). You have a claim because you mixed your labour with the world, or because the work expresses your personality. These theories need a human. A machine fits poorly. This is the theory behind the moral rights reason for naming people and the reason a machine doesn’t care.
  5. Moral rights of inventors. Paris Convention Art. 4ter: the inventor has the right to be named. Abbott admits the right may have little “Direct commercial value” but inventors cling to it. He argues the transparency version of the same rule can apply to a machine.

The thread. Abbott picks the first three theories and sets aside the fourth. The courts and the offices, in contrast, started from the statute and its human-centred rights, which fit the fourth. The disagreement is partly about which theory the Act expresses.

Where others push back

  • Mere tool. The UK Supreme Court, the Federal Circuit and the EPO treat AI as a tool of a human. A tool does not get named, like a microscope.
  • No gap in incentive. A human who selects, runs, and recognises the output can be named, so the incentive survives. (Abbott accepts this when a human truly conceives.)
  • Moral rights need a person. Naming exists partly to credit a person’s creative act.

What they leave out. What the named human should say. After the BGH and T 528/25, an applicant may add a note that an AI was used. Neither Abbott nor the host discusses whether such a note helps or hurts the patent’s later strength.

Check yourself

  1. Which of his four reasons would still hold if every AI-assisted application named a human inventor? Which would fail?
  2. Why does he say it matters that “firms care about patents” even if machines do not?
  3. What is the difference between incentive and transparency as reasons to name the machine?

2. Strategy: litigate, don’t lobby, and how the cases were built

In short. He chose to litigate rather than lobby, saying “I’m more a litigator than a lobbyist”. The cases have still worked as lobbying. They filed at the UK IPO and the EPO first to get preliminary examinations on novelty, inventive step and utility. Refusals came only on formalities. The aim was also to give industry guidance and to start an open discussion.

🎙 In the talk. The idea of a test case came from a colleague who asked why there was no case. Abbott had interviewed Thaler for his academic work, and Thaler was “one of the earliest people to say they had had AI making new inventions”. They filed applications internationally. When the host asks why not lobby, Abbott says “I’m more a litigator than a lobbyist”, and adds that the cases “have proven pretty useful in a lobbying function”.

On the choice of offices: they filed first at the UK IPO and the EPO “so that we would get… preliminary examinations from those offices as to novelty inventive step and utility” before correcting the inventorship, to make sure the inventions were really patentable. (The quote above is trimmed; the exact words are “preliminary examinations from those offices as to novelty inventive step and utility”.) He says those offices denied on form, and he knows of no other case where a patent was denied for failing to designate an inventor “in this manner”. Where refusals occurred, “it’s been on a formalities basis”.

On the reasons for the project beyond winning: “to provide some guidance for industry”, to “start a discussion”, and to stop people “whispering about it”. Also, “robots are fun to talk about”.

📚 Deeper theory

  • A test case is a legal instrument with a known design. Think of strategic litigation in public-interest law: pick facts that isolate a single question. Here the facts were set so that no human had conceived the invention. The cost is that the facts are unusual. In real practice a human nearly always has some claim. Abbott concedes “almost all the time certainly there is someone you could make a colourable claim as an inventor”. So the cases test the edge, not the middle.
  • Why pre-clear patentability first. If the offices had refused for lack of novelty, the inventorship issue would never have been reached. The sequencing is a standard way to keep a test case “clean”. It also shows the filings were not a bid for the patents’ commercial value.
  • Forum shopping across 15-20 offices. The Artificial Inventor Project filed in many systems (the project’s own site lists them; I did not check). This produces a comparison across legal texts: same facts, different words. It is a rare natural experiment, and the table in Part 1 is the result.
  • Strategic litigation can move a policy debate even when it loses. Abbott lists consultations by the UK IPO, India, and a South Korean president’s remarks. The UK government’s 2022 reply and the US/EU/WIPO attention fit this. But the policy reaction did not change any statute that I could find (Part 1).
  • The “publicity stunt” charge. Public-interest litigation always faces it. The reply is that losing cases still produce rulings. By 2026 there are rulings from the UKSC, the Federal Circuit, the BGH and the EPO Boards.

Where others push back

  • The facts are artificial. Critics said the claim that no human was involved could not be tested, because a human built DABUS, chose the training and decided to file. Courts did not need to resolve the question and decided on law.
  • Wrong forum. Some think a legislature or international body (WIPO) is the right place, since a change in one country alone is awkward for the Paris/EPC rights.

What they leave out. The cost and the risk to Thaler’s applications. Also: what the project’s loss means for the incentive argument; a series of losses is itself evidence about the law.

Check yourself

  1. Why file at the UK IPO and the EPO first?
  2. What does Abbott gain from a test case that he would not from lobbying?
  3. Why does the Part 1 table show a stronger result for his opponents than the talk implied?

3. DABUS and the two inventions

In short. Thaler’s DABUS is built from hundreds of neural networks, each encoding a concept. A person mentors it, and then it combines simple ideas into complex ones on its own and spots which have value. It produced two inventions: a flashing light to draw attention in an emergency, and a beverage container. These were filed as the test cases.

