The Reasonable Robot — Ryan Abbott on AI and the Law

About this talk

Speaker: Ryan Abbott
Event: Center for Data Innovation book talk series, “The Reasonable Robot: Artificial Intelligence and the Law” (webinar, Fri 13 Nov 2020, 12:00 EDT)
Date: 2020-11-13
Length: 59:30
Channel: DataInnovation
Watch: Book Talk on ‘The Reasonable Robot: AI and the Law’ with Ryan Abbott

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

• 🎙 In the talk. What Abbott or the host says, with [MM:SS] timestamps. Only these parts are his views. Quotation marks mean his exact words from the captions.
• 📚 Deeper theory / added context. Background from the traditions he draws on, and later events. It explains why he argues as he does and who disagrees. It is not his claim, and facts dated after Nov 2020 are not his.

Suggested route. Summary, then Map, then Key terms, then Part 2: The intellectual map (read once, slowly), then the Deep dives in order, then “Putting it together”.

See also. Same author, other talks: abbott-future-of-ai-human-vs-machine (2021 broad interview) and abbott-patent-law-ai-inventors (2024 patent interview, the depth page for DABUS and inventorship). This page is the home for his general framework: AI legal neutrality, tort, criminal law, tax and personhood.

---

Speakers

Label	Who	Role
Host	Daniel Castro (named by himself [00:03])	Director, Center for Data Innovation (a think tank on data, technology and public policy). Introduces, then moderates the Q&A
Abbott	Ryan Abbott (named by the host)	Professor of law and health sciences, University of Surrey School of Law; adjunct assistant professor of medicine, UCLA. Author of The Reasonable Robot (2020). Main speaker [00:03]
Audience	Questions submitted on Slido, read out by the host	Topics: patent as economic tool, robot tax, inventorship, who is liable if the creator dies, averaging humans, China

Confidence: speaker names are stated in the talk itself, so high.

Caption warnings. Auto-captions, lower case, no punctuation. Corrections used on this page:

• “davis” / “davis” = DABUS (the AI in the test case) [20:09]
• “burro giles b cerrone” = Burrow-Giles Lithographic Co. v. Sarony [18:34]
• “heliograph” = Heliograf, the Washington Post’s article-writing system [17:02]
• “port law”, “torque reasons” = tort law, “tort reasons” [37:16]
• “covet”, “covet vaccine” = COVID [24:48]
• “a person’s field of the art” = a person skilled in the art [26:21]
• “ad researchers” = probably average researchers [24:48]
• “slide it” = Slido; “adventure”, “invention” in the inventor answers = inventor [49:37]–[51:09]
• “nai”, “ami” = AI [32:38]
• “oral b cross action toothbrush” is an example, not a name slip: Oral-B CrossAction [20:09]
• “cryptic” at [34:10] cannot be matched to a known chess engine (see slips below)
• “we robot” at [43:27] = the We Robot conference on robot law
• “monkeys” at [18:34] refers to the monkey-selfie case (Naruto v. Slater), not named in the talk

Factual slips by the speaker (verified):

1. [18:34] He says the photograph case was “1988 1886”. The case, Burrow-Giles v. Sarony, was decided in 1884 (111 U.S. 53). From memory; not checked against a primary source.
2. [32:38] “in 2007 nai beat the world’s grand master at the board game go”. The famous result is AlphaGo over Lee Sedol, March 2016, 4-1 (source). He himself says at [34:10] that 2017 was “the year that alphabet’s deep mind ai beat the world go champion” (Ke Jie, 2017, from memory; not checked). So “2007” is a slip or a caption error.
3. [34:10] “deep blue … beat gary kasparov in 1998”. The deciding rematch was May 1997 (Wikipedia).
4. [34:10] He says Kasparov, “the next year”, “placed first in the first human ai centaur chess tournament”. Search confirmed only that Kasparov introduced advanced chess in 1998. I could not confirm that he won a tournament; from memory, the 1998 event was an exhibition match against Veselin Topalov. Treat the claim as unverified and probably wrong in detail; the broader point (human plus machine beat either alone for a time) is well attested.
5. [34:10] “in 2017 … the chess engine cryptic beat the world’s best human plus ai team”. Could not identify any engine called “cryptic”. Unverified.
6. [10:50] “human error accounts for about 94 percent of auto accidents”. The number is real but is widely misread. See deep dive 3 for the NHTSA caveat. This is added context, not a slip.
7. [06:12] “about 90 percent of federal tax revenue comes … from labor taxes … less than 10 comes from corporate taxes”. From memory (not checked): US federal receipts in recent years were roughly half individual income tax, about a third payroll tax and about a tenth corporate tax. Individual income tax is not all labour income. So “about 90” is loose, but the direction is right.

---

Summary

Abbott’s central question: when machines do what only people used to do, should the law treat the machine’s behaviour differently from the person’s? His answer is mostly no. The default should be AI legal neutrality: do less to discriminate in law between AI and human behaviour, because current differences produce bad social outcomes [03:07].

He is clear about what the principle is not. It does not say AI deserves rights. It is one principle among several for AI regulation (he names transparency, fairness and accountability as others) [01:34]. It is about behaviour and outcomes, not about the moral status of machines.

He then runs the principle through four fields:

1. Tax. Businesses pay payroll tax on workers but not on machines, so tax pushes firms to automate, and revenue falls when they do. Fix: neutral treatment, mainly by cutting taxes on labour and taxing capital more, not by taxing robots literally [06:12]–[09:18].
2. Tort. A human driver faces a negligence test (“would a reasonable person have done the same?”). A self-driving car, as a product, faces strict liability. That discourages safer technology. Fix: judge both by negligence. Later, once AI is much safer, judge humans by a “reasonable robot” standard [10:50]–[15:28].
3. Intellectual property. Copyright and patents assume a human creator. Abbott leads a test case listing an AI as inventor, and argues that AI-owner-owned patents serve the incentive purpose of patent law. He also argues that “person skilled in the art” will need to become “machine skilled in the art” [17:02]–[26:21].
4. Criminal law. The hardest fit, because criminal law cares about mental states and moral blame. He concludes that punishing a machine “probably isn’t a good idea”, though it is less foreign to doctrine than it sounds, because we already punish corporations [27:54]–[32:38].

The Q&A adds: explainability should depend on context (essential in sentencing, less so in driving); the “reasonable robot” standard needs wide availability and affordability; personhood should be granted only if it produces better outcomes for people; and regulation designed for AI helps rather than hurts industry [37:16]–[58:53].

Stance to note: he is a pragmatic consequentialist with a plain preference for outcome-based tests over process-based ones. The host says he disagrees with some positions but values that the book takes “concrete positions” [35:42].

