AI's Impact on IP Laws: Mammen at PARC (2021)

About this talk

Speaker: Dr Christian Mammen (IP litigator, San Francisco); host probably Chris Colvin, founder of PARC and In the House
Event: PARC Featured Speakers series (monthly, third Friday), online seminar for lawyers
Date: 2021-10-15 (inferred, see below)
Length: 26:34
Channel: PARC — Raising the Bar
Watch: PARC Seminar — AI’s Impact on IP Laws with Chris Mammen

How to read this page. It has two kinds of content, always marked:

🎙 In the talk. What Mammen or the host say, with [MM:SS] timestamps. Only these parts are their views.
📚 Deeper theory. Background from the traditions behind the talk, plus what happened after 2021. It is not the speaker’s claim unless a timestamp says so.

The date is the main fact about this video. The upload date is 2 Nov 2021, but the talk was recorded earlier. The host says PARC holds its guest talk on the third Friday of each month [00:02], and Mammen says the UK Court of Appeal ruled “about three weeks ago” [18:39]. That ruling came in September 2021. So the likely date is Friday 15 October 2021 (inference, not stated). The talk is a snapshot taken in the middle of the first wave of DABUS cases, before any appeal court in the US, Australia or the UK had spoken. Read it as a time capsule, then score it with the table below.

Suggested route. Summary → Map → Scorecard → Part 2: The intellectual map → Deep dives → Putting it together.

See also. samuelson-generative-ai-meets-copyright covers the copyright side in depth (2023), with later case history.

Verification note. Links come from live web search on 30 Oct 2026. For each linked claim I fetched the page with curl and checked that the page contains the key terms (case name, date, holding). That shows the page discusses the claim; it does not mean I read every word, and I did not read the primary judgments. Where a page blocked scripts (HTTP 403) or I could not confirm, the claim says (search summary; source not confirmed) and carries no link. Everything else not linked, including every thinker, date and book in Part 2 and all legal doctrine background, is from memory; not checked. Case outcomes after 2021 are reported as law-firm commentary reports them. Nothing here is legal advice. This page was written by Claude, an Anthropic model; Anthropic appears in the training-data litigation material.

Speakers

Label	Who	Role in this talk	Confidence
Mammen	Dr Christian Mammen	IP litigator in San Francisco and Silicon Valley for about 20 to 25 years, with a background in legal philosophy; PARC member and the featured speaker. The video description calls him Dr Christian Mammen and uses “he”	High (self-introduced at [00:02]–[01:36]; description)
Host	Chris, founder of PARC (law partners) and In the House (in-house counsel)	Introduces, chairs, asks the one clarifying question at [13:58]	Name probable: one search summary names a Chris Colvin as the person behind those groups (search summary; source not confirmed). Captions say “chris coven”
Audience	Unseen	Q&A was held after the recording stopped	n/a

Caption warnings. Auto-captions. Corrections used on this page: “chris coven” = probably Colvin; “chris mammon” = Mammen; “park” = PARC; “dabos” / “dabus” = DABUS; “dr peter thaler” = Dr Stephen Thaler (Mammen’s own slip, see below); “sienna” = Siena; “brinkman” = Brinkema (the EDVA judge); “judge burst” = Birss LJ; “judge arnold” = Arnold LJ; “judge lang” = probably Laing LJ (the UK Court of Appeal had Arnold, Birss and Laing; from memory, and search summaries name Laing LJ and Arnold LJ in the majority with Birss dissenting); “owned by failure on these davos inventions” [18:39] = probably owned by Thaler on these DABUS inventions; “nextrembrand.com” = The Next Rembrandt project (from memory); “dictionary act” is the real name of the US statute (1 U.S.C. §1).

Factual slips and soft spots in the talk.

[10:54] / [12:27] Mammen calls him “dr peter thaler”. The inventor and applicant in the cases is Dr Stephen L. Thaler, as the Indian Copyright Office report and the US cert-denial summary both give.

[18:39] He puts the South African grant and the Australian ruling in “august”. The Australian first-instance decision is dated 30 July 2021, and the South African patent notice 28 July 2021 (the page confirms the 2021/03242 number and the lack of substantive examination; the exact journal date comes from a search summary; source not confirmed). So both were July, a few days apart, not a day apart in August. Small, but it matters for the story of DABUS gaining ground.

[15:30] He says the US, UK and European offices all said no in “the summer of 2019”. The UK and EPO refusals date from late 2019 and early 2020 (from memory; not checked), and the USPTO petition denial from 2020 (the US summary says the USPTO denied the application in May 2020). The sequence is right, the season is loose.

[13:58] Mammen says case law has held, uniformly and since the 1950s, that an inventor must be a natural person [12:27]. The Federal Circuit would later rely mainly on the word individual in the statute, not on those older cases, and said the text was clear (see Deep dive 2). Not an error, but the weight he gives old precedent differs from how the appeal court ended up reasoning.

[12:27] He says the Thaler applications are “the only two patent applications i’m aware of” claiming an AI alone invented. By his own wording this is a limit of his knowledge, not a fact about the world.

[20:13] The Australian court “made a whole bunch of assumptions”. That is his characterisation. The Full Federal Court later set aside the first-instance decision (see Scorecard).

Summary

Central question. Can an AI system be an inventor under patent law, or an author under copyright law, and if the law says no, why does it say no and who could change it?

Main thesis. Mammen sorts AI and inventions into three kinds: the AI is the invention, the AI helps a human inventor, and the AI is the inventor. He then sets policy aside and argues from “first principles”: in the US the Constitution speaks of authors and inventors, the Patent Act speaks of an individual, and every US court so far has read that as a natural person. The same wording and the same answer hold, so far, in the UK and Europe. Then he walks through the DABUS cases that had just come out: yes in South Africa and Australia (both for reasons he thinks are weak), no in the Eastern District of Virginia, and a split 2 to 1 “no” from the UK Court of Appeal on a tangle of formalities and the old doctrine of accession (owning the tree means owning the apples).

The close. Many more countries are still pending. He leaves open a back door: Thaler could refile, say he does not know who the inventor is, and claim the patent as owner of the machine. He gives the UK a “non-zero chance” of accepting that.

Stance to note. He presents himself as a litigator and a philosopher. He is cool towards the Australian reasoning, doubtful about the Thaler side’s coherence (they do not claim the AI has legal personhood), and treats the economic argument for protecting all innovation as one of three live arguments rather than his own. Copyright gets only an icebreaker. The one thing the talk almost never touches is the question that dominates 2026: training AI on other people’s work.

Map — talking points

#	Point	Time
1	Host intro. Mammen: litigator, legal-philosophy background, speech recognition cases	[00:02]–[01:36]
2	Icebreaker: AI art, a fake Rembrandt, who owns copyright, the monkey selfie	[01:36]–[06:14]
3	Three kinds of “AI invention”: AI as invention, AI as tool, AI as inventor	[04:41]–[07:48]
4	Where AI-as-inventor shows up: drug discovery, circuit design	[07:48]
5	First principles: Constitution Art. I s.8 cl.8, Patent Act “inventor”, “individual”, “person”	[07:48]–[09:20]
6	Hard side questions: programmers as inventors; dual standard of obviousness; AI as infringer	[09:20]–[10:54]
7	The Artificial Inventor Project, Thaler and Abbott, the fractal container and neural flame	[10:54]–[12:27]
8	US case law says inventor = natural person; corporation analogy	[12:27]–[13:58]
9	Host’s question: Supreme Court? constitutional or statutory? The Dictionary Act and the Venn diagram	[13:58]–[15:30]
10	”Individual” in other statutes; 2019 refusals by USPTO, UKIPO, EPO	[15:30]–[17:06]
11	Reasons for refusal; the Thaler side’s replies; economic and industrial-policy argument	[17:06]–[18:39]
12	Summer 2021: South Africa, Australia, Virginia, UK Court of Appeal	[18:39]–[20:13]
13	Why South Africa and Australia said yes; the Australian assumptions	[20:13]–[21:46]
14	Virginia ruling, appeal to the Federal Circuit; no claim of AI personhood	[21:46]–[23:19]
15	UK Court of Appeal: form, accession, Blackstone, the apple tree	[23:19]–[24:54]
16	A refiling route; 10 to 12 more countries pending; close	[24:54]–[26:27]

Scorecard: five years on

This is not in the talk. 📚 It scores Mammen’s statements and implicit bets against 30 Oct 2026. Outcomes are as reported by the linked pages unless marked.