🎙 In the talk. In the 1990s Thaler’s system used “at least two neural networks”: one trained on data (his example: car suspension designs) was made to corrupt its own weights to output new designs; a second judged them against a goal. That was “parametric optimization”, “pretty narrow”. DABUS “has hundreds of artificial neural networks each one basically encoding for a concept”, mentored by a person at “combining individual complex into simple ideas”, and then “in unsupervised runs it combines simple ideas into complex ones and identifies what it has an idea that has value”.

Example: the operator tells the machine to find something that “prevents death”; the machine proposes a light that flashes in a certain pattern, designed to draw human attention in an emergency (the caption says “attack”; it is almost certainly “attract”). That became the neural flame application. While they were doing a prior-art search, the machine volunteered an idea for a beverage container. “Why not take two shots just in case one of these inventions doesn’t work out”.

📚 Deeper theory

  • What the two applications are. From memory (not checked against the filings): one is a food/beverage container with a fractal-patterned wall, the other a signalling light. The talk’s descriptions match (container, flashing light).
  • The “perturb and judge” architecture is the same family as later generative-adversarial and evolutionary search (Roots). It is not a new kind of invention method, though its use for conceptual ideas is a claim the courts did not test.
  • Where the “invention” lies. For patent law the key is the claim: a container shape, a flash pattern. A claim is a specific, testable thing. Whether the machine or a person first formed a “definite and permanent idea” of that claim is a factual question about the process. No court heard evidence on how DABUS works. All the courts took the applicant’s statement as given and ruled on whether a machine can be named.
  • The weak point: human steering. Abbott accepts “the machine is mentored by a person” and that people set the goals (“prevents death”). Under US conception law, supplying a goal is not conception (idea 4), so this does not give the human a claim. But it is what critics point at.
  • The prior-art step. Choosing to file two applications for coverage (“better to have more chances”) is ordinary practice. It matters because the second invention (the container) came as a spontaneous output, which supports his claim that DABUS “identified these two ideas of value before a person saw them”.

📚 Roots: Machine creativity — what machines have actually invented

The problem. Is the claim “a machine invented this” an empirical fact or a legal label? The history shows the term is older than DABUS.

  1. Early search programs. Dendral (Stanford, from 1965; Feigenbaum, Lederberg, Djerassi) proposed chemical structures from mass-spectra data. Lenat’s Eurisko (1980s) produced novel designs and game strategies†.
  2. Genetic programming (John Koza, 1990s-2000s). Koza reported evolved circuits and controllers that were “human-competitive”, some re-inventing patented designs, and one patent-worthy new design†. A NASA evolved antenna flew in 2006†. Abbott says others used genetic programming and expert systems “before and after” Thaler.
  3. Thaler’s Creativity Machine (1990s). Two linked neural networks: one is perturbed to generate novel outputs, the other judges them. This is the architecture Abbott describes. Thaler held US patent 5,659,666 on the device†.
  4. DABUS. Abbott describes it as “hundreds of artificial neural networks each one basically encoding for a concept” that combines simple ideas into complex ones and flags the ones that “have value”. The flagging step is what makes it, in his telling, the inventor: the machine identified value “before a person saw them”. Whether the system works as described is not tested by any court. Offices decided on law, not on engineering. Treat the claim as the applicant’s.
  5. Margaret Boden’s three kinds of creativity (combinational, exploratory, transformational). DABUS-type combination of concepts is mostly combinational. Whether that is “inventive” in the legal sense is the question; a patent only needs novelty and non-obviousness, not transformational leaps.
  6. Since 2021. Drug-discovery systems have identified candidate compounds (the halicin antibiotic from an MIT group, 2020)† and protein-structure systems like AlphaFold have changed what is routine. In every case a human team set the goal and filed the patent. The central practical point is that patent practice has moved toward naming the humans.

The thread. Machines have produced candidate designs for decades. What is new is not machine output but firms naming the machine’s role in applications and a legal test case. The patent attorneys told clients to list yourself practice is how the issue stayed hidden.

Where others push back

  • The UK and European court records do not accept a statement that the machine alone conceived. The EPO, in T 528/25, said the applicant’s addendum (“I caused DABUS to conceive”) did not clearly designate a human inventor (Part 1).
  • Technical commentators doubt that a neural combiner is a “creative” machine as opposed to a search procedure. (General view; no source checked.)

What they leave out. The data and training. What concepts the networks encoded, how outputs were judged, and what a human did at each step. These would be the facts a court needs to apply the conception test.

Check yourself

  1. What does “identifies what it has an idea that has value” claim about the machine?
  2. Why does supplying the goal (“prevent death”) not make the human an inventor under the conception test?
  3. Which part of this description is the applicant’s statement, and which part did a court find?

4. Who is the inventor when a human is “all over this thing”?

In short. The host argues that people prompt and judge the output. Abbott replies that there are people all over the process, but the test for inventorship is not whether someone was involved. In the US the test is conception. A person can reach it by finding the problem, by training the machine, or by recognising the value of the output. A person who just receives the best answer may not qualify.