---

Map — talking points

#	Point	Time
1	Host intro; the book’s principle in one line	[00:03]
2	Thesis: AI legal neutrality. Not about AI rights; corporations as the model of “artificial persons”	[01:34]–[03:07]
3	Tax I: automation is not new; history of jobs; two roles of tax	[03:07]–[06:12]
4	Tax II: payroll tax gap; robots do not pay taxes; revenue share from labour	[06:12]–[07:44]
5	Tax fix: not a literal robot tax; cut labour taxes, tax capital more	[07:44]–[09:18]
6	Tort I: self-driving cars; are AI drivers safer than human drivers?	[10:50]
7	Tort II: negligence vs strict product liability; focus on behaviour, not why	[10:50]–[12:23]
8	Tort III: “reasonable robot” standard applied to humans	[13:54]–[15:28]
9	Surgical robots; doctors and lawyers	[15:57]
10	Creative robots: Washington Post, Christie’s sale; copyright needs a human author; Burrow-Giles	[16:50]–[19:11]
11	The DABUS test case; neural networks; Siemens anecdote; patent rejections	[20:05]–[22:38]
12	Why patents on AI output; why list the AI; AI owner owns the patent	[22:38]–[24:48]
13	”Person skilled in the art” becomes “machine skilled in the art”; AGI aside	[25:10]–[27:54]
14	Criminal law: mental states, retribution; corporate criminal liability	[27:54]–[31:05]
15	Punishing an AI: benefits, costs, verdict	[31:05]–[32:38]
16	Board games: Go, Kasparov, centaurs, “people may just get in the way”	[32:38]–[34:10]
17	Host’s reaction; Q: is strict liability wrong in general?	[34:10]–[37:16]
18	A: product liability is right in general; outcomes over explanations; trolley problem	[37:16]–[40:22]
19	EU vs US on explainability; context dependence (sentencing vs driving)	[38:50]
20	Q: how to define the “reasonable robot” across price points; global standards	[40:22]–[43:27]
21	Q: cruelty to robots; Kate Darling; animals analogy	[43:27]–[45:00]
22	Q: AI personhood, Sophia, EU, Japan, India	[45:00]–[48:05]
23	Q: patents as an economic tool	[48:05]
24	Q: should robots be a “protected class”? Robot tax again	[49:37]
25	Q: inventorship in the US, UK, EU over five years; “person” in the UK Act	[49:37]–[52:40]
26	Q: how can a human contribute to an AI invention	[52:40]–[54:13]
27	Q: AI commits a crime after its creator dies	[54:13]
28	Q: judging machines against the “average human”	[55:44]
29	Q: China, regulation and industrial development	[57:18]–[58:53]

---

Part 1 — Before you read

Key terms

• AI legal neutrality. Abbott’s principle: the law should not treat AI and human behaviour differently without a good reason, because differences tend to produce worse social outcomes. It is a policy principle about behaviour. It is silent on whether machines have moral status.
• Artificial person. A legal entity that is not a human being but can hold rights and duties, such as a company or a government. The law creates it because it is useful, not because the entity is morally deserving. Abbott says this directly [03:07].
• Negligence vs strict liability. Negligence: you are liable only if you failed to take the care a reasonable person would have taken. Strict liability: you are liable for the harm even if you took all reasonable care. Product liability for defects is strict in most US states.
• Reasonable person (also “reasonable man”). A fictional average, careful person used as the yardstick for negligence. Abbott points out that in practice a juror imagines how they would have acted [55:44].
• Reasonable robot. Abbott’s coinage, a twist on the above: once AI is much safer than people, the human is judged against what a reasonable AI would have done [13:56].
• Mens rea. Latin for “guilty mind”. The mental state (intent, knowledge, recklessness) that many crimes require, on top of the act. Abbott’s phrase is “a wrongful mental state” [27:54].
• Strict-liability offence. A crime that needs no mental state. Abbott points out that corporations are already punished for these [31:05].
• Person having ordinary skill in the art (PHOSITA). The patent-law yardstick for “obviousness”: an imagined average researcher in the field. If that person would find the invention obvious, no patent [24:48]. Abbott’s proposal: “a machine skilled in the art” [26:21].
• Inventorship vs ownership. Who is named as the inventor is a separate question from who owns the patent. In his proposal the AI is named, and the AI’s owner owns [24:48].
• Explainability. The ability to say why an AI system reached a result. Legally salient in Europe. Abbott wants it to be “very context dependent” [38:50].
• AGI / artificial superintelligence. AGI: AI able to do anything a person can. Abbott uses it as a thought experiment, with the line “i digress” [26:21]–[27:54].
• Centaur (chess). A human plus a machine playing as one team.

People and cases named (a reading list)

Person / case	Why they come up	Start with
Kate Darling	Robot cruelty, argued on the grounds of human dignity	Her papers on robot ethics and the “kids, animals and robots” analogy (from memory; check titles)
Burrow-Giles v. Sarony	Photograph as authored work; the copyright office’s reasoning now	The 1884 Supreme Court opinion (111 U.S. 53)
DABUS / Stephen Thaler	The test case behind the talk	See abbott-patent-law-ai-inventors
Gary Kasparov	Centaur chess; human plus machine	Deep Thinking (2017)
Siemens	Engineers who would not sign as inventors (anecdote, unverified)	Abbott’s book, patent chapters
Sophia the robot	”Citizenship” in Saudi Arabia, 2017	See the personhood deep dive
Washington Post / Heliograf	Newsroom AI	Search “Heliograf”
Christie’s sale	First AI-generated work sold at major auction	Portrait of Edmond de Belamy (2018)

Abbott’s own frame

Abbott is a lawyer and physician-adjacent academic, and the book is The Reasonable Robot: Artificial Intelligence and the Law (Cambridge University Press, 2020, from memory; not checked). He leads the Artificial Inventor Project, which filed the DABUS applications in many countries [20:09]. This matters for reading him: part of the talk is advocacy for a case he is running, and he says so (“we filed”, “we announced”) [21:42]. He also argues from consequences (safety, innovation, revenue) more than from principles of rights.

The current debate (added context, as of 30 Oct 2026; not the speaker’s view)

The talk is six years old. Where later events have overtaken a claim, they are noted here and in the deep dives.