Claim or bet in the talk	Time	2026 verdict
US law: an inventor must be a natural person, “pretty conclusively resolved”	[12:27]	Right. The Federal Circuit affirmed in 2022 and the Supreme Court denied certiorari on 24 April 2023 (No. 22-919). A cert denial is not an endorsement of the reasoning, but the rule stands
EDVA appeal “now on appeal to the federal circuit”	[21:46]	Affirmed (Thaler v. Vidal, 2022), on the word individual. Case reference at the cert-denial summary
Australia held an AI could be an inventor	[20:13]	Reversed. The Full Federal Court (13 April 2022, five judges) set aside the Beach J decision, and the High Court refused special leave on 11 November 2022
South Africa granted a patent; “they don’t examine the patents down there”	[20:13]	Right on the cause, and the grant stands as the outlier. Commentary says South Africa has no substantive examination and the Act does not define inventor (Mondaq page). The CIPC said it would review the Act (search summary; source not confirmed). I did not find whether the patent was later challenged
UK: “an inventor in the uk has to be a human”, but ambiguity and a refiling route with a “non-zero chance”	[18:39], [24:54]	The UK Supreme Court closed it. Judgment of 20 December 2023, [2023] UKSC 49 (page checked for the court and inventor point). A search summary reports a unanimous dismissal, no right to a patent from owning the machine, and accession rejected (search summary; source not confirmed). That kills the owner-of-the-machine route in the UK
EPO and national offices say no	[15:30]	Right. EPO Legal Board of Appeal, J 8/20 and J 9/20, December 2021: an inventor must be a natural person; the fallback request (no inventor named, applicant claims as owner of the machine) was also refused
Germany (not in the talk)	n/a	A workaround was allowed. A JUVE report shows the Federal Patent Court line; per a search summary the Federal Court of Justice (BGH, X ZB 5/22, 11 June 2024) accepted a form wording where the human applicant is named inventor and the AI is mentioned as having generated the invention (search summary; source not confirmed). The EPO’s T 528/25 abstract refers to the BGH decision and says wording of the form was the problem
”about 10 12 more countries where applications are still pending”	[24:54]	India decided. Reported: the Controller refused the DABUS application on 15 April 2026. Other countries: from memory; not checked
Copyright icebreaker: who owns an AI picture?	[04:41]	US answer given. D.C. Circuit, March 2025: human authorship required; cert denied 2 March 2026 (docket checked)
Can an AI be an infringer? “beyond the horizon”	[10:54]	Still beyond the horizon in law. The litigation moved to the developers and users instead (see Part 1)
Dual standard of obviousness	[09:20]	Open. No court or office I found adopted one. The USPTO’s November 2025 guidance says one legal standard for inventorship whether or not AI was used (Federal Register page checked for the rescission and the single standard)
Missing from the talk	n/a	Training-data litigation became the main event: Bartz, Kadrey, Ross, Getty, GEMA, NYT, ANI (Part 1)

Part 1 — Before you read

Key terms

Inventor vs applicant vs owner. ⚠ Three roles, often one person. The inventor is the person who conceived the invention. The applicant files. The owner holds the right. Under most systems the right starts with the inventor and passes by assignment or employment. Most DABUS fights are about how a machine’s “invention” gets from the machine to a human applicant.
Natural person vs legal person. A natural person is a human being. A legal person is anything the law treats as able to hold rights and duties: a company, sometimes a river or an idol. Mammen’s Venn diagram at [13:58] puts human inside individual inside person. Whether person includes companies, and individual includes only humans, is the knot.
AI as inventor / AI as tool / AI as invention. Mammen’s three categories [04:41]–[06:14]. Only the third is the DABUS question. ⚠ The middle one is the one most real inventions fall in, and the law treats it quite differently: the human is the inventor and the AI is equipment.
Conception. In US doctrine, the formation in the mind of the inventor of a definite and permanent idea of the complete invention. It gives the law a way to ask who thought of it. Machine-generated ideas strain it (from memory; not checked).
Non-obviousness / inventive step. The test that the invention was not obvious to a person having ordinary skill in the art (US) or a person skilled in the art (UK, EPO, India). Mammen asks whether this changes if the skilled person is a Watson [09:20].
Originality and authorship (copyright). Copyright needs a work that is original and made by an author. US law ties authorship to humans. India’s Act has a special clause for computer-generated works that names the person who causes the work to be created (Part 2, tradition 6).
Accession. A property rule: the owner of a thing owns what it produces or what is added to it. Mammen: “accession is the notion that if you own the apple tree then you also own the apples that grow from the tree” [23:19].
Dictionary Act. US statute (1 U.S.C. §1) that gives default meanings to words in other federal statutes, including person. Mammen says it is circular here [13:58].
Legal personhood. The status of being able to hold rights and duties. The Thaler side told the courts it does not claim it for DABUS [21:46].
Training data / fair use / fair dealing. Not in the talk. The doctrines that decide whether copying works to train a model needs the owner’s permission. US fair use is an open-ended four-factor defence; Indian fair dealing (Copyright Act s.52) is a list of purposes. Terms differ by country, so a US ruling on training does not move India directly.

People and cases named (reading list)

Name	Why it matters here
Dr Stephen Thaler; DABUS (Device for the Autonomous Bootstrapping of Unified Sentience)	The applicant and his system. Mammen says “peter” by mistake. Names the AI as inventor on a fractal-shaped container and a flashing-light beacon (the neural flame)
Ryan Abbott (Surrey) and the Artificial Inventor Project	Lawyer-led campaign that filed the same two applications in a dozen or more countries [10:54]–[12:27]. Abbott’s writing on whether AI-made inventions should be patentable: from memory; not checked
Judge Brinkema (E.D. Va.)	Ruled against Thaler in September 2021, as the Mondaq summary reports (page checked for the court and the natural-person holding)
Beach J (Federal Court of Australia)	Held for Thaler at first instance on 30 July 2021, reversed on appeal
Birss, Arnold, Laing LJJ	UK Court of Appeal. Birss dissented, as Osborne Clarke reports
Blackstone, Commentaries (1765–69)	The authority on accession that the UK judges read [23:19] (from memory; not checked)
Naruto v. Slater (the monkey selfie case)	Raised at [04:41] as a copyright analogy: could a non-human own copyright. Holding: from memory; not checked (a US appeals court held animals lack statutory standing under the Copyright Act, 2018)
The Next Rembrandt (2016)	The AI Rembrandt portrait Mammen shows [04:41] (from memory; not checked)

The speaker’s own frame

Mammen is a litigator who also studied legal philosophy [01:36]. That gives the talk its method: first principles over policy, statutory text before purpose, and a taste for the odd concept (accession, personhood, the oath that an AI cannot swear). He reasons like a lawyer deciding a case, and explicitly resists “jump to the last question” of whether AI inventors should be allowed [07:48]. His audience is lawyers, so there is little explanation of what an AI system is.

The current debate (live search, 30 Oct 2026)

This is not in the talk. 📚

Patent side: settled for now, workable in practice.

United States. Thaler v. Vidal stands. On 28 Nov 2025 the USPTO rescinded its February 2024 AI-assisted inventorship guidance and replaced it, saying the Pannu joint-inventorship factors apply only among multiple natural persons and that AI systems are not persons. Firm summaries describe a single standard for all inventions (search summary; source not confirmed on the single standard phrasing, though the notice page confirms the rescission and the Pannu point). In practice: name the humans who conceived the invention, treat the AI as a tool.
UK. UKSC 2023: a machine cannot be an inventor, and owning the machine gives no right. Per summaries, the court left open whether AI-generated inventions are patentable at all (search summary; source not confirmed).
Europe. EPO J 8/20 (2021) and, in February 2026, T 528/25, which (per the abstract page checked) is about the form wording under Art. 81 EPC and R. 19 EPC not being a clear and unambiguous designation of the inventor. A search summary reads its headnote as allowing inventor designation in principle for AI-assisted inventions (search summary; source not confirmed).
Australia. Reversed on appeal; High Court refused special leave.
India. See the India note below.

Copyright side: output.

US. D.C. Circuit 2025, cert denied 2 March 2026. The Copyright Office page confirms Part 2 (Copyrightability, published 29 Jan 2025) and a pre-publication Part 3 (Generative AI Training, 9 May 2025). The Part 2 summary I found says wholly AI-made output is not protected even from detailed prompts, while human selection, arrangement or modification can be (search summary; source not confirmed). A related challenge, Allen v. Perlmutter (a Midjourney image, 600+ prompts), was a pending case in Colorado as of the last report I found; I did not find a ruling.

The bigger fight: training data. Not in Mammen’s talk at all.