🎙 In the talk. The host plays devil’s advocate: someone must prompt DABUS and judge the output. Abbott replies: “there’s people all over this thing”, and DABUS did not “wake up one morning” and decide to invent, “but nor really did any of us”. His central claim: “the test for inventorship is not was someone involved in this process”. In the US the test is conception. There are three ways to get there:

  1. Find the problem. Sometimes this makes you an inventor, if the clever step was the problem. But usually we know what we want (a better suspension, an antibody for covid).
  2. Train the machine to solve the problem, “carefully selects training data” and adjusts it. That gives “a pretty good claim”. It gets harder with teams of hundreds, and when “the people who program or train the AI aren’t the people using it”. Example: a Google engineer who licenses a suspension AI to Siemens would not be an inventor of the output, “because you wouldn’t even know what the output is”.
  3. Recognise the value. If the AI gives ten designs and you test and find the best, “I think that pretty clearly makes you an inventor”. But if you just ask for a better suspension and receive “the best conceivable car suspension”, the recipient “has done anything involving conception or inventive skill” - his words: “I don’t think that that person receiving that output has done anything involving conception or inventive skill”.

For DABUS: it “was built and trained to invent it was not built or train to invent any particular thing” (the exact words: “it was not built or train to invent any particular thing”) and it flagged both ideas before any person saw them. He says “it depends where you are in the world because” “Germany and the US have different test for this”.

📚 Deeper theory

  • His three routes map onto US case law (Roots). Route 1 is “problem identification as inventive act” (a doctrine that exists but is narrow). Route 2 is contribution to conception by design of the process. Route 3 is “recognition”. Each is arguable, and each comes with the caution that a person who merely supplies a goal or runs a tool is not an inventor.
  • The Google/Siemens example shows a causal chain with no conception. A chain of human acts upstream of the output is not enough. This matches the principle that conception needs knowledge of the specific idea.
  • The consequence for modern AI-assisted work is a spectrum. At one end, a chemist who designs the search, narrows candidates and tests them is an inventor. At the other end, a user who types a request and files the first answer has no claim. Most real work sits between, and that is where the Feb 2024 USPTO guidance tried to help (it set out principles for a “significant contribution”) and where the Nov 2025 guidance went back to the plain conception test (Part 1).
  • Abbott’s dilemma for his own side. If route 3 works for humans, the human who spots the best of ten AI outputs is an inventor, so most AI-assisted work has a human inventor. That shrinks the set of cases where the machine would need to be named. He accepts this (“almost all the time” a human has a claim). It is also why the BGH and EPO treat naming the human as the practical answer.
  • A philosophical point: attribution of agency. The talk’s phrase “nor really did any of us” is a determinist jab: human inventors also do not self-start. In legal terms, the conception test sorts causal contributions into inventive and non-inventive, and it has done so for human teams for a century.

📚 Roots: Inventorship doctrine — what makes someone an inventor

The problem. Patent law gives rights to inventors, but statutes rarely define the act of inventing. Courts built the definition. Every argument in the talk about “the human pushing the buttons” depends on it.

The chain

  1. Conception as the core (US).
    • Mergenthaler v. Scudder (D.C. Cir. 1897) is often cited for the idea that the “mental part” of the inventive act is the inventor’s work, and reducing it to practice is not.
    • Burroughs Wellcome v. Barr Labs (Fed. Cir. 1994): conception requires a “definite and permanent idea of the complete and operative invention”. A hunch or a research goal is not enough. Wording†
    • Key point: conception is a mental act that has to be complete enough that a skilled person could reduce it to practice without more inventive work. This is why “I asked for a better car suspension” is not conception, and why Abbott says the user who receives the best output “has done” nothing “involving conception or inventive skill”.
  2. Not an inventor if you merely carry out instructions. Ethicon v. U.S. Surgical (Fed. Cir. 1998) and Burroughs Wellcome hold that someone who merely follows another’s instructions, or who supplies only well-known knowledge, is not a joint inventor. This is the legal root of “was really just a skilled technician”. (Cases†)
  3. Joint inventorship: Pannu v. Iolab (Fed. Cir. 1998). A joint inventor must: (a) contribute in some significant manner to the conception or reduction to practice of the invention; (b) make a contribution that is not insignificant in quality compared with the full invention; (c) do more than explain well-known concepts or the state of the art. A later case, Dana-Farber v. Ono (2020), stresses that a person must contribute to the conception of the claimed invention†. The Feb 2024 USPTO guidance tried to apply these to a person working with AI. The Nov 2025 guidance dropped that (Part 1).
  4. Recognition can count. Abbott’s third route, “recognizing the value of something can make you an inventor”. US law has older cases saying that recognising and appreciating the inventive result can be part of conception. It is contested, and the details vary by case†. Note what this means: in a world of AI outputs, the person who picks the winner may be the only human with a claim.
  5. UK and Europe.
    • UK Patents Act 1977: s7 says the “inventor” is the actual deviser of the invention; s13 gives a right to be mentioned. In Yeda v. Rhone-Poulenc [2007] UKHL 43 the House of Lords said an inventor is the person who devised the “inventive concept”, which is the claimed idea, not everyone who helped†.
    • EPC: Art. 60 (right to the patent), Art. 62 (right to be mentioned as inventor), Art. 81 and Rule 19 (the application must designate the inventor). The EPO does not itself examine who is the true inventor. It checks the designation is a clear, lawful one. That is why the EPO refused on form and why the 2026 T 528/25 appeal failed on the wording of a designation.
    • Germany: Abbott says the German and US tests differ. In German law the inventor is the one whose contribution shaped the “teaching”, and a person who merely provides the problem or checks results is generally not†.
  6. The statutory point about people. The inventors’ oath and declaration (35 USC 115) and the definition of “inventor” (35 USC 100(f), “the individual or, if a joint invention, the individuals collectively who invented”) are written for people. This fed the US result (idea 9).