• Patents: the test case lost everywhere that matters, and the open question moved. The UK Supreme Court held on 20 Dec 2023 that an inventor under the Patents Act 1977 “must be a natural person”, but it did not decide whether AI-assisted inventions are patentable in general (IPWatchdog, CNBC). The Abbott talk’s prediction that “some jurisdictions are going to allow patents on this” [51:09] has, as far as I could check, not happened for AI-as-named-inventor in the big jurisdictions. For the US, EPO, Australia and Germany results, from memory and not checked: Federal Circuit (Aug 2022) and EPO Board of Appeal (Dec 2021) also refused. Go to abbott-patent-law-ai-inventors for depth.
• Copyright: the human-author rule was confirmed. The US Supreme Court denied review in Thaler v. Perlmutter on 2 Mar 2026, leaving the D.C. Circuit’s ruling that copyright needs human authorship in place (SCOTUSblog case page, Reed Smith, Patently-O on the government’s brief). The denial carries no opinion. The live question now is how much human input is enough, with a case (Allen v. Perlmutter, Colorado) pending at the time of the sources; I did not check whether it has been decided.
• Tort: the EU took the other fork. The Commission withdrew its proposed AI Liability Directive in Feb 2025 for lack of agreement (Bird & Bird). The revised Product Liability Directive, adopted Oct 2024, treats software and AI-integrated products as “products”, which keeps the strict-liability track Abbott criticised (Taylor Wessing). Sources differ on the application date; from memory, member states must apply it from December 2026 (not checked).
• Safety evidence is better than in 2020, with caveats. A Swiss Re study with Waymo, using insurance claims, reported 76% fewer property-damage claims and no bodily-injury claims over 3.8 million driverless miles (Waymo blog); a 2025 follow-up reported larger reductions (Repairer Driven News). My search did not show whether those studies were peer-reviewed, and Waymo commissioned the first. I found no jurisdiction that has adopted a “reasonable robot” test for human drivers (from memory; not checked).
• Personhood: the trend is the opposite of what he feared or hoped. Several US states have passed or proposed laws that deny legal personhood to AI. Reports name Idaho, Utah, North Dakota and Tennessee; Ohio’s House Bill 469 was still a pending bill at the time of the source (SAN, Fox News). These are secondary sources; the statutes were not read.
• Tax: the robot-tax idea is alive but no major regime exists. A 2026 arXiv position paper argues automation erodes the tax base and discusses token taxes as an alternative to robot taxes (arXiv 2603.04555). A law-firm summary lists the standing problems, the same ones Abbott gives: defining a “robot”, who bears the tax, and the damage to innovation (Miller & Chevalier). One blog claims an April 2026 OpenAI policy paper proposed shifting tax from payroll to capital; this rests on a single secondary source and is unverified.
• The 94% statistic is contested. See deep dive 3.

---

Part 2 — The intellectual map

Abbott spoke for 35 minutes to a policy audience. He names almost none of the thinkers behind his positions. Seven traditions feed the talk.

 LAW & ECONOMICS OF ACCIDENTS ──┐                     ┌── THEORIES OF PUNISHMENT & MENS REA
 Hand · Calabresi · Shavell     │                     │   Kant · Bentham · Hart · Hallevy
 negligence vs strict liability │                     │   responsibility gap (Matthias)
                                ▼                     ▼
                       ┌─────────────────────────────────────┐
 REASONABLE-PERSON ──► │  ABBOTT: judge BEHAVIOUR, not the    │ ◄── LEGAL PERSONHOOD
 STANDARD              │  kind of actor; outcomes first       │     Salomon · Stone · Solum
 Vaughan · Holmes      └─────────────────────────────────────┘     EU 2017 · Whanganui
                          ▲                  ▲                ▲
   IP INCENTIVE THEORY    │   TAX & AUTOMATION│   REGULATING AI: │
   Locke · Burrow-Giles   │   Keynes · Autor ·│   technology-    │
   CONTU · CDPA s9(3)     │   Acemoglu        │   neutrality ·   │
                          │                   │   explainability │

2.1 The law and economics of accidents — why the liability standard matters

The problem it addresses. Accidents cost money and lives. Liability rules are a price system for risk. Which rule produces the right amount of care and the right amount of activity?

The chain of thinkers

1. Learned Hand, United States v. Carroll Towing (2d Cir. 1947). The “Hand formula”: a party is negligent if the burden of precaution (B) is less than the probability of harm (P) times the loss (L). This turns “reasonable care” into cost-benefit. Feeds Abbott: his case for negligence is a case that machine behaviour should be tested by this cost-benefit logic, not by a rule that asks only whether harm happened.
2. Guido Calabresi, The Costs of Accidents (1970). Liability should fall on the cheapest cost avoider, the party best placed to prevent harm at the lowest cost. Strict liability for manufacturers rests on this: the maker can improve design across millions of units, while a buyer cannot. Abbott accepts this in general [37:16] (“manufacturers are much better situated than consumers”), then argues the exception: when the product replaces a more dangerous human activity.
3. Roger Traynor, Greenman v. Yuba Power Products (Cal. 1963), and Restatement (Second) of Torts §402A (1965). The birth of strict product liability in the US. Reasons: consumers cannot inspect products, makers can spread the cost through prices and insurance. From memory; not checked.
4. Steven Shavell and others on activity levels. Strict liability is efficient in controlling how much of a risky activity people do, but negligence is better when the question is how carefully an activity is done. If a safer technology is taxed with higher liability, the activity level shifts toward the more dangerous alternative. This is exactly Abbott’s “discourages the use of ai even if it may be safer than a person” [12:23]. From memory; not checked.

The thread. Abbott’s tort argument is not exotic law and economics. It is standard theory applied to a case the original theories did not foresee: a product that competes directly with a human on the same task.

2.2 The reasonable-person standard — an old fiction doing hard work

The problem it addresses. Courts need one yardstick for careful behaviour without writing a rule for every situation.

1. Vaughan v. Menlove (1837) moved English law from a subjective “best of his own judgment” standard to an objective one. A man who stacked hay badly could not escape liability by saying he was not clever. Key point: the standard ignores individual variation by design.
2. Blyth v. Birmingham Waterworks (1856). Baron Alderson’s definition: negligence is the omission to do what a reasonable person would do, or doing what such a person would not do. From memory; wording not checked.
3. Oliver Wendell Holmes, The Common Law (1881). Defended the objective standard: the law cannot weigh each person’s talents, and society needs predictable minimum conduct. Holmes accepted that some people would be held to a standard they could not meet.
4. The critique that Abbott raises himself [55:44]. The “average person” does not exist. A juror imagines their own driving. Feminist and critical scholars have long argued the “reasonable man” encodes a particular kind of person. Abbott’s answer in the talk is the pragmatic one: no better solution has been found, safe drivers enjoy the benefit, bad drivers “too bad for you”.
5. Where the “reasonable robot” fits. Standards have always moved with technology. Courts judge a motorist by the standards of motoring, not of horse-riding. What is new in Abbott is the ratchet: the yardstick would be set by a machine, and the human would have to meet it or stop. Compare the T.J. Hooper case (Hand, 1932), where a whole industry’s custom was not enough to defeat negligence if a cheap safety device (radio receivers on tugs) was available. From memory; not checked. It is the closest existing precedent for Abbott’s “widespread availability and affordability” condition [41:54].