Bartz v. Anthropic. On 20 July 2026 the court gave final approval to a $1.5 billion class settlement (page checked). The earlier June 2025 ruling treated training on lawfully obtained books as fair use, but not building a library from pirated copies (Kilpatrick Townsend, page checked). Anthropic is the author of this page’s model; read this item as a party’s case.
Kadrey v. Meta. June 2025, Judge Chhabria, found for Meta on that record while saying a market-dilution theory could win in other cases, according to the Manatt analysis (page checked for the terms Alsup, Chhabria and dilution).
Thomson Reuters v. Ross. The Third Circuit affirmed that Ross’s training use was not fair use (page checked; a decision unsealed 30 Sep 2026; the commentator thinks the reach is limited because the system was a non-generative competitor).
NYT v. OpenAI and Microsoft. Still in discovery and pleading changes; one report says the Times narrowed its claims on 25 June 2026 (page checked; secondary source).
UK and Germany. Getty v. Stability AI: the UK High Court (Nov 2025) rejected the secondary-infringement claim because the model is not an infringing copy (page checked; appeal on that point reported). GEMA v. OpenAI: the Munich court (Nov 2025) found infringement on song lyrics (page checked for Munich and lyrics; details from a search summary).

India: a short, labelled note. 📚

Authorship. The Copyright Act, 1957 s.2(d)(vi) names, for a computer-generated literary, dramatic, musical or artistic work, the person who causes the work to be created. A King Stubb & Kasiva report (21 Sep 2026) describes the Registrar of Copyrights refusing Thaler’s application for the DABUS artwork A Recent Entrance to Paradise: the work met the low originality threshold under s.13, but DABUS cannot be the author, and the person who caused the work was the human who built, configured and fed the system. The Registrar said legislation, not interpretation, would be needed to give an AI authorship. Thaler declined to name himself. This is an administrative order on one application, not a High Court ruling.
Inventorship. Per Intepat, the Indian Patent Office refused application 202017019068 on 15 April 2026 on the ground that AI is not the true and first inventor under the Patents Act, 1970 (page checked for the date, number, wording and s.15). Summaries add a separate inventive-step objection under s.2(1)(ja) and a ruling against a pre-grant opponent’s separate order (search summary; source not confirmed). A Controller’s order is not a court judgment.
Training. ANI v. OpenAI: on 24 July 2026 Justice Amit Bansal of the Delhi High Court refused ANI an interim injunction, on a prima facie view that training falls within fair dealing under s.52(1)(a), per a Mondaq summary (page checked). ANI’s appeal: a LiveLaw report of 15 Sep 2026 says a division bench issued notice. Next hearing date (8 December) is from a search summary; source not confirmed. No final ruling that I found.
Policy. DPIIT’s committee released Part I of a working paper on generative AI and copyright on 8 December 2025 proposing a hybrid statutory-licensing model (page checked for Hybrid Model and the date). Authorship of outputs was left to Part 2. A Bar & Bench report of a Lok Sabha answer says that issue is expected in Part 2, still under review. I did not find Part 2.
For a company that files patent provisionals and publishes AI-assisted video and text. From the above, and not legal advice: (1) For inventions, the working position in India, the US, the UK and the EPO is to name the human who contributed to conception; an AI cannot be named. How much human contribution is enough is the open edge. (2) For content, the Indian clause points at the person who causes the work to be created, but who that is (prompter, tool-builder, commissioner) is contested; contracts that assign rights between the people involved are the only dependable cover. (3) Protection for the AI-generated portion itself is uncertain in the US (human authorship needed) and untested at appellate level in India. Take advice from Indian counsel before relying on any of this.

Part 2 — The intellectual map

The talk looks like a case report but rests on eight traditions. Mammen names few of them: Blackstone, the Constitution, the philosophy of mind in passing. The rest he leaves implicit. For each I give the problem, the chain of thinkers with dates, how it feeds the talk, and then the core argument in some depth. All chains and dates are from memory; not checked unless linked. Where I am unsure of a name or year I say so and keep the wording general.

1. Natural-rights and personality theories of property: why anyone owns what they make

Problem. Why should a creator own anything at all? If an idea is free to copy, why give someone a right to stop others? Before you can ask whether a machine can have such a right, you have to know what the right is for. The talk’s authority is the constitutional clause about authors and inventors, which already assumes a person who makes things and earns something by doing so.

Chain.

John Locke, Second Treatise of Government (1689): a person owns their own body and so their own labour, and mixing labour with unowned things makes them property, with two famous provisos (enough and as good left for others; no spoilage).
Immanuel Kant, “On the Wrongfulness of Unauthorised Publication of Books” (1785): a book is a speech act, an address from author to public. The author’s right is a right in the speech, which cannot be sold away like a physical copy.
G. W. F. Hegel, Philosophy of Right (1820): property is the way a will gives itself external existence. Things that carry a person’s will, such as a letter or a design, are part of their personality.
Romantic authorship, 18th–19th century: the author as a singular genius. Scholars of copyright history, such as Martha Woodmansee and Peter Jaszi, argue the idea was built up alongside copyright law rather than preceding it (from memory; not checked).
Burrow-Giles Lithographic Co. v. Sarony (US Supreme Court, 1884): the photographer is the author because the work originates in his mental conception. The 2025 D.C. Circuit decision rested on this line (search summary; source not confirmed).
Margaret Radin, “Property and Personhood” (1982), and Justin Hughes, “The Philosophy of Intellectual Property” (1988): modern statements of the personality and labour justifications for IP (from memory; not checked).

How it feeds the talk. It supplies the unspoken premise of the Constitution’s clause and of the economic and “human flourishing” language at [18:39]. When Mammen says IP could be viewed as “policy aimed at enhancing human flourishing” he is reaching for this family of ideas, and he names the rival (industrial policy) on the other side.

The core argument. Locke’s labour theory has four moving parts. First, a self-ownership premise: I own myself. Second, labour is an extension of the self, so what I mix my labour with, I partly extend myself into. Third, the proviso: appropriation is legitimate if enough and as good is left for others. Fourth, the right is a natural right, antecedent to the state, which the state only recognises. Applied to a machine, the first premise fails at once. A machine does not own itself in any sense Locke would accept, and so it has no labour to mix. A defender of machine inventors has two replies. One: the labour belongs to whoever made the machine, so the programmer or owner stands in for the self. That is a reconstruction of Locke through the doctrine of accession (tradition 4), and the UK courts rejected it. Two: drop the labour theory and say rights serve social ends (tradition 2).

Hegel’s personality theory is the cleaner basis for the human-only rule. If a work or invention matters because it embodies the will and character of its maker, then the right protects the maker’s personhood, and an object cannot have a personhood to protect. The theory also explains why authors in civil-law systems often have moral rights, such as attribution and integrity, which cannot be sold. It struggles with a different set of facts: the corporate employer who owns the patent for something its engineers built. There, personality theory goes quiet, and the case for ownership rests on contract and incentive. This is why the debate about machine inventors tends to switch theories mid-argument.

Kant’s account gives the copyright version. A book is the author’s address to the public, and an address needs an addresser. A model’s output is addressed to nobody by anybody, which is part of why the Copyright Office and courts speak of human authorship as essential, not as a technicality (from memory; not checked, and the extension to AI is my gloss).

Where it breaks. These theories were built for the artisan and the writer. They give no clean answer for large, collective, tool-heavy R&D, which is where most patents come from. A drug company’s patent is not an extension of an individual’s personality. That is where the second tradition takes over.

2. Utilitarian incentive theory and industrial policy: why law grants monopolies at all

Problem. Information is non-rival: once made, anyone can use it at almost no cost, so makers may under-invest unless they can exclude copiers. The question is what rule produces the most invention at the lowest cost to users. A machine needs no reward, so what does an incentive rule say about machine inventors?

Chain.

Statute of Monopolies (England, 1624) and Statute of Anne (1710): early statutory monopolies for new manufactures and for authors (from memory; not checked).
Jeremy Bentham, early 19th century: ridiculed natural-law claims; defended patents only as a reward for useful invention (from memory; not checked).
John Stuart Mill, Principles of Political Economy (1848): defended temporary patents on utilitarian grounds, objecting to a reward funded by a levy on users (from memory; not checked).
Fritz Machlup, An Economic Review of the Patent System (1958): argued the evidence for patents’ net benefit was too thin to prove or disprove them (from memory; not checked).
William Landes and Richard Posner, The Economic Structure of Intellectual Property Law (2003): the standard law-and-economics treatment, with an optimal term and scope balancing incentive against access (from memory; not checked).
Kenneth Arrow (1962) and Joseph Schumpeter (mid-20th c.): information as public good; innovation under competition (from memory; not checked).
Industrial-policy tradition: IP as a lever for national development, from 19th-century catching-up states to the TRIPS bargain (1994) and modern debates in South Africa, India and others.