The thread. Everywhere, inventorship is about who conceived the claimed idea. None of the doctrine asks “who was in the room”. Abbott’s argument takes that doctrine seriously and then asks: if no human meets it, what then?

Where it strains.

  • Conception was designed when ideas came from heads. A system that produces candidate ideas by search or generation shifts the mental act. If the user only specifies a goal, the “definite and permanent idea” arguably appears only after the machine makes it.
  • Teams: the law has long tolerated crude attribution. As Abbott says, it “illuminates how mushy inventorship is as a concept”.

Where others push back

  • The Federal Circuit and the USPTO (Nov 2025) say the human who sets up and uses the AI as a tool, and then conceives the claimed result, is the inventor. They deny that a “no-conception” gap needs filling by the machine.
  • Some argue “recognition” should not be enough on its own, since it rewards the person who happens to be first to look. Abbott himself floats giving the first person to see a bare AI output inventorship “regardless of whether it involved inventive skill” as a policy choice.

What they leave out. How to prove, in a dispute, what the machine did and what the human did. Audit trails and logs matter, but nobody in the talk raises record-keeping.

Check yourself

  1. State the three routes to human inventorship and give one example of each.
  2. Why would the Google engineer not be an inventor of Siemens’s suspension?
  3. Why does route 3 weaken the need to name the machine?

5. Why inventorship matters: mushy, but useful

In short. The cases show how mushy inventorship is, even for human teams. Abbott says abolishing inventor naming would make life easier for patent attorneys, but it matters for chain of title and for moral rights. A machine does not care about credit. Naming it still tells the public how an invention was made, and it acknowledges the people who built the AI.

🎙 In the talk. Teams analogy: a lab of five, and an attorney asked to list all five. Some will check “who did what”. The case “illuminates how mushy inventorship is as a concept”, and where a team spans Pfizer and Novartis under a collaboration agreement, “who was really just a skilled technician” matters. As teams grow and use more AI, “moral inventorship issues are going to take on more economic importance for entitlement reasons”.

Host: should we abolish inventor naming? Abbott: it “would certainly make life easier for patent attorneys”, but it matters for chain of title and for moral rights, which inventors guard. A machine “doesn’t care about being acknowledged”, but naming it is “being transparent to the public”. In 10 or 20 years, “when we get covid 25 Pfizer and GSK and Johnson and Johnson just go to AI”, and then “I think it will be good for the public to know how invention is made”. He also notes it acknowledges the people who built the AI, as a PhD supervisor is acknowledged on a student’s patent.

📚 Deeper theory

  • Inventorship as an ownership gate. In many systems rights flow from the inventor to the owner. If a research partner owns “the inventions of its employees”, the list of inventors decides who owns. This is why inventor lists are litigated in joint ventures far more than in single-firm work (Abbott’s point).
  • Employee-inventor compensation. Some systems give inventors extra rights. Abbott mentions “the EU where there’s special rights for inventors for very profitable inventions” (Germany’s Employee Inventions Act is a known example)†. If a machine were the inventor, no human could claim those payments. Naming a human would revive them. This is an unexamined side effect of the whole project.
  • Moral rights in the Paris Convention (Art. 4ter): a right to be mentioned. In practice it is a weak right (often a name on the front page), which fits Abbott’s remark that it has little direct commercial value.
  • The PhD-supervisor analogy raises a subtle point: crediting those who built the system. Contemporary practice does not list a tool’s builder as an inventor of its outputs unless they contributed to the conception. Abbott offers credit as “acknowledgement”, not inventorship.
  • Alternatives to naming. Some jurisdictions (the host mentions “companies could be inventors”) could remove the requirement. Argument for: simpler. Against: loss of the audit trail and of moral rights, and mismatch with international rules (Roots).

📚 Roots: Comparative law and the copyright parallel

The problem. Abbott uses copyright to predict patent. How good is the parallel?

  1. UK s9(3) CDPA 1988. For a computer-generated literary, dramatic, musical or artistic work, the author is the person who undertakes “the arrangements necessary for the creation of the work”. The term is shorter (50 years from creation, per s12(7))†. There is almost no case law, a point Abbott explains.
  2. US human authorship. Burrow-Giles v. Sarony (1884) described the author as the one who gives “mental conception” to a photograph. Naruto v. Slater (9th Cir. 2018) held a monkey cannot sue for copyright in a selfie. The Copyright Office’s practice refused registration for non-human works. Thaler v. Perlmutter (D.C. Cir. 2025; cert. denied March 2026) applied the rule to Thaler’s own AI art. See Part 1.
  3. Why the parallel is imperfect.
    • Copyright arises automatically on creation, with no examiner. Patents are granted after examination, with an inventor named. A false inventor is a bigger formal problem than a false author.
    • Copyright’s “author” can be a company under work-for-hire. Patent law does not allow a company as inventor (the host’s and Abbott’s point about “IBM invents”).
    • Copyright bars are low. Patents need non-obviousness.
  4. International rules. The Paris Convention (Art. 4ter) and the EPC give inventors a right to be named, which is part of why a change cannot be done in one country alone. The UK government’s 2022 reply cited this concern about unilateral change, per Mishcon.