2.3 Intellectual property incentive theory — why patents on AI output might make sense

The problem it addresses. Why does law protect creations at all, and what happens when the creator is not a person?

1. Two justifications. Natural-rights / labour theories (Locke, Hegel): a creator has a claim to what they make. Utilitarian / incentive theories (the US Constitution’s Progress Clause): protection exists to encourage production and disclosure. Abbott says he is working in the “anglo-saxon” economic tradition [48:05].
2. Burrow-Giles v. Sarony (1884). The Supreme Court held a photograph could be copyrighted because the photographer’s choices (pose, lighting) express ideas “in the mind of the author”. The Copyright Office now cites this line to say a machine has no mind [18:34]. Abbott quotes the thrust of the case, not the full text.
3. Naruto v. Slater (9th Cir. 2018). The monkey-selfie case; the court held animals lacked statutory standing. The Office’s guidance has long grouped monkeys and machines together. Not named by Abbott, who only says “neither can monkeys” [18:34].
4. CONTU (1978) and computer-generated works. The US commission treated the computer as a tool of the human author. The UK’s Copyright, Designs and Patents Act 1988, s.9(3), goes further: for a computer-generated work the author is the person who made the arrangements for its creation. From memory; not checked, and I did not check whether the UK has since changed this. It is the best existing evidence that AI-owner-owns is not a crazy idea [24:48].
5. Work-for-hire. The corporation can be the legal “author” of an employee’s work [18:34]. Abbott’s analogy: the AI’s owner stands to the AI as the employer stands to the employee.
6. Obviousness and the skilled person. Abbott [24:48] points out that the benchmark for non-obvious inventions is an imagined ordinary researcher. If the ordinary researcher routinely uses AI, the bar rises. For the scale of this idea and the later case law, see abbott-patent-law-ai-inventors.

2.4 Legal personhood — the artificial-person tradition

The problem it addresses. Who or what can hold rights and duties?

1. The corporation. Salomon v. Salomon (1897) fixed the separate legal personality of a company in English law. Corporations hold property, sign contracts, sue and are sued. Abbott uses this as his proof that granting personhood is a tool, not a moral verdict [03:07].
2. Christopher Stone, “Should Trees Have Standing?” (1972). Argued that natural objects could have legal standing through guardians. Realised in the 2017 laws giving the Whanganui River in New Zealand legal personhood (also mentioned in the Citton page for a different argument). From memory; not checked.
3. Lawrence Solum, “Legal Personhood for Artificial Intelligences” (1992). The classic early paper. It separates the legal question (can we give an AI the functions of a legal person?) from the moral one (does it deserve them?). This is the same split Abbott makes. From memory; not checked.
4. European Parliament resolution (Feb 2017) on civil-law rules for robotics, which floated an “electronic personality” for the most advanced autonomous robots. An open letter from AI and law experts in 2018 opposed it. Abbott’s remark that the idea “lost traction” in the EU [45:00] reflects this. From memory; not checked.
5. The contrary view: personhood as a liability shield. Critics (for example Bryson, Diamantis and Grant, “Of, for, and by the people”, 2017) warn that giving AI legal personality would let owners hide behind it. This is the standard objection to Abbott’s own self-driving-car insurance example. From memory; not checked.

2.5 Theories of punishment and the responsibility gap

The problem it addresses. What is punishment for, and what must exist before it is just?

1. Retributivism (Kant). Punishment is deserved because of wrongdoing, and it requires a responsible agent. Abbott: “culpability and retribution are really fundamental to criminal law” [29:28].
2. Utilitarianism (Bentham). Punishment is justified by deterrence and incapacitation. Abbott’s case for punishing a machine is entirely this kind: it would shape the behaviour of “owners, users and developers” [31:05].
3. H.L.A. Hart on responsibility. Capacity and fair opportunity to do otherwise are conditions of just punishment. Machines fail the test on most accounts.
4. Corporate criminal liability. New York Central & Hudson River Railroad v. United States (1909) let a company be convicted via the acts of its employees. This is the “impute the mental states of a corporation’s agents” doctrine Abbott invokes [29:28]. From memory; not checked.
5. Gabriel Hallevy’s three models of criminal liability for AI: perpetration-by-another (the AI is an innocent tool), natural-probable-consequence (the developer or user is liable for foreseeable results) and direct liability (the AI itself). Abbott’s answers about the cryptocurrency-stealing AI [54:13] sit in the first two. From memory; not checked.
6. Andreas Matthias, “The responsibility gap” (2004). When a learning machine acts in ways its makers could not foresee, nobody may be fairly blameable. Abbott names the same gap: it is “harder and harder to reduce machine behavior to a particular human act” [27:54].

2.6 Tax and automation — the labour-capital wedge

The problem it addresses. Tax systems evolved when labour was the main source of both income and revenue.

1. Technological unemployment. Keynes coined the phrase in 1930 (“Economic Possibilities for our Grandchildren”). The 19th-century Luddite fear is the one Abbott’s agricultural example answers [04:40]. David Autor’s “Why are there still so many jobs?” (2015) is the modern statement of why displacement has been matched by new work so far. From memory; not checked.
2. The tax wedge. In the US, labour is taxed through payroll and income taxes, capital (including equipment) is typically taxed lightly through depreciation and interest deductions. Acemoglu, Manera and Restrepo (“Does the US tax code favor automation?”, 2020) estimated that the effective tax rate on equipment was much lower than on labour, tilting firms toward automation. This is Abbott’s first point in modern form [06:12]. From memory; not checked.
3. The robot-tax debate. Bill Gates proposed a robot tax in 2017. The European Parliament rejected the idea in its 2017 robotics resolution. Critics (including those who warn about definitions and innovation) match Abbott’s own list of drawbacks. From memory; not checked.
4. Abbott’s own tax writing. With Bret Bogenschneider he published “Should Robots Pay Taxes?” in the Harvard Law & Policy Review (2018). The talk compresses it. From memory; not checked.
5. Capital mobility. The textbook reason to tax labour over capital is that capital can leave. Abbott’s counter [09:18] (“people are not rushing to invest in businesses in the cayman islands”) is an empirical bet on agglomeration: big markets and infrastructure keep capital in place. This is the weakest-evidenced step in the argument and should be read as such.