How it feeds the talk. Two of the four reasons Mammen lists come from here. The first is the “economic theory” argument at [18:39]: “it’s really a matter of uh economic competitiveness in the in the global economy”, and so protect whatever is innovative however it came about. The second is the South African explanation at [20:13], where he says there was “a lot of discussion” about changing patent law to serve industrial policy. The incentive argument also underlies the D.C. Circuit’s copyright reasoning that machines do not respond to ownership incentives (search summary; source not confirmed).

The core argument. Start with the trade-off. A patent raises the price of an invention to users (deadweight loss) in return for a stream of rewards that makes inventors invest. Society gains when the extra invention is worth more than that cost. The key condition is that the right must change behaviour. A right that rewards activity that would have happened anyway is pure cost.

Now apply it to machines. If DABUS cannot respond to incentives, giving it, or its owner, a right changes nothing about DABUS. The owner might respond: a person who funds machine R&D might invest more if the output is patentable, whether or not the machine is named inventor. This is the pro-Thaler incentive case, and note that it does not need the machine to be the inventor; it needs the output to be patentable and the owner to be able to claim it. That is why the DABUS litigation attacks the naming rule rather than patentability. Courts that refuse to name the machine can still reach the same incentive outcome by allowing the owner to be the applicant, which is exactly what the German workaround and the “non-zero chance” refiling route at [24:54] try to do.

The opposing incentive argument is that a flood of machine-made patents could raise the cost to every user and crowd out human work, so that extending protection to AI output lowers welfare. It also cuts on the skilled-person side (tradition 5): if everything becomes obvious to a machine, fewer patents issue at all.

Industrial policy adds a different logic: IP is a lever, not a neutral prize. A country may reasonably decide to offer or withhold protection for a new class of output to attract investment. That is the best reading of the South African outcome, though the page I checked also stresses a formalities-only system. It also explains why the question is open in new places. The Delhi DPIIT paper’s licensing proposal for training data is a good example of policy-first design. It starts from the question of what bargain between AI developers and rights holders is best for the country, not from first principles about authorship.

Where it breaks. The empirical base is thin (Machlup’s point survives). Patents may matter for pharmaceuticals and little for software. And an incentive story offers no help with the moral intuition, shared by many, that an invention should be credited to those who made it.

3. Legal personhood: who or what can hold a right

Problem. Law distinguishes who may hold rights and duties from what may only be owned. The vocabulary (person, individual, human being, natural person, juridical person) carries the whole argument at [09:20] and [13:58]. If the Constitution and statute speak of inventors and individuals, who is in the class?

Chain.

Roman law: persona and the distinction between persons and things, with slaves treated as things (from memory; not checked).
William Blackstone, Commentaries (1765–69): natural persons and “artificial persons” created by law, such as corporations (from memory; not checked).
Corporate personhood in US law, 19th century, including Santa Clara County v. Southern Pacific (1886) (from memory; not checked).
Wesley Hohfeld (1913–17): rights and duties analysed as jural relations. Personhood is a bundle of capacities, not a single yes-or-no property.
Christopher Stone, “Should Trees Have Standing?” (1972): argued for guardian-held rights for natural objects.
Recent examples: some jurisdictions have recognised rivers and similar entities as legal persons; Indian courts have treated certain idols as juristic persons (from memory; not checked).
The EU Parliament’s 2017 resolution on robotics floated electronic personality for sophisticated robots, and drew heavy criticism in an open letter by experts (from memory; not checked). Joanna Bryson and colleagues, “Of, For, and By the People: The Legal Lacuna of Synthetic Persons” (2017), warned of shell-company liability dodges (from memory; not checked).

How it feeds the talk. The whole discussion at [09:20]–[15:30]: individual, person, the two Dictionary Act definitions, the Venn diagram, the corporation analogy at [12:27], the question of whether the AI whether it fits within the legal meanings of these terms. It also explains the pointed observation at [21:46] that the Thaler side disclaims AI personhood.

The core argument. Legal personhood is a tool of legal bookkeeping, not a metaphysical status. It exists so that someone can be the subject of rules: sued, taxed, required to sign. A company is made a person for these reasons. A natural person comes with it, as default. Two things follow.

First, the bookkeeping view is why Mammen’s Venn diagram is only “a hypothesis”. [13:58] The inner circles are not guaranteed: statutes can use individual more or less widely. The host’s question about the Supreme Court and Mammen’s answer show it: the Supreme Court has read individual as human-only in some statutes, a circuit court has done so in another, and other courts have given the word “a broader interpretation” in contexts like bankruptcy and crime-victim and computer-damage statutes [15:30]. His line is exact: “it’s not out of the question that individual could be defined somewhat more broadly than than natural persons” [15:30]. The interpretation depends on the statute’s purpose, which is why the answer in the Patent Act could be fixed by the Act’s own oath and declaration provisions.

Second, the bookkeeping view says what an AI would need. To hold a patent, an entity must be able to sign an oath, assign rights, be sued over validity, pay fees, and bear consequences. The oath problem is the one Mammen leans on: “how does the ai sign under penalty of perjury” [17:06]. A system that cannot be punished or bound cannot hold the role. This does not show that no AI could ever hold a right; it shows that the current administrative structure has no way to attach duties to one. A legal-personality proposal for AI would have to build that structure, typically by tying the AI to a human guardian or owner. Which is the same as naming the owner, and so brings the argument round to tradition 4.

The reverse risk, noted in the academic writing above, is that personhood for machines becomes a way for human principals to shed responsibility. That is one reason policy bodies were cool on electronic personality, and one reason Thaler’s team stayed away from it.

Where it breaks. Personhood-as-bookkeeping makes the ordinary-meaning argument look flimsy: if the category is a legislative choice, then the courts’ plain meaning of individual hides a judgement. Brinkema reportedly said the question was for Congress (search summary; source not confirmed). The move is institutional: let the legislature decide.

4. Acquisition and accession: how rights pass from a thing to its owner

Problem. Even if DABUS cannot be an inventor, perhaps its owner has the right to what it makes. The law of property has several ways to give an owner rights in new things: accession (what a thing produces), specification (making a new thing from another’s material), employment (what an employee makes), and contract. The issue is whether those rules, built for tangible things, extend to intangible inventions.

Chain.

Roman law (Gaius, Justinian’s Institutes): rules of accessio (fruits and additions go to the owner of the principal thing) and specificatio (the maker who transforms materials may acquire the new thing) (from memory; not checked).
Blackstone (1765–69): property by accession, including the increase of animals and the produce of trees; the classic apple-tree illustration. Mammen says this goes “all the way back to blackstone” [23:19].
Locke (1689), again: the labour theory as a rule of acquisition. Compare tradition 1.
Employer rules: US work-made-for-hire, the UK rules on employee inventions (Patents Act 1977 s.39), India’s Copyright Act rules on first ownership (ss.17–19). All from memory; not checked.
The UK Court of Appeal in 2021 and the UK Supreme Court in 2023 on accession and s.7 of the Patents Act 1977: see below.

How it feeds the talk. The UK Court of Appeal portion, [23:19]–[24:54]: Thaler had not become the inventor, so how could he be the applicant? His answer was that he owned DABUS and so, by accession, owned what it produced. Mammen: “these judges in the uk court of appeal judges were citing and discussing in-depth blackstone 17th century jurisprudence”. He tells the audience the three judges split on whether “an invention that’s output from a computer algorithm is the same thing as an apple that grows on a tree” [23:19].

The core argument. The accession rule, in its simple form, says the owner of an object owns its natural increase. It works for fruit, young animals and rents because the increase is a physical, identifiable thing that appears on the owner’s land, from the owner’s property, without anyone else’s labour mixed in. Four features carry the rule: (1) the product is physical or at least separable; (2) the source is identifiable and owned; (3) the product arises naturally from the source; and (4) nobody else’s rights compete.

An invention is different on each count. It is intangible, it can be copied by anyone who learns of it, and the right to a patent is a creature of statute, not property found in nature. So the question is whether the statute makes the right derive from inventorship only, or whether other routes (such as ownership of a machine) can lead to it. In the UK that is a question of the construction of s.7 of the 1977 Act, which lists who may apply for a patent: the inventor, the person entitled by an agreement or by rule of law before the invention, and successors in title. In the UK Supreme Court’s reported reasoning (search summary; source not confirmed), accession cannot create a right the statute does not confer, and ownership of the machine does not make you the person entitled.