The thread. Copyright is the older, smaller test bed, and so far it has said the same thing as patent: human required, AI-assisted work protectable where a human contributes.

Where others push back

  • Naming a tool is odd. No one lists the microscope, the spreadsheet or the supercomputer. The line that makes AI different is its autonomy, which courts did not accept as proven.
  • Disclosure of AI use is a different instrument. If the aim is transparency, a duty to disclose AI use could do it without naming a machine as inventor. The USPTO, in Nov 2025, declined to impose a general duty (Part 1).

What they leave out. Fraud and accuracy risks of the other direction: a human who names themself on work they only supervised (the “false credit” he worries about). The Nov 2025 USPTO guidance addresses this indirectly by saying direction is not enough.

Check yourself

  1. When does inventorship matter commercially, according to Abbott? Give his example.
  2. Why does he say naming the machine can still be useful if the machine does not care?
  3. Who would lose rights if the machine were the only inventor?

6. Ownership by accession and the court record

In short. No statute says that owning a machine means owning what it makes, but he says common law property rules point that way. He cites accession, which goes back to Roman law, with examples like the 3D printer, the tree and its fruit, and the cow and its calf. The record by late 2021 was mixed: refusals from the EPO, UK and USPTO, a grant in South Africa, and a pro-patent ruling in Australia.

🎙 In the talk. “There isn’t a statute that says if you own a machine that makes an invention you own that invention, but it is consistent with common law rules of property ownership”. Examples: “if I own a 3D printer that prints a beverage container for me I own that”, and “a principle of property ownership that dates back to Roman law and the concept of accession”. He adds the tree and fruit, cow and calf, and the Irish to each book its copy story. Justice Birss raised it in the Court of Appeal. Allens in Australia found “seven 18th century Blackstone references”.

On the record so far (late 2021): the EPO, UK and USPTO had refused. The refusal “would have been the path of least resistance”. South Africa granted; Australia’s federal court wrote a “41 page” decision that an AI-generated invention “should be patentable”. The UK Court of Appeal “split” and “we have now put in an appeal to the Supreme Court that is discretionary”. He describes waking at 3:00 a.m. for the Australian ruling. He says of the E.D. Va. judge “you couldn’t have had a judge tip much more” (his words: “you couldn’t have had a judge tip much more you know how she was feeling”).

📚 Deeper theory

  • Accession’s analogies break at two points (Roots). (1) Accession gives you the thing produced. A patent is not the produced thing; it is a state-granted right to exclude, which the statute gives to the inventor. (2) The 3D-printer example is about the physical container, not the design right in it. The analogy is better for ownership of a physical output than of an idea.
  • UK Supreme Court result. The court rejected accession as a basis for title (Part 1, press summary). The ground was that the doctrine does not reach the invention because the Act ties the right to an inventor who is a person (summary from the press summary; I did not read the full judgment).
  • The case timeline in the Part 1 table is the best single teaching device. Read it as a legal “forest plot”: different texts, same facts, nearly the same outcome. The two outliers are explained: South Africa checks forms only; Australia’s first-instance ruling was reversed.
  • What a win would have meant. Even a win on naming would not have decided validity. Examination on novelty and inventive step is separate. The project pre-checked that (idea 2).

📚 Roots: Property and accession — who owns what a machine makes?

The problem. Even if a machine cannot be an inventor, who holds title to the invention? Patent law’s default is “the inventor, then whoever the inventor assigns to”. If the inventor is a machine, the chain breaks.

  1. Accession in Roman law (Gaius, Justinian’s Institutes). The owner of a thing owns its natural increase: partus sequitur ventrem (the offspring follows the mother), fruit with the tree. Abbott uses this.
  2. Blackstone, Commentaries, Book II. The English statement of property by accession, which Allens’s team found “seven 18th century” references to.
  3. The limit. Accession was built for tangible things with natural increase. A patent right is a statutory, intangible right, created by registering. The UK Supreme Court held that accession does not give the owner of DABUS title to inventions, because there was no right to the invention to attach to when the inventor is not a person under the Act (press summary). The Australian first-instance court had accepted an owner-title analogy†.
  4. Employment and assignment law as the real alternative. Normally title comes from the inventor by assignment or by employment rule (UK s39, “employer owns” employee inventions; US assignment and “hired to invent”). A machine has no employment contract. So title must come from a statute that says “owner of the machine”. Abbott admits “there isn’t a statute that says” this.
  5. The copyright parallel. UK s9(3) deems the person “by whom the arrangements necessary for the creation of the work are undertaken” to be the author. That is a statute doing explicitly what accession does by analogy. Paul Samuelson’s 1986 article argued the same type of allocation for computer-generated works†.

The thread. Abbott separates who is the inventor (a naming question) from who owns (a title question). Courts largely merge them, because the statute links rights to the inventor. This is why his clean split did not carry.