2.7 Regulating AI — technology neutrality and the explanation question

The problem it addresses. Should law regulate the technology, the behaviour, or the process?

1. Technology neutrality. A drafting principle: write rules around functions and outcomes, not around specific technologies, so they last. Abbott’s AI legal neutrality is this idea applied between actor types.
2. The explainability debate. The GDPR (Article 22, 2016) restricts solely automated decisions with significant effects and is read by many as giving a right to meaningful information about the logic. Cynthia Rudin (“Stop explaining black box machine learning models for high stakes decisions”, 2019) argues that for high-stakes uses we should build inherently interpretable models, not explain opaque ones after the fact. Abbott’s view [38:50] is context-sensitive and differs from both. From memory; not checked.
3. State v. Loomis (Wis. 2016). The COMPAS risk-assessment case in sentencing, which is almost certainly the “criminal sentencing” worry Abbott mentions [38:50], though he does not name it.
4. The collingridge dilemma (1980): early on a technology is easy to regulate but poorly understood; later it is understood but hard to change. Abbott’s last answer [57:18] (“ai has already regulated plenty”) is a way out: the technology is already governed by laws made for other things.
5. The EU AI Act (in force Aug 2024, from memory) came after the talk. It is risk-based and process-heavy, the opposite of Abbott’s outcome-first instinct. Added context.

2.8 One shorter thread: humans and machines together

Centaurs and complementarity [32:38]–[34:10]. Brynjolfsson and McAfee’s Race Against the Machine (2011) and later work on complementary skills; Kasparov’s Deep Thinking (2017) on advanced chess. The chess evidence (that centaurs beat both humans and engines) held for a limited period and then did not, which is Abbott’s own closing point: “in the long term people may just get in the way” [34:10]. The caveat is that chess is a closed, rule-bound game. From memory; not checked.

---

Part 3 — Deep dives

1. AI legal neutrality — the organising principle

🎙 In the talk. The book argues that the law distinguishes between a machine and a person doing the same thing, and that “we would get better social outcomes if we did less to legally discriminate between ai and human behavior” [03:07]. He heads off a misreading at once: “an ai can behave like a person but an ai is not like a person”. It does not morally deserve rights. Rights for artificial persons already exist (corporations, governments), but “we do grant rights to artificial persons but we do so because we think that this produces good social outcomes” [03:07]. He gives the example of an autonomous car that might need to hold its own insurance policy, and says that would not be because the AI “was morally deserving of rights” [03:07]. At the end, he adds that AI is “already regulated plenty” by laws that “weren’t designed to regulate ai”, and that this makes no sense in places [57:18]–[58:53].

📚 Deeper theory

• A functional, not a metaphysical, test. Abbott borrows the structure of Solum’s 1992 split (2.4): decide the legal question by asking what outcomes a rule would produce, not whether a machine is “really” intelligent. This lets him avoid the philosophy of mind completely.
• Neutrality is not identity. The principle does not say identical treatment. It says differences need a justification tied to social outcomes. This is the structure of equal-protection or non-discrimination analysis: like cases alike, relevant differences allowed. What counts as “relevant” is where all the work happens.
• Where the principle bites. Wherever law sorts by actor type: tax (employee vs machine), tort (driver vs product), IP (human author vs none), criminal law (natural vs artificial person). He treats each as a separate application of one rule.
• A hidden asymmetry. In tax and tort, neutrality means lowering a burden on one side (tax, liability). In IP, it means extending a protection. In criminal law, he concludes against extending it. So the principle is not applied mechanically: the outcome test decides.

Where others push back

• Behaviour is not the only thing that differs. A person can be deterred, can understand a duty, can suffer a sanction. A product can be recalled and redesigned. Different tools for different actors can be a feature (the strict-liability defender’s point, 2.1).
• Neutrality can launder harm. Treating an AI as “just a reasonable driver” may hide systemic failure modes: one bug affects a whole fleet at once, while human errors are independent. This is a standard reply to equating one AI error with one human error (from memory; not checked).
• Category worries. Whether “AI” is a stable category to legislate around at all.

What he leaves out. A theory of when a difference is justified. The talk gives instances (criminal punishment) but no test.

Check yourself

• Why does Abbott say neutrality is compatible with denying that AI deserves rights?
• Give one area where neutrality means lowering a burden, and one where it means extending a protection.
• What would count as a good reason to treat an AI differently?

---

2. Tax — the payroll wedge and the missing revenue

🎙 In the talk. He starts with history: in the first industrial revolution people feared machines would take all jobs, and that did not happen; farm work fell from “about 40 percent” of the workforce to “about 3 percent” with no mass unemployment [04:40]. But “it’s not entirely clear whether or not history is going to repeat itself”, because there may be “fewer and fewer limits to what machines can do” [04:40].

Tax plays two roles. First, it treats a machine and a person differently: McDonald’s pays no payroll tax on a self-service kiosk, plus gets other benefits such as accelerated deductions, so even at equal cost and quality it saves tax by automating [06:12]. Second, “robots don’t pay taxes”. About 90% of federal revenue comes from labour in one way or another, and under 10% from corporate tax, so when a firm swaps a cashier for a machine we lose payroll and income tax and receive little extra corporate tax even if profit rises [06:12]–[07:44]. His summary: “tax policies are not only inefficiently subsidizing automation they are also reducing tax revenue” [07:44].

His remedy: neutrality. A literal robot tax has problems: “it’s hard to say what a robot is”, robots do not replace people one for one, and there is “a lot of administrative burden” and gaming [07:44]. Better: cut labour taxes such as payroll tax and make up the revenue by rethinking “how we’ve been taxing labor versus capital” [09:18]. Capital will not simply flee: he cites US infrastructure, workforce and market size (“people are not rushing to invest in businesses in the cayman islands”) [09:18]. When asked later whether we should tax planes, trains and pacemakers too, he says that is why he is “against a literal robot tax”; raising business taxes while abolishing payroll tax would be “effectively a robot tax” without W-2s for robots [49:37].