The three-way split below the Supreme Court shows how close the tangible-intangible analogy was perceived to be. On the report Mammen gives, one judge said the invention is like an apple, one said no, and the third “went along with” the no side while seeing the issue differently. [23:19], [24:54] Birss LJ (dissenting) is reported to have given the purpose of the application provisions more weight (search summary; source not confirmed on the wording).

The deeper tension is with the labour theory. Accession treats the machine as a productive asset and the output as its fruit; but the same accession logic would give the machine owner the fruit even if a different human operator did all the prompting. The more the human operator matters, the weaker accession becomes. That is why the live debate has shifted to human contribution (tradition 5, and the USPTO guidance). Where there is a human who conceived, the question is about that human’s role, not about the machine’s ownership.

Where it breaks. Accession is a doctrine for property in things. It could be rescued by contract, the way employers acquire employee inventions. This is the practical advice in the India note: in the absence of a statutory route, put assignments in writing among the humans involved.

5. Patent doctrine: invention, conception and the skilled person

Problem. A patent requires an invention that is new, non-obvious and useful. Each criterion assumes a human reference point. Conception assumes a mind. Non-obviousness assumes a skilled reader. If machines do the conceiving and the reading, the reference points move.

Chain.

Statute of Monopolies (1624): the “manner of new manufacture” (from memory; not checked).
Hotchkiss v. Greenwood (US, 1851): an improvement must show more than the skill of an ordinary mechanic. Origin of non-obviousness (from memory; not checked).
Patent Act 1952 (US), s.103: obviousness to “a person having ordinary skill in the art”. Graham v. John Deere (1966) set out the factual inquiries; KSR v. Teleflex (2007) loosened rigid tests (from memory; not checked).
Conception as the test of inventorship: Burroughs Wellcome v. Barr Labs (Fed. Cir., 1994) and Pannu v. Iolab (Fed. Cir., 1998), the latter setting the factors for joint inventors (from memory; the Federal Register page above confirms the Pannu reference).
European approach: the EPO’s problem-and-solution approach to inventive step, with the notional skilled person; UK Windsurfing and Pozzoli steps (from memory; not checked).
India: Patents Act 1970, s.2(1)(ja) defines inventive step as a technical advance or economic significance that makes the invention non-obvious to a person skilled in the art; s.2(1)(y) defines true and first inventor; s.6 says who may apply; s.10 requires disclosure (the Intepat page above confirms the s.2(1)(ja), s.6, s.7 and s.10 references as reported).
Ryan Abbott, The Reasonable Robot (2020) and the article “Everything Is Obvious” (2018) argued that as machine capability grows, the skilled person should be treated as a machine and nearly everything becomes obvious (from memory; not checked).

How it feeds the talk. The side questions at [09:20]–[10:54]: “what’s not obvious to to watson does this change the level of obviousness”, a dual standard of obviousness for human and machine inventions, using algorithms to “connect the dots and show that everything is obvious”, and whether an AI could be an infringer. Also the list of what the offices said at [17:06] about the oath of inventor and the theory of mind behind invention.

The core argument. The person skilled in the art is a legal fiction: a normal, unimaginative practitioner who has the common general knowledge of the field and ordinary tools. The test for obviousness asks whether this person, faced with the problem, would have arrived at the claimed solution without inventive effort. The standard works because the fiction is stable. It is anchored to what human practitioners of that time and field routinely know and do.

AI disturbs that anchor in three ways. First, the knowledge of the skilled person: if standard practice includes running a model over the literature, then what counts as common general knowledge expands to include whatever such a model would readily surface. Second, the effort: routine experiment is not inventive. If machines make exhaustive search cheap, many results that once took inventive insight now look like routine search, so the bar for a human to clear rises. Third, the inventor’s side: if the machine generates a candidate that no human would have thought of, the inventive step is the machine’s, and the human’s contribution may be only recognising or selecting it. Is recognition enough for conception?

Mammen’s phrase “dual standard” names an obvious but costly way out: judge machine-made inventions by what is obvious to a machine, and human-made by what is obvious to a human. The problem is administrative: you would need to know how an invention was made, and applicants would conceal the answer. The more practical route, now visible in the USPTO guidance, is to avoid the question by insisting that only humans are inventors and the usual tests apply; AI use is then a fact about how the human worked. In that setting the doctrinal questions become sharper: did the human conceive, or only run the tool? The USPTO’s Federal Register notice reads the situation as a single-inventor case in which the Pannu factors do not apply because AI systems are not persons and so there is no joint-inventorship question Federal Register notice. The conception question remains, and the page does not answer it.

For India the same structure applies, with a statutory definition of inventive step that is already expressed in terms of technical advance or economic significance. The Controller’s reported additional objection under s.2(1)(ja) in the DABUS case (search summary; source not confirmed) suggests an examiner can reject on inventive step even when the inventorship question is the headline.

Where it breaks. The skilled-person fiction was never fully empirical; it is a legal judgement dressed as a fact. So the proposals to update it (Abbott’s) are in effect a call to change how patents are granted in a machine-saturated world, not a discovery of what is obvious. The counterargument is that raising the bar sharply would end patents in fields where machines dominate, which is a policy choice for legislatures.

6. Copyright doctrine: authorship, originality and the idea-expression line

Problem. Copyright protects original works of authorship. What counts as authorship when a human supplies a prompt, a selection or a curated dataset and the system supplies the detail? And what is protectable at all: ideas, or only the particular expression?

Chain.

Statute of Anne (1710); the US Constitution’s copyright and patent clause (1787), the one Mammen reads at [07:48]: “authors and inventors” (the article and clause numbers are from the talk).
Baker v. Selden (US, 1879): the idea-expression distinction. A system described in a book is not protected by the copyright in the book (from memory; not checked).
Burrow-Giles v. Sarony (1884): the photograph as the author’s intellectual conception; Feist v. Rural Telephone (1991): originality needs independent creation and a modicum of creativity (from memory; not checked).
Berne Convention (1886 onward): authorship attributed to natural persons in most readings; moral rights (from memory; not checked).
The UK’s Copyright, Designs and Patents Act 1988 s.9(3): for a computer-generated work, the author is the person who made the arrangements necessary for its creation (from memory; not checked).
India: Copyright Act 1957, as amended; s.2(d)(vi) uses the phrase the person who causes the work to be created for computer-generated works (confirmed as the text the Registrar applied, per the King Stubb & Kasiva report); s.13 originality; ss.17–19 first ownership; s.52(1)(a) fair dealing (the ANI summary confirms s.52(1)(a)).
Cases on non-human authors: Naruto v. Slater (the monkey selfie), raised by Mammen at [04:41]; Thaler v. Perlmutter (D.C. Cir. 2025).
US Copyright Office registration practice from 2023: works with AI-generated parts are registrable only for the human-authored parts, per Zarya of the Dawn and later guidance (from memory; not checked). The Part 2 report’s summary is in Part 1.

How it feeds the talk. The icebreaker at [03:08]–[04:41] is the copyright half: he shows images, asks the audience to guess which is AI-made, and sets the three questions “can this picture be copyrighted uh if so who is the artist and who should own the copyright” [04:41]. He then explicitly sets copyright aside: “i’m not gonna focus that much today on copyright law” [01:36] because patents were the field moving fastest.

The core argument. Copyright’s authorship rule has two layers. The first is originality: a work must originate with the author and show a minimal creative spark. The Indian Registrar reportedly found an AI picture could meet that low bar. The second is who is the author, a question about the source of the creative choices. US law answers: a human being. The reasoning is textual (several provisions of the Copyright Act make sense only for a human author, per the D.C. Circuit as reported) and structural (the incentive logic of tradition 2).

The idea-expression line is the reason this does not wipe out all AI-assisted creation. An idea, however novel, is not protected; the specific expression is. When a human gives a prompt, the prompt is a description of an idea; the model’s particular rendering is the expression. The US Copyright Office’s Part 2 summary says that prompts alone do not give enough control over the expression (search summary; source not confirmed). The line cuts differently for works where the human arranges, selects, edits or transforms the output: there the human contributes expression, and protection can attach to that contribution. The 600-prompt Midjourney image in Allen v. Perlmutter tests how much iteration changes this, and I found no final ruling.