Where others push back

  • Courts said ownership rules cannot create a right that the statute assigns to inventors. Property doctrine is a background, not a source of title to a statutory right.
  • Several commentators say the right fix for owner-title is a statute like UK s9(3), not common-law analogy.

What they leave out. Employment-style default rules for who owns AI-assisted inventions, and how licences (for example a model vendor’s terms) allocate rights to outputs. These matter more in practice than accession.

Check yourself

  1. What is accession, and what example does Abbott give?
  2. Why did the UK Supreme Court not accept it?
  3. Which of the Part 1 table’s outcomes does his “path of least resistance” remark predict?

7. “It isn’t really AI inventing”: three objections and a stunt charge

In short. Abbott hears “it isn’t really AI inventing” much less than he did two years ago. He answers three reasons for it. Being causally involved does not make someone an inventor. Inventive skill is poorly defined. And whether a machine understands what it does is irrelevant. What matters is that the rules create the right incentives, whether a person, an AI or monkeys on typewriters cure covid.

🎙 In the talk. A Wall Street Journal quote from a patent attorney doubts “we truly have AI yet where machines are doing the inventing”. Abbott says he hears this “a lot less now than I did two years ago”, and that industry now says “we’ve just spent a billion dollars on our new AI Division”. He gives three reasons the objection persists:

  1. Definitions. A European Commission report said AI was not inventing because a person was always causally involved. That is true, but “being causally involved in an invention is not a grounds of being an inventor under us or really any kind of patent law I’m familiar with”.
  2. Inventive skill undefined. If an AI outputs the best car suspension, he thinks “it’s fine that the first person to see that should be an inventor regardless of whether it involved inventive skill”, but “what it means to be an inventor… hasn’t been that well fleshed out”.
  3. No understanding. Some say an AI cannot invent because it lacks semantic understanding. He disagrees: what matters is that the system is “incentive creating the right sort of behavior”, and “whether I’m curing covid an AI is curing covid or room full of monkeys on typewriters is curing covid”. “whether a machine is thinking like a person or semantically understands what it does is totally irrelevant”. His Tesla comparison: “I don’t really care if it knows that it’s driving me in a philosophical sense”. On Turing: “we don’t really know what thinking is”.

On the stunt charge: some people “didn’t get it”; the press wrote that “machines were getting rights and AIS were going to own patents and of course that has never been part of this case”. He says he no longer hears the charge.

📚 Deeper theory

  • Two senses of “invent”. Causal: the system produced the idea. Legal: a person met the conception test. Abbott works to hold these apart. His critics argue the legal sense already includes agency and intent, so the two cannot be pulled apart.
  • The “monkeys on typewriters” line is about patent policy, not metaphysics. It is a standard utilitarian stance: only outputs and incentives matter (idea 1). It clashes with natural-rights theory, in which inventing is a person’s act.
  • Searle and Turing (Roots). Abbott picks Turing’s behaviour test. Searle-style objections are not refuted by it, only set aside as irrelevant. This works for patent policy but not for moral-rights arguments, which he also uses. There is a tension: if the machine “doesn’t care” about acknowledgement, the moral-rights reason does not support naming it.
  • The “publicity” question is a public-choice point. The press misread “listed as inventor” as “owns the patent”. Courts and offices dealt only with the first.
  • The timeline. His claim that sceptical patent lawyers had changed their minds is anecdotal. It is consistent with wide corporate AI use but is not evidence on the legal question, and the courts did not move.

📚 Roots: Philosophy of mind and legal personhood — “does it have to understand?”

The problem. Skeptics say a system that lacks understanding or intent cannot invent. Abbott says this is irrelevant. Two old debates sit under that exchange.

  1. Turing (1950). Replace “can machines think?” with the imitation game: if you cannot tell the difference by behaviour, the question loses its edge. Abbott’s use (“we don’t really know what thinking is”) is the behaviourist-functional move.
  2. Searle’s Chinese Room (1980). Syntax is not semantics: a system manipulating symbols may produce right answers without understanding them. This is the position of those who say AI “lacks semantic understanding”, which Abbott names. He rejects it as a standard for patents.
  3. Functionalism. What matters is the function a system performs, not what it is made of. Abbott’s analogy: it does not matter if a Tesla “knows” it is driving.
  4. Legal personhood is not mental personhood. Companies, ships (in admiralty), idols and in some places rivers are legal persons. Lawrence Solum’s “Legal Personhood for Artificial Intelligences” (1992) argues the test for legal status is a practical one†. Note a subtle point: Abbott does not ask for AI personhood. He says it was “never part of this case” that machines own patents.
  5. Why the patent system might not care. Patent law grants a right on the basis of the output (new, inventive, useful) and a human act of conceiving. It never tests the inventor’s mind except in the narrow sense of conception. A functional approach to mind fits this. But a moral rights theory of inventorship does not.

The thread. Abbott avoids the metaphysics by tying law to function: “whether a machine is thinking like a person or semantically understands what it does is totally irrelevant”. That is a legal-policy choice, and one that opponents answer with text and history, not with philosophy.

Where others push back

  • On “irrelevant”: the Federal Circuit ruled the text answers the question, and the UK Supreme Court said the question for the court was the meaning of the Act, not whether DABUS is creative.
  • On the stunt charge: even sympathetic commentators say the cases lacked facts that would force a hard choice, because humans were clearly involved in building DABUS.