📚 Deeper theory

• Two goals, one instrument. Tax here does two jobs: it shapes behaviour (automate or not) and raises revenue. Abbott separates them: neutrality fixes the first; a shift from labour to capital fixes the second. Public finance calls the first an excess burden problem (Pigou-style distortions) and the second a base erosion problem (2.6).
• Evidence on the wedge (Acemoglu, Manera and Restrepo, 2020; from memory; not checked): effective US tax rates were far lower on equipment than on labour, which supports his first point. Whether this causes “excessive” automation is contested, because a rate wedge is not proof of socially wasteful adoption.
• Is the revenue claim right? Labour taxes are a large share of US federal revenue, but “about 90 percent” is loose (see slip 7). Other countries rely on VAT and consumption taxes, which a labour-to-capital shift does not touch in the same way.
• The distribution argument [09:18]: automation “is going to result in a lot of wealth” in “narrowly controlled” hands. This links the tax proposal to inequality, not only to efficiency. It is the part of his case closest to ordinary progressive tax politics.
• Optimal-tax objection. Standard theory (Atkinson-Stiglitz; Chamley-Judd, both from memory; not checked) says capital should not be taxed in the long run because it harms growth. Abbott implicitly rejects that for a world where capital substitutes for labour. This is the deepest theoretical dispute under his tax section, and the talk does not engage it.

Where others push back

• Taxing labour less and capital more is not obviously “neutral”. It changes the relative price of machines in the other direction. Neutrality in the sense of equal treatment of equal-productivity inputs is hard to define.
• Automation may raise wages. If machines raise productivity, labour income may rise with them; the revenue loss is then smaller than the talk implies. Evidence is mixed and the debate is live.
• Added context (2026). A literature survey and law-firm summary (links in Part 1) still list the same practical obstacles to a robot tax, and I found no major jurisdiction with one.

What he leaves out. International tax competition in detail; VAT; the political fact that a labour-to-capital shift is hard to pass. Also, how to tax the firm that uses AI services bought from a vendor (no on-premises machine to point at).

Check yourself

• What are the two roles of tax in the automation debate, per Abbott?
• Why does he reject a literal robot tax but say a corporate-tax rise plus payroll-tax cut is “effectively” one?
• Which step of the argument depends most on capital not leaving the country?

---

3. Tort — from strict liability to the “reasonable robot”

🎙 In the talk. The framing question is not whether AI drivers will cause accidents (“clearly yes”) but whether they will cause fewer than humans. He argues they will, citing that “human error accounts for about 94 percent of auto accidents” and that “people are not very good drivers” [10:50]. Today a human driver is judged by negligence, while a self-driving car, as a commercial product, is judged by strict product liability, “a higher standard of liability” [10:50]–[12:23]. His reasoning: “it doesn’t make sense to have a higher liability standard if you have a technology that is going to be safer than the alternative”, because it “discourages the use of ai even if it may be safer than a person” [12:23].

Proposal one: hold both to negligence. A second benefit is practical. Working out why a self-driving car crashed is expensive, so under negligence “you can just focus on the ai’s behavior did it run a red light not why did it run a red light” [12:23].

Proposal two, later: “the liability standard you apply to ai is less important than the liability standard you apply to people” because AI keeps improving and humans do not [13:56]. Once self-driving cars are standard, judge the human driver by whether a “reasonable ai or a reasonable robot” would have caused the accident, even though the AI could “see the child … through buildings using radar” [13:56]. The fairness argument: the human has the option of using the AI, and by not doing so internalizes the cost of accidents [13:56]. The alternative, banning human driving, “would limit human freedom and autonomy” [15:28]. He extends this to surgery, with the line “it may not even be ethical for human doctors to do those surgeries” if machines do them better [15:28], and to lawyers.

In the Q&A he says strict product liability is right in general, because manufacturers can best manage safety. The difference here is that the comparison is not “a new product versus nothing” but “versus a very concrete substitute”, so the goal is “not some absolute level of safety” but something “safer than the substitute” [37:16]. On opacity: “by focusing on the behavior rather than what’s going on inside the ai we’re focused on the right issues from a tort perspective” [38:50]. He says explainability “has to be very context dependent”: vital in bail or sentencing, where there is no obvious right answer, but of little use in driving, where the right outcome is known [38:50]. On the trolley problem: self-driving cars will mostly make it moot, since “it’s going to get rid of the trolley problem” by braking faster [40:22]. On defining the standard across price points: it must be “widely available and affordable”, not only a $100,000 car [40:22]–[41:54]; “widespread availability and affordability” is the key factor [41:54]. For an inconsistent human “average”: “pretty soon we are all going to be bad drivers compared to ai”, and we should keep the right to drive [55:44].

📚 Deeper theory

• The activity-level argument (2.1). If strict liability makes the AI more expensive to deploy than its human substitute, the market shifts toward the riskier option. This is a clean economic point and is Abbott’s strongest one.
• Why strict liability for products exists. Consumers cannot assess design; makers can spread costs. Abbott concedes this and argues only that it should not apply when the product is a substitute for human conduct. A harder question is whether it should apply when the human alternative is itself unreasonably risky.
• The 94% figure needs care (added context). It comes from NHTSA’s crash-causation survey of 2005–2007: researchers assigned a “critical reason”, the last event in the causal chain, to the driver in about 94% of crashes, to the vehicle in about 2% and to the environment in about 2% (NHTSA brief). NHTSA states the assignment is not meant as the cause or as blame. NTSB’s chair has called the way the figure is quoted misleading (Transportation Topics). It supports “humans are involved in most crashes”. It does not show that AI would remove 94%. Abbott’s conclusion that AI could be safer does not depend on it, but his phrasing does.
• Evidence since 2020. Insurance-claims data from Waymo’s operations (Part 1) is the strongest published support for “AI drivers cause fewer accidents”. Limits: the first study was commissioned by Waymo; the cities and conditions are narrow; I did not check peer review.
• Negligence for machines is awkward in doctrine. A reasonable-person test asks about foresight and care. A machine has no foresight in that sense. Courts would have to ask what a reasonably designed system would do. That drifts back toward design-defect analysis, the very strict-liability territory Abbott wants to leave (from memory; consistent with the product-liability literature; not checked). His answer is to focus on the behaviour only [12:23].
• The ratchet. The “reasonable robot” for humans is a one-way standard: as AI improves, the human duty rises. T.J. Hooper (2.2) is the nearest precedent, and it asks for availability and affordability, which Abbott adds [41:54].
• EU path (added context). The EU kept product-style liability and extended it to software, and dropped its separate AI liability directive (Part 1). Abbott’s “negligence for AI” proposal is not the direction taken there.

Where others push back

• Fleet-level correlation. One software defect can hurt many people at once, which strict liability and recalls are built for.
• Victim compensation. Under negligence, a victim of a car that behaved “reasonably” may go uncompensated, even if the car had a defect that cannot be shown. Strict liability pays out more reliably. Insurance or no-fault schemes could fill the gap; Abbott does not address them.
• The human “reasonable robot” standard punishes the individual for a market failure. It may be unfair to a poor driver who cannot afford the technology. Abbott’s affordability condition answers this in principle.
• Moral hazard in the other direction. Manufacturers advertising safety is not a substitute for liability pressure.