Compare three legislative models. UK (s.9(3)): a rule that vests authorship of computer-generated works in the arranger, made in 1988 for tools like early software, not generative AI; government consultation on whether to keep it has taken place (from memory; not checked). India (s.2(d)(vi)): the person who causes the work to be created, so AI is never an author but a human always is, at least on the face of the clause. The test is wide open: who caused? The Registrar reportedly looked at who built, configured and fed the system; commentary says a prompting user may also qualify (search summary; source not confirmed). US: no author means no copyright for purely machine-made output, so no one owns it, and it falls to the public domain.

There is a policy twist. The US rule does not give any right to the user’s output unless a human contribution is separable, but the developers rely on contracts and terms of service to assign what rights exist. That is how the market fills the gap (from memory; not checked).

Where it breaks. The idea-expression line is a vague tool, and control over expression is a sliding scale. It also does not answer the training problem, where the question is not who owns the output but whether copying the inputs needed a licence. That is the subject of tradition 7’s institutional discussion and of Part 1’s current-debate list.

7. Interpretation and legal philosophy: how a court decides what a word means

Problem. Mammen says his background blends IP litigation with legal philosophy [01:36]. The philosophy is not a decoration. The central move in the talk, “go all the way back to first principles and see what we’re limited in talking about” [07:48], is a position on how law should be reasoned about. Courts must say what a statutory word means when the drafters never imagined the case.

Chain.

Natural law and positivism: Aquinas and the natural-law tradition; Jeremy Bentham and John Austin on law as command; H. L. A. Hart, The Concept of Law (1961), with open texture in rules (from memory; not checked).
Lon Fuller and Hart, the debate on positivism and morality (1958) (from memory; not checked).
Ronald Dworkin, Law’s Empire (1986): law as interpretive; judges should read statutes in the light that makes the best sense of the whole practice (from memory; not checked).
Textualism and ordinary-meaning canons in the US, popularised by Antonin Scalia and Bryan Garner, Reading Law (2012) (from memory; not checked).
Dictionary and “ordinary meaning” practice, and the canon against surplusage; the Dictionary Act’s default definitions as a statutory aid.
Purposive interpretation in UK and Commonwealth courts; the “always speaking” principle of construing statutes to meet new technology (from memory; not checked).
Separation of powers: courts interpret, legislatures decide policy.

How it feeds the talk. Everything in [07:48]–[15:30] and the court summaries after. Mammen deliberately separates the question whether the existing law can be read to include an AI inventor from “should it”. He reports that the Eastern District of Virginia held “because we’ve got such a clear legal doctrine we don’t even have to talk about any of these policy issues” [21:46]; that is the textualist posture. The Australian court read inventor in its ordinary sense, undefined in the Act, and found nothing barring an AI. The UK Court of Appeal fell into the formalities and accession. Each court is making an interpretive choice.

The core argument. Start with Hart’s insight. Any rule has a core of clear cases and a penumbra of doubtful ones. Vehicles in the park clearly bans cars and clearly does not ban a baby’s pram; skateboards sit in the penumbra. Inventor has a core (a human who thought of the thing) and a penumbra (a team, a company, a machine). In 1952 nobody had to decide the penumbra, because nothing was in it. The question for the court in 2021 is how to decide a penumbra case when the legislature never chose.

Three styles answer differently. A textualist says: use the ordinary meaning at the time of enactment, check how the statute uses the word elsewhere, and leave extension to the legislature. Individual means a human, and the oath and declaration provisions of the Act presuppose a human. This is the Federal Circuit’s route, as reported, and it is why the US courts could say that policy is not their job. A purposivist says: ask what the provision is for, and read it so as to serve that aim. If the aim of patent law is to encourage disclosure of useful inventions (tradition 2), then an AI-generated invention should be capable of patent. This is Birss LJ’s reported dissent, and the thrust of the Australian first-instance reasoning. A Dworkinian says: pick the reading that fits the whole practice and puts it in its best light. The practice of patent law is built around human creative effort and the reward of persons; so the best reading keeps inventors human. All three can be defended, which is why courts split.

A related issue is whether the question is constitutional or statutory. The host’s question at [13:58] gets to it. In the US the Constitution’s clause is a grant of power to Congress, and the Patent Act implements it. Mammen’s answer is that the cases so far construe the statute, which means Congress could change the answer by amending the Act without touching the Constitution’s wording about inventors, though the clause might constrain how far it could go (my reading; from memory; not checked).

The speaker’s choice of first principles also reflects the lawyer’s habit of reasoning from the narrowest ground. Courts resolve a case on the statute if they can and avoid larger questions. That is why the Virginia court did not need policy, the High Court of Australia declined to take a broad appeal, and the UK Supreme Court’s reported scope limit left patentability of AI-generated inventions open.

Where it breaks. Textualism is not neutral: choosing the ordinary meaning at enactment favours the past. Purposivism can read into the statute whatever the judge thinks it ought to do. And the clean line between interpretation and policy is itself a legal-philosophy position, disputed since Hart and Dworkin.

8. Mind, agency and machine creativity: what inventing is

Problem. The patent offices, Mammen reports, said that “the active invention is really requires a theory of mind and only humans have minds” [17:06]. Does inventing require a mind? Can an algorithm that searches a space and flags novelty be said to invent? The Thaler side says DABUS decided; the courts say DABUS is a tool.

Chain.

Ada Lovelace, notes on Babbage’s Analytical Engine (1843): the engine cannot originate anything, only do what it is ordered to do (paraphrase from memory; not checked).
Alan Turing, “Computing Machinery and Intelligence” (1950): answers the Lovelace objection by arguing machines can surprise their makers.
John Searle, “Minds, Brains, and Programs” (1980): the Chinese Room. Syntax alone does not yield understanding.
Margaret Boden, The Creative Mind (1990): three kinds of creativity: combinational, exploratory, transformational. She argues computers can in principle do all three (from memory; not checked).
Herbert Simon and Allen Newell: early programs that discover (the Logic Theorist, 1956); later discovery systems in chemistry and mathematics (from memory; not checked).
Genetic programming and evolutionary design, 1990s–: machine-designed antennas and circuits (from memory; not checked).
AlphaFold and AI drug discovery, 2020s: the practical outcomes behind the drug and circuit examples at [07:48].
Philosophy of action: Elizabeth Anscombe, Intention (1957) and Donald Davidson (1963) on what makes an event an action done for a reason.

How it feeds the talk. The “theory of mind” objection at [17:06]; the three-category scheme at [04:41]–[06:14] (is the AI a tool, or does it rise to co-inventor); the Australian court’s assumptions at [20:13], which he lists as “the ai sets sets its own goals”, that it has “freedom of choice”, that it chooses the input data and is autonomous. His comment that these go “well beyond” what Thaler and Abbott would claim [20:13] is a point about agency: the court attributed more autonomy to DABUS than its owners did.

The core argument. The debate turns on what the word invent requires. There are three positions.

Output view. To invent is to produce something new, useful and non-obvious. If DABUS produced the container design and nobody before had it, DABUS invented. This is the view behind the Thaler campaign. It fits economics (tradition 2) and sits uneasily with law, which looks at the process as well as the product.

Process view. To invent is to conceive: to form an idea of a solution to a problem, with a grasp of the problem and the criteria. This is the law’s view, and it explains the American reliance on conception. A machine can search a space, but the problem was set by the humans, the evaluation function by the humans, the training data by the humans. Where does the invention sit in that chain? Mammen’s list of the Australian assumptions maps onto this: the court assumed DABUS set the goal, chose the options and chose the data. If a human did any of those, the claim of machine-only invention weakens, and the argument becomes one about degree, not kind. Because of that, the line between AI as “tool” and AI as “inventor” in Mammen’s categories is a continuum, not a border.

Mind view. To invent is an act of a thinking subject: you need understanding, intention and, in the patent offices’ phrase, a theory of mind. This is Searle’s and Anscombe’s side. On this view no existing system invents, whatever it outputs. This view is why the offices could refuse DABUS without deciding whether the outputs were good.

Boden’s taxonomy matters because it suggests where the machine gets close. Combinational creativity (unusual combinations of familiar ideas) and exploratory creativity (searching a defined space) are things search algorithms do well; transformational creativity (changing the space itself) is harder to credit to a program. DABUS-style outputs, such as a container shaped by fractal geometry, look combinational. That is the best case for the law’s cautious position: even on a generous view of machine creativity, the transformational kind that patent law most prizes remains the human’s.

The oath problem in the law, which Mammen raises at [17:06], is the action-theory problem in a legal form. To swear under penalty of perjury that one is the inventor, one must be able to own a claim: to assert it, to answer for it. That is intention and responsibility, and it is what the philosophy of action says marks a doer from a mere cause.