What they leave out. Safety and responsibility. If a machine is the named inventor and the invention is harmful or the patent is later found invalid for misconduct, there is no one to hold to account. That is the human-centred “duty of candour” side of patents.

Check yourself

  1. What three reasons does Abbott give for the “it isn’t really inventing” view?
  2. Why does the “monkeys on typewriters” line fit incentive theory?
  3. How does his Tesla example relate to his general framework (see abbott-reasonable-robot-ai-and-the-law)?

8. Obviousness, hindsight and the skilled person

In short. The host asks whether obviousness will be argued in hindsight, with challengers saying anyone could have had an AI do it. Abbott is a bit concerned. He says hindsight bias is a real problem for the obviousness test. He suggests leaning more on economic factors, such as long-felt but unsolved needs and evidence of investment. The fuller argument is in his 2019 article.

🎙 In the talk. The host’s question: “arguments about obviousness are made in hindsight”. With AI, a challenger can say “anybody could have programmed this computer to look at these two types of technologies”. Is there a worrying period when firms pay for R&D, and afterwards others say they could have had an AI do it? Abbott: “I am I am a bit concerned”; “hindsight bias is really problematic for obviousness analysis” and the test “is already… more objective than any other part of the patentability analysis”. His fix: maybe “greater Reliance on… economic factors” such as “long-felt but unsolved needs” or evidence of investment, and a move “more explicitly to an economic incentive sort of analysis”. (Exact quote for the last phrase: “more explicitly to an economic incentive sort of analysis”.) The chapter title in the video details says “whether everything will be obvious in the future”; that framing is the description’s, not a quote from the talk.

The talk itself spends one question on this. The fuller version is in his 2019 article (Roots).

📚 Deeper theory

  • The host’s scenario is a special case of KSR. Once a tool can generate and test many candidates cheaply, the “obvious to try” door opens wider. A “finite number of identified, predictable solutions” (KSR wording)† can include the space the AI searches.
  • Two timelines. The inventor invested before the AI was cheap; the challenger judges after. US law requires the judgment to be made as of the filing date, with the skilled person of that date. So the “hindsight” worry is about future tools being imputed back. The statute guards against this in text, and the real risk is in how the judge reconstructs the past.
  • Which skilled person? Three options in the literature: (a) human, no AI (stays as now); (b) human with AI as a tool; (c) the AI itself (Abbott’s article). Option (b) is the most discussed in recent commentary (current debate, Part 1). It raises the level of skill and so the bar. Option (c) can lead to “everything is obvious”. Abbott accepted this possibility in the article, per the search results.
  • Why economic factors can help and why they can fail.
    • Help: long-felt need and failure of others are real-world evidence that did not depend on the argument’s hindsight. Investment shows effort and risk.
    • Fail: the courts require a nexus between the secondary evidence and the claimed features. Commercial success can come from marketing or other patents. And the AI-era problem is future cheapness, which these factors do not answer. They say “it was hard then”, not “it would be easy now”.
  • A harder structural point. If AI keeps lowering the cost of invention, the bar rises and the reward falls. Is that the right outcome? Under incentive theory, it may be fine: patents are needed less when inventions cost less. Abbott’s point is about transition fairness for those who paid before.
  • Evidence on offices. I found no guidance from the EPO, USPTO or UKIPO changing the skilled person for AI (Part 1). This is a search gap.

📚 Roots: Non-obviousness and the skilled person — the last question of the talk

The problem. A patent must not be obvious. But obviousness is judged by a fiction: a hypothetical person with ordinary skill. If AI is a standard research tool, what does that person know and do?

  1. Hotchkiss v. Greenwood (1851). The Supreme Court said an invention must show more “ingenuity and skill” than an ordinary mechanic would. This is the origin of the non-obviousness idea.
  2. Statute (35 USC 103, 1952). Obviousness is judged “before the effective filing date” from the viewpoint of a person having ordinary skill.
  3. Graham v. John Deere (1966). The four-step inquiry: scope and content of the prior art, the differences, the level of ordinary skill, and secondary considerations (commercial success, long-felt but unsolved need, failure of others, copying, praise). Abbott’s suggested fix for AI is to lean on exactly these: “long-felt but unsolved needs” and evidence of investment.
  4. KSR v. Teleflex (2007). The Court relaxed rigid tests: a skilled person is not an automaton, has “ordinary creativity”, and may find an obvious combination through common sense or “obvious to try” reasoning. The Court also warned against hindsight. After KSR, an examiner can say a combination was predictable. If a standard AI tool can enumerate and test combinations, “obvious to try” gets larger. That is the mechanism of the host’s worry.
  5. Hindsight bias. Psychology: Baruch Fischhoff (1975) showed people, once they know an outcome, overestimate how predictable it was. The law accepts the problem and tries to guard against it with the secondary considerations and the “before filing date” rule. Abbott says hindsight is “really problematic for obviousness analysis”.
  6. Europe and the UK. The EPO uses the problem-solution approach and a “skilled person” who is a skilled practitioner with common general knowledge. UK courts use the Windsurfing/Pozzoli steps†. The structure is similar. Secondary evidence matters less in the EPO than in US practice†.
  7. Abbott’s own proposal. In “Everything Is Obvious” (UCLA L. Rev. 2019) he argues that when inventive machines are standard in a field, the skilled person should be treated as an inventive machine, and that this may make nearly everything obvious over time. He also suggests that secondary factors or “reproducibility” tests avoid the difficulty of reasoning about what a machine would find obvious. The talk gives only the second idea, not the first. The first is in the article, per search results; I did not read the article in full.
  8. Opposing views. Some argue the skilled person should remain human and the inventive-step threshold should adjust instead; others argue AI is just a tool in the skilled person’s kit (see current debate). A hard question: if the skilled person uses the best available AI, and that AI can find the invention, is the invention then obvious? Economic evidence (investment, unmet need) can answer “no, because it took years and money”, but it cannot answer “but a cheap tool would now do it”.