What he leaves out. Insurance structures and no-fault compensation; data access for plaintiffs; who proves what when the system is a black box; and drivers’ loss of skill when they rely on automation.

Check yourself

• Why can a higher liability standard on a safer product make roads less safe?
• What does the 94% figure measure, and what does it not measure?
• Name two conditions Abbott gives before humans could be held to the “reasonable robot” standard.

---

4. Intellectual property — what the talk says (patents in short)

🎙 In the talk. Machines now create commercially relevant work. The Washington Post’s Heliograf writes short articles and “people are often unable to distinguish between the two” [17:02]. Christie’s sold an AI-generated painting for “about half a million dollars” [17:02]. Copyright usually requires a human author. In the US a corporation can be the legal author of an employee’s work (work for hire), but the Copyright Office does not allow the same for AI output. It cites Burrow-Giles (the Oscar Wilde photograph) and argues “machines can’t think don’t have minds and therefore can’t be authors and neither can monkeys” [18:34].

On patents, he leads a test case on two inventions made by an AI (DABUS): a fractal-geometry beverage container and a flashing light for emergencies. The creator “hadn’t thought of these things and was surprised” [20:09]. He adds that Siemens had an AI-made car part, but its engineers would not name themselves as inventors, and “falsely reporting yourself as a patent inventor in the united states is a crime” [21:42]. The applications were filed in the UK and Europe, then publicly stated the AI had invented them; they were rejected in the US, UK and Europe, with appeals pending and applications in about a dozen more jurisdictions [21:42].

His argument: machines do not care about patents, but “the people who make use and build artificial intelligence care about patents”, and if patents exist to encourage innovation they should cover AI output [21:42]. This matters most in life sciences (the antibody-screening example) [23:16]. They listed the AI as inventor “we didn’t do it to give the ai credit”, but to inform the public and to stop someone claiming credit for work they did not do [23:16]. Ownership: “the ai’s owner should own any patents on its output” [24:48]. The “person skilled in the art” test should change, eventually to “a machine skilled in the art” [26:21]. In Q&A: the patent system is mainly an economic tool in the Anglo-American tradition, so the economic rationale “supports protecting ai generated output” [48:05]. For the UK result, he says it “came down to whether or not the word person was used in the patents act”, which he finds “kind of doesn’t make sense” for a 1970s statute [51:09]. Humans can often be named: someone who programmed, trained or used the AI, or recognised the output as inventive; the boundary is “murky” but already murky for large teams [52:40]–[54:13].

📚 Deeper theory

• Predictions vs outcomes (added context). He predicted that “some jurisdictions are going to allow patents on this” within five years [51:09]. As far as my search showed, the leading courts went the other way: the UK Supreme Court (Dec 2023) required a natural-person inventor; the US Supreme Court declined Thaler v. Perlmutter on copyright in March 2026 (Part 1). See abbott-patent-law-ai-inventors for the full patent record and his later views.
• Incentive theory has a gap here. If a machine produces inventions at near-zero marginal cost, the incentive to invest comes from the machine’s price and the firm’s costs, not from a monopoly on each output. Critics ask whether patents on AI-made inventions over-reward (2.3, from memory; not checked). Abbott’s reply is the life-sciences point: where patents are needed to recoup trial costs, no patent means no use of the machine.
• The “inventor” doctrine already has a human-contribution test. Conception is the touchstone in US law. Abbott’s point that human contribution is often easy to find [52:40] and already murky in teams [54:13] is a fair description of current practice.
• The mind requirement. Courts read “author” in Burrow-Giles as needing a mind. Abbott’s neutrality argument turns on purpose (what the system is for) rather than the metaphysics of mind (2.3).

Where others push back

• Honesty of the inventor. Naming a machine as inventor may hide who contributed. Abbott argues the reverse: it reveals it [23:16].
• Flooding the patent system. If AI can invent “10,000 things”, the thicket of patents could block competitors. He raises this himself as a reason to keep credit honest [23:16].
• The statute-text view. Courts and the Copyright Office say Congress and Parliament, not judges, should decide. The UK court said as much in effect (“take a step back” is Abbott’s own wording on policy [51:09]).

What he leaves out. Data inputs and training-data copyright; infringement by AI; whether AI-made works should get shorter terms.

Check yourself

• Why does Abbott want the AI named as inventor but its owner to own the patent?
• What does “machine skilled in the art” change in practice?
• Which of his five-year predictions did not come true, and what does that say about the argument?

---

5. Criminal law — can a machine commit a crime?

🎙 In the talk. Criminal law differs from tort, tax and IP because it is not only utilitarian: “we care a lot about why people did certain things”, and a wrongful mental state can turn a tort into a crime [27:54]. So far machines are tools: “if i pick up a computer and hit you with it that’s me committing a battery” [27:54]. Later, as autonomy grows and behaviour is “harder and harder” to reduce to a human act, machines may “functionally” commit crimes [27:54]. Problem: “culpability and retribution are really fundamental to criminal law” [29:28]. But corporations already face criminal liability: “we impute the mental states of a corporation’s agents to it” [29:28]. Some question whether corporate liability makes sense at all (it punishes innocent shareholders), others hold that companies have a life above their agents [29:28]. Attributing a machine’s act to a person gets harder with complex training and open-source contribution over time [31:05].

Benefits of punishing an AI: it would affect owners, users and developers (“we’re influencing the behavior of the people who make ai”), for instance by confiscating or destroying a valuable AI, and it would have expressive value (a car “deliberately running over investment bankers”) [31:05]. Costs: changing the law for a problem not yet significant, risk to “the rule of law”, and a message “that machines are on par with human beings” [31:05]. Verdict: “punishing machine probably isn’t a good idea”, but it is not so inconsistent with doctrine, because we already punish corporations for strict-liability offences and for failing to act where there is a duty, even “without an adverse mental state” [31:05]–[32:38]. Q&A: if a creator trained an AI to steal cryptocurrency and then died, “it is less what the ai has done and more what you have done”; the criminal law follows the human’s act and state of mind, and a dead person cannot be prosecuted [54:13]. On sentencing AI he says explainability is essential [38:50].

📚 Deeper theory

• Why the corporation is the bridge, and why it is weak. The corporation analogy (2.5, 2.4) works because a corporation has human agents whose acts and minds can be attributed. A truly autonomous AI has no such agent. Abbott acknowledges this by saying attribution becomes hard [29:28].
• The responsibility gap (2.5). The talk comes close to naming this and moves on quickly. It is the most important unresolved issue in AI criminal law. Options in the literature: strengthen developer and operator duties (negligence-type offences), create new offences for deploying unsafe systems, or accept some gaps.
• Deterrence on people, not machines. His utilitarian case for punishing the AI really targets humans. This is a point critics make: if so, punish the humans directly and skip the fiction.
• Expressive function. Criminal law condemns. Some theorists (Feinberg, expressive theory of punishment; from memory; not checked) argue condemnation of a non-moral agent is category error, which is Abbott’s “troubling message” cost.
• Added context. The AI-personhood bans in US states (Part 1) read as the legislative form of Abbott’s conclusion in this section, though adopted for different reasons than his.

Where others push back

• Direct liability advocates (Hallevy) argue an AI meeting the elements of a crime could be liable in principle, mostly as a tool for deterrence and for modelling corporate-style liability.
• Strict-liability offences for deployers could capture what Abbott wants without any machine punishment, and that is arguably a simpler path.

What he leaves out. Crimes where the AI is the victim or instrument at scale (fraud, deepfakes); mens rea for developers who accept foreseeable misuse; how to treat autonomous weapons.

Check yourself

• Why is criminal law “different” from tort and IP in Abbott’s account?
• What does the corporate analogy get right and where does it fail for AI?
• In the cryptocurrency example, what is the crime and who commits it?

---

6. Personhood, rights and cruelty

🎙 In the talk. At the start: giving an AI rights, if ever, is about outcomes for people [03:07]. In Q&A on cruelty to robots, he says he has not thought much about it. He cites Kate Darling’s point: the UK law against animal cruelty is not based on animals having rights but on cruelty being “demeaning to human beings” [43:27]–[45:00]. His own view: “animals do have rights”, so there are two reasons to avoid cruelty to animals (the animal’s sake and human dignity). With robots only the second applies, and “there is some validity” to the idea that letting people do anything to robots is demeaning and “may encourage certain sorts of anti-social behaviors” [45:00]. On personhood: it is discussed more than before, but “i don’t think it’s going to catch on in the eu”; Saudi Arabia and Sophia may have been “a publicity matter”; Japan and India are more flexible [46:31]. His line: “it would only be because this somehow resulted in better outcomes from people”. For each example (AI inventor assigns patent; AI with its own insurance), it is “better and easier and more efficient not to do it” [46:31]. A self-driving car with no commercial maker is unlikely; “there’s almost certainly someone who’s going to be making a self-driving car for commercial benefit and that’s the party who should be liable” [46:31].

📚 Deeper theory

• Direct vs indirect duties. Kant held we have no direct duties to animals but should not be cruel because it corrupts our character toward humans (from memory; not checked). This is the “demeaning to humans” argument Abbott attributes to the UK law via Darling. Abbott rejects it for animals but accepts it for robots, which is a coherent position.
• Moral patients vs legal persons. Whether a being can be harmed (moral patient) is different from whether it can hold legal rights (legal person) (2.4). Abbott keeps them apart, and so his robot-cruelty answer is a human-dignity argument only.
• Personhood as a tool. The corporation shows personhood can be granted without any moral claim. Abbott’s test (better outcomes for people) is the same test. The risk (2.4, point 5) is shielding owners from liability.
• Added context (2025–26). AI labs have model-welfare programmes and philosophers debate machine moral status; this was a minor topic in 2020. At the same time, US states moved to bar AI personhood by statute (Part 1). Both are developments after the talk and neither is Abbott’s view.
• Added context: Sophia. The 2017 “citizenship” was widely treated as a publicity event; Abbott hedges the same way [46:31].

Where others push back

• Precaution. If there is a real chance advanced AI can suffer, waiting for evidence of harm may be too late. This challenges Abbott’s behaviour-only stance.
• Slippery slope worries about banning cruelty to robots: it could weaken the line between persons and things.

What he leaves out. Any view on AI sentience; the possibility that AI systems may in future hold economic assets or be the beneficiaries of trusts; and the status of AI “agents” that sign contracts on behalf of people (a live issue after 2020).

Check yourself

• On what grounds does Abbott say a ban on robot cruelty could be justified?
• Why does he think AI patent ownership by the AI itself is “unnecessary complexity”?
• How does the corporation example both support and limit his argument?

---

7. Regulation, explainability and industry

🎙 In the talk. Host: the EU has “a really strong focus on explainability”, the US less so [38:50]. Abbott: the EU is “a bit more conservative”; explainability must be “very context dependent”. It is “critical” in sentencing or bail, where it is “not clear that there is a right answer”, and of “a lot less use” in driving or medicine, where he “absolutely know[s] what the right outcome is” [38:50]. On global standards: he expects cars to be expensive and available in developed countries first, then cheaper, and worries about a “developed world divide” that should be handled by “technology transfer and international trade” policy, “aside from ai legal neutrality” [41:54]–[43:27]. On China: legal frameworks will be “a critical factor in their industrial development”. Values on privacy, data and automated decisions differ and “reasonable people could differ”, but laws that make “the law function better for everyone” help development. The idea that regulation hinders AI “misses the point” because AI “is just regulated by laws that weren’t designed to regulate ai” [57:18]–[58:53]. The AGI aside: an AI told to improve itself would soon be superintelligent and “we’ll get to spend the rest of our days living in virtual reality but i digress” [26:21]–[27:54]. Board games: “in the long term people may just get in the way” [34:10].

📚 Deeper theory

• Outcomes vs process regulation (2.7). Abbott’s view fits domains where the outcome can be measured and agreed (crashes, surgical complications). It fits less in domains where the outcome is contested (sentencing, hiring), which he concedes in the bail example. The EU AI Act’s risk tiers, adopted after the talk, apply more process duties to the contested domains. In that sense, later law partly matches his split by context.
• The “regulation vs innovation” framing. His claim that AI is already regulated by ill-fitting law is the “legal uncertainty” argument: unclear rules deter investment as much as strict ones. It is plausible and not proven; from memory, empirical evidence on this is thin (not checked).
• China. At the time, China had no comprehensive AI statute. Since then it has issued rules on recommendation algorithms, deepfakes and generative AI (from memory; not checked). That is later and added context.
• Centaur evidence (2.8): human plus machine beats either alone only while the human adds something. His chess example is a caution against assuming augmentation is the stable end state.

Where others push back

• Explainability is also a right of the affected person, not only a tool for finding the right answer. A person harmed by an opaque system may have a dignity claim to a reason, even when the outcome is “right” on average.
• Outcome metrics can be gamed or biased (a system may be safe on average and unsafe for particular groups).

What he leaves out. Auditing and testing regimes that would make “outcome-based” regulation operational; who measures outcomes; data access.

Check yourself

• Where does Abbott want explainability and where not, and why?
• What is his reply to the claim that regulation harms AI development