Where it breaks. The mind view proves too much: if conception requires a thinking subject, the law has to examine human inventors’ minds, and it does not, except crudely. The process view has a practical problem: patent offices cannot see how an invention was made. That is why disclosure of AI use, and the human-contribution tests in guidance, have become the main enforcement tools.

How they meet

(1) Natural rights / personality ──┐
   why anyone owns what they make │
                                    ├──> Who is the INVENTOR / AUTHOR? ──> (3) Legal personhood
(2) Incentives + industrial policy ─┘        │                                  "individual", "person",
   what rule yields most invention         │                                  the oath, the oath-taker
                                              v                                       │
(8) Mind and agency ────────────> Does the machine INVENT? ───────────────────────────┤
   "theory of mind", Boden                                                            │
                                                                                       v
(5) Patent doctrine ──> conception, skilled person ──┐                    (4) Accession: can the OWNER
   "everything is obvious?"                          │                         take what the machine makes?
(6) Copyright doctrine ──> authorship, expression ───┤                                  │
                                                     v                                  v
                          (7) Interpretation: text, purpose, or leave it to the legislature
                                                     │
                                                     v
                         Outcomes: no (US, UK, EPO, Australia, India) / formal workaround (Germany)
                                  / yes only by a formalities-only office (South Africa)

Two loops matter. Person-loop: natural-rights theory (1), personhood (3) and the oath (8) all point to a human inventor; incentive theory (2) alone could favour machine ownership, and accession (4) is the route that converts that into a claim. Human-contribution loop: because the law will not name a machine, the live work is on how much a human must contribute (5, 6, 8) — the thread of the 2024–2026 guidance and of the copyright prompt cases.

Part 3 — Deep dives

1. Three kinds of “AI invention”

🎙 In the talk.

Three categories: the AI is the invention, the AI helps a human inventor, and the AI is the inventor [04:41]–[06:14].
On the second: sometimes “there there are lab assistants who maybe don’t contribute enough to the research process to themselves be named an inventor” [06:14].
Where this arises: “drug discovery” and “design new circuits that are different than the ones that humans have have previously designed” [07:48].
The method: “it’s helpful to go all the way back to first principles” [07:48].

📚 Deeper theory.

The lab-assistant analogy comes from joint-inventorship law: a person who only follows instructions or supplies standard equipment is not an inventor. See Part 2, tradition 5 (conception, Pannu). The USPTO’s 2025 notice carries this over to AI: the tool is not a joint inventor.
The first category (AI as the invention) is a different problem again: patent eligibility of software and algorithms (in the US, the Alice framework; in India, s.3(k) and the 2025 CRI guidelines for computer-related inventions, which the Intepat page mentions only in passing). From memory; not checked.
In 2021, category two was the commercially important one. In 2026 it is overwhelmingly so: nearly every filing involves some AI assistance.

Where others push back. Some argue that the three-part scheme is a false trichotomy. Autonomy is a spectrum, and that is where the real disputes lie (Part 2, tradition 8).

What they leave out. Trade secrets. Many AI-generated outputs are protected, if at all, by secrecy and contract rather than by patent or copyright. Not discussed, and not asked.

Check yourself.

Which of the three categories do the DABUS cases concern, and which category covers most real filings?
What is the difference between an inventor and a lab assistant under joint-inventorship law?
Why might the same company be in category one for its product and category two for its R&D?

2. “Inventor” in the Patent Act: individual, person, human

🎙 In the talk.

“so constitutionally we can only give patents whether it’s an inventor” [07:48] (the caption garbles; his point is that the Constitution speaks of authors and inventors).
The Act “defines an inventor as the individual who invented” and the question is “does it fit within the legal meanings of these terms” [09:20].
The case law: “an inventor under u.s patent law has to be a natural person a human being” [12:27].
The host asks whether the Supreme Court has addressed it. Mammen’s answer is the Venn diagram: “the big circle is person uh within that is individual and within that is human” [13:58].
But “it’s not out of the question that individual could be defined somewhat more broadly than than natural persons” [15:30].

📚 Deeper theory.

See Part 2, traditions 3 and 7. The Federal Circuit later gave the statute-first answer. The cert-denial summary says the court treated the text as clear and did not need to consider metaphysical questions about invention or AI rights. A search summary says it relied on the Supreme Court’s ordinary meaning of individual as a human being (search summary; source not confirmed).
Comparison across systems: the EPO Board of Appeal held in J 8/20 that an inventor must be a natural person under the European Patent Convention, and refused the fallback of leaving the inventor blank and claiming as owner of the machine (EPO communique). The UK Supreme Court (2023) and the Indian Controller (2026, reported) reached the same result by their own statutes.
The Dictionary Act point: Mammen says its two definitions “seem to be a little bit circular” [13:58]. The Federal Circuit did not need them.

Where others push back. Abbott and the Artificial Inventor Project argued that the statute does not say only human, that inventorship rules came before AI, and that rewarding the output serves the Act’s purposes (the arguments are as Mammen gives them at [17:06]–[18:39]). Some scholars argue a purposive reading should apply (from memory; not checked).

What they leave out. What happens when the answer is partly human. The talk treats the machine as sole inventor, which is rare. The 2025 guidance and the German workaround show that practice has moved to the human-contribution question.

Check yourself.

Why did the Federal Circuit not need the Dictionary Act?
What does it mean that a cert denial leaves the rule in place but not the reasoning approved?
Which words in the Patent Act would have to change for an AI to be named inventor, and who has power to change them?

3. Why the offices said no, and the replies

🎙 In the talk.

The refusals: USPTO, UK and European offices said no to the fractal container and the neural flame [15:30].
Reasons: “statutes require that an inventor be human case law requires it” and “in the u.s you have to submit an oath of inventor that’s that’s sworn under penalty of perjury” [17:06].
“how does the ai sign under penalty of perjury” [17:06].
Transfer of rights: “if thaler is filing the application how did thaler get his rights from this ai” [17:06].
The mind point: “only humans have minds” [17:06].
The replies: “if thaler put his own name down on there that would be fraudulent” [17:06]; and the economic argument, “it’s really a matter of uh economic competitiveness in the in the global economy” [18:39].

📚 Deeper theory.

The four official reasons map onto four traditions: statutory text (7), formalities and personhood (3), title transfer (4), mind (8). The reply from fraud and the economic reply map onto policy (2).
The fraud argument is a trap in form: naming a human who did not invent is false, and naming the machine is refused. The way out the courts offered was transparency about what happened. The German BGH workaround and the EPO’s T 528/25 discussion are about drafting the form so it tells the truth: a human is named as the inventor, and the AI’s role is declared (search summary on the BGH; source not confirmed; the EPO abstract page confirms the form-wording point).
The economic argument reappears in 2026 India: DPIIT’s training-data proposal is explicitly policy design.

Where others push back. Critics of the human-only rule say it forces dishonest practice: applicants will name a human for work a machine did. Critics of the Thaler side say the campaign was designed to lose, to generate a legislative debate (a view held by some commentators; from memory; not checked).

What they leave out. Enforcement of truthful inventorship. The offices cannot tell which humans did what. That is a problem in 2021 and bigger in 2026.

Check yourself.

Why is the oath a good argument and a bad one?
What happens to the transfer-of-rights problem if the applicant is named owner and the inventor is left unknown?
Which of the replies survives into the 2026 guidance, and which does not?

4. South Africa and Australia: the two yeses

🎙 In the talk.

“the patent office of south africa agreed to issue a patent” and the Australian court “likewise held that those applications should be deemed patent-eligible” [18:39].
South Africa: “inventor was not defined by statute as being a person or an individual” and “they don’t examine the patents down there” [20:13].
Australia is “a little more puzzling because inventor is defined as a person there” yet the court “held it wasn’t prohibited” [20:13] .
The assumptions he lists: “the ai sets sets its own goals” [20:13], and he says they go “well beyond what thaler and abbott uh would even claim” [20:13].

📚 Deeper theory.

South Africa: see the Mondaq page. The explanation Mammen gives (no examination, no definition) is the commentary’s too. A search summary adds that the patent number is 2021/03242 and that critics argued a substantive-examination system would have rejected it (search summary; source not confirmed). The grant is a fact about a registration system, not a judicial holding.
Australia: the first-instance decision reasoned that inventor is an ordinary word and an AI can be one (see the Mondaq page). The Full Federal Court reversed. The High Court refused special leave. Commentators note the refusal was narrow and did not say a human is always required (search summary; source not confirmed).
Tradition 8 explains Mammen’s scepticism: the court’s list of assumptions about autonomy was a factual finding about DABUS that the owners themselves did not press.

Where others push back. South Africa’s defenders point to its innovation policy environment (search summary; source not confirmed). Australian supporters of the first-instance reasoning say the Act’s object clause favours a broad reading.

What they leave out. What happened to the South African patent afterwards, which I could not find. Also that in a formalities-only system the grant carries little legal weight until tested.

Check yourself.

Why does the absence of examination weaken the South African precedent?
What facts about DABUS did the Australian first-instance court assume?
Why would a High Court refusal of special leave be weaker than a ruling on the merits?

5. Virginia and the appeal

🎙 In the talk.

“judge brinkman in the eastern district of virginia ruling on the previous u.s patent office ruling affirmed that no uh there shouldn’t be ai inventor patent eligibility in the u.s” [18:39] (the caption spells the judge Brinkman; the judge is Brinkema).
Reasoning: “inventors must be humans under the case law under the definition of individual” and “because we’ve got such a clear legal doctrine we don’t even have to talk about any of these policy issues” [21:46].
“that case is now on appeal to the federal circuit” [21:46].
A tension: “they don’t contend that the their ai has legal personhood” [21:46], which he finds “a little bit puzzling” [21:46].

📚 Deeper theory.

The ruling is reported in Mondaq (page checked for Brinkema’s court and the natural-person holding). The Federal Circuit affirmed (Thaler v. Vidal, 2022), and certiorari was denied on 24 April 2023.
Mammen’s puzzle is an incentive-theory and personhood puzzle at once. If DABUS is not a legal person it cannot be an individual in the Act, and the Thaler side cannot get rights through it; if it were claimed as a person, the claim would invite the problems in Part 2, tradition 3. The Thaler position is best read as “inventor” being a status that does not need personhood, which the courts rejected.
The Federal Circuit’s choice to rest on text is an instance of the narrow-ground habit in tradition 7.

Where others push back. Some commentators say individual in the Patent Act was meant only to distinguish inventors from companies, and so the statute’s silence on machines leaves room (from memory; not checked).

What they leave out. The Copyright Office parallel, which went the same way in the D.C. Circuit in 2025 and ended at the Supreme Court’s cert denial in March 2026.

Check yourself.

What does it mean to rely on the plain meaning of a statutory word?
What is the tension in a claim for inventor status without personhood?
Why might courts prefer not to discuss policy?

6. The UK Court of Appeal and the apple tree

🎙 In the talk.

The structure: “you’re required to fill out a form the applicant’s required to fill out a form that identifies the inventor” [21:46]; the slot “left blank you can still get a patent theoretically” [21:46].
The result: Thaler named DABUS in the slot, “that was wrong and so his application was canceled” [23:19].
Accession: “accession is the notion that if you own the apple tree then you also own the apples that grow from the tree” [23:19].
The split: “one judge said no one judge said yes” [23:19] and the third “went along with the judges and said no they’re different” [23:19].
The refiling idea: “i own the computer system so i should get the patent” [24:54]; “non-zero chance that that kind of approach could work in the uk” [24:54].

📚 Deeper theory.

See Part 2, tradition 4. The Court of Appeal decision is [2021] EWCA Civ 1374 and Birss LJ dissented, as Osborne Clarke reports (page checked).
The UK Supreme Court’s 2023 judgment: William Fry (page checked for the court and the inventor point). The panel, the lead judgment by Lord Kitchin, the rejection of accession and the point that the decision does not concern the wider issue of whether AI-generated inventions are patentable come from a search summary (search summary; source not confirmed).
Mammen’s refiling route is the very one the EPO rejected as an auxiliary request, as the EPO communique shows. The UK Supreme Court’s reported reasoning that ownership of a machine gives no right closes it.

Where others push back. Birss LJ (dissenting) is reported to have read the provision more purposively. Supporters of accession note it worked for the Australian first-instance judge, who held Thaler could derive title that way (per the Australian appeal summary).

What they leave out. That the UK had, and still has, a separate rule for computer-generated copyright works (s.9(3), from memory), which shows the UK legislature chose a different answer for copyright than for patents.

Check yourself.

Why did naming DABUS in the form sink the application?
What would you need to show to say that an invention is like an apple?
Why is the refiling route weaker after the Supreme Court?

7. Training data: the question the talk does not ask

🎙 In the talk.

Nothing on training data. The only gesture is the Rembrandt: the algorithm took “all of rembrandt’s photos” and “used those to generate a picture” [04:41] (his word is “photos”).
He mentions in passing that he is litigating AI cases “in the areas of speech recognition” [01:36].

📚 Deeper theory.

Since 2021 the main AI-and-IP disputes moved to inputs. The US fair-use rulings in 2025, the $1.5 billion Bartz settlement, the Third Circuit’s Ross decision, the UK Getty ruling that a model is not an infringing copy, the Munich GEMA ruling on lyrics, and India’s ANI order are all in Part 1.
The doctrinal tools are different. Output questions turn on authorship (tradition 6) and personhood (3). Input questions turn on copying, fair use or fair dealing, and market harm, a mix of tradition 2 (incentives and substitution) and tradition 6 (the idea-expression line).
The two sets of questions meet in one practical place: the output of a model trained on works can be denied protection (no human author) and still infringe (if it reproduces protected expression). Both can be true at once.

Where others push back. Developers argue training is transformative, an internal process that does not substitute for the originals. Rights holders argue it is mass copying that dilutes markets. The Chhabria and Alsup rulings, as reported, disagree about dilution.

What they leave out. Everything. Mammen’s talk is dated before the generative image and text boom of 2022 and later.

Check yourself.

Why can an output be unprotectable and still infringing?
Which four fair-use factors does market dilution relate to?
Why does a US ruling on training not directly control India?

Putting it together

Mental model. Think of the DABUS question as a gate with three locks. The first lock is who can be a person in the statute: the answer so far is humans, because the statutes were written around oaths, signatures and minds. The second lock is how rights travel: accession, employment and contract could carry a right from a machine to its owner, but the courts held the statute does not let them. The third lock is whether the law cares about process or only product: the policy arguments for protecting machine output are strong, but courts leave them to legislatures. What changed after 2021 is that the argument moved from the gate to the room behind it. The question is no longer whether a machine can be named, but how much a named human must have done, and whether the training of the machine needed permission.

Open tensions.

If AI-assisted invention is routine, what makes a human’s contribution enough for conception?
If purely machine-made output has no owner, who benefits: the public, or whoever controls access to the machine?
If everything becomes obvious to a machine, does the skilled-person fiction need a new anchor?
Should the answer for copyright and patent be the same? India’s clause and the UK’s s.9(3) say copyright can differ; patents have not.
If an inventor’s declaration is only a form, how do offices detect a human named for work a machine did?
Does training on works need a licence, a levy (DPIIT’s hybrid model) or nothing?

Reading path.

Start: the talk itself (26 minutes); the cert-denial summary for the US outcome; the Indian Copyright Office report.
Then: the EPO communique on J 8/20; the USPTO notice of November 2025; the DPIIT working paper summary.
Deeper: Locke, Second Treatise ch. V; Hegel, Philosophy of Right on property; Hart, The Concept of Law; Boden, The Creative Mind; Abbott, The Reasonable Robot (all from memory; not checked); the primary judgments, which I did not read: Thaler v. Vidal (Fed. Cir. 2022), [2023] UKSC 49, J 8/20, Thaler v. Perlmutter (D.C. Cir. 2025).

Talks

Explorer

AI's Impact on IP Laws: Mammen at PARC (2021)

Speakers

Summary

Map — talking points

Scorecard: five years on

Part 1 — Before you read

Key terms

People and cases named (reading list)

The speaker’s own frame

The current debate (live search, 30 Oct 2026)

Part 2 — The intellectual map

1. Natural-rights and personality theories of property: why anyone owns what they make

2. Utilitarian incentive theory and industrial policy: why law grants monopolies at all

3. Legal personhood: who or what can hold a right

4. Acquisition and accession: how rights pass from a thing to its owner

5. Patent doctrine: invention, conception and the skilled person

6. Copyright doctrine: authorship, originality and the idea-expression line

7. Interpretation and legal philosophy: how a court decides what a word means

8. Mind, agency and machine creativity: what inventing is

How they meet

Part 3 — Deep dives

1. Three kinds of “AI invention”

2. “Inventor” in the Patent Act: individual, person, human

3. Why the offices said no, and the replies

4. South Africa and Australia: the two yeses

5. Virginia and the appeal

6. The UK Court of Appeal and the apple tree

7. Training data: the question the talk does not ask

Putting it together

Graph View

Table of Contents

Backlinks