The thread. The skilled person is a moving benchmark of what a field can do routinely. AI changes what is routine. Abbott’s two answers: move the benchmark (inventive machine standard), or lean on evidence that does not depend on the benchmark (economic factors). The talk shows the second.

Where others push back

  • Some say that the skilled person should remain human, and that when everything is easy, grant fewer patents; the system self-corrects.
  • Others say that a high bar for AI-found inventions is the correct outcome, since patent exclusivity then ends where tools make discovery cheap.
  • Some argue secondary considerations are too noisy and should not be central.

What they leave out. The disclosure link: if a patent must enable the invention, and a model can fill gaps in a disclosure, the enablement standard also shifts. Also, who bears the cost of proof that an AI would have found the answer: a challenger must show the specific tool, and the patent owner can argue it was not available.

Check yourself

  1. Why is hindsight bias a special risk if AI tools exist?
  2. Name two Graham secondary considerations and say what each tests.
  3. What are the three candidates for “the skilled person”, and which one does Abbott’s 2019 article choose?

9. The wider frame, briefly: AI stepping into human shoes

In short. He sees the project as one case in a wider question: how should law adapt when machines do things that people used to do. He points to policy interest in the EU, the UK and India. He also raises tax and liability, where the rules can create perverse incentives not to automate. His advice to young lawyers is to learn the technology well enough to know how the law should engage with it.

🎙 In the talk. The project is “a case in a much broader discussion about… the role of technology and the law” and asks “how should the law adapt as we move from a paradigm where people do things to people building machines to do things”. Policy attention rose: the European Commission’s draft AI rules, UK IPO consultations, an Indian parliamentary consultation “recommending the law be changed to protect this”. Near the end: tax (a university can replace a lecturer with a robot and avoid National Insurance contributions) and liability (a human Uber driver under negligence, a self-driving car under strict liability), which produces “perverse incentives for Uber to avoid automating”. On the book’s Chinese edition, a machine back-translation turned the title into “rational robots”. His advice to young lawyers: learn the technologies well enough “to understand how the law should engage with them”.

📚 Deeper theory. The principle behind this is Abbott’s “AI legal neutrality” (similar acts, similar treatment). Its full statement, the tax and tort cases, and the objections are in abbott-reasonable-robot-ai-and-the-law. Two links to this page’s topic: (1) in patents the neutrality claim says “treat an inventing machine like an inventing person” (idea 1, idea 7); (2) the courts’ reply is that neutrality cannot override statutory words about persons (Roots).

📚 Roots: Statute and interpretation — what is an “individual”?

The problem. If the Act says “individual” or “person”, can a machine fit? The three outcomes differ by method.

  1. Textualism. Ordinary meaning at enactment. Under the Dictionary Act (1 USC 1), “person” can include corporations, but “individual” has been read as a human. In Mohamad v. Palestinian Authority (2012) the Supreme Court read “individual” in the Torture Victim Protection Act as a natural person. The Federal Circuit used that reasoning in 2022 and added that its precedent already held that corporations and sovereigns cannot be inventors (Thaler v. Vidal). The court rejected Thaler’s policy arguments as beside the point once the text was clear.
  2. Purposive reading. What does the Act want? Abbott’s argument: the Act’s purpose is to encourage invention, “individual” is ambiguous, and a machine reading gives “all the right outcomes”. He points to “whoever” in §271 (infringement), which covers corporations and states, so “individual” could also stretch.
  3. Why his §271 argument is weaker than it sounds. “Whoever” is a broad pronoun in a liability section. “Individual” appears in a definitional section next to the inventors’ oath. Context matters, and this is the sort of context a textualist reads against him.
  4. Different statutes, different words. UK law says “person” and the UKSC read the 1977 Act as limiting inventors to natural persons, rejecting the argument that the Act’s purpose required more (press summary). The Australian first-instance judge went the other way, reading “inventor” as an agent noun (like “computer” or “dishwasher”) that does not itself demand a human†. The Full Court reversed.

The thread. In Abbott’s own words, the plain reading was “the path of least resistance”. The cases show it was also the winning one. Where a statute is written around human-centred words and rights (an oath, moral rights), courts hold that only the legislature can change them. This is the strongest reason the courts did not follow him, and he acknowledges it is why policy-makers matter.

Where others push back / What they leave out / Check yourself. See the sibling page. One question here: why is the patent case a harder test of neutrality than the Uber example?

Q&A

(Add questions and answers here only if worth keeping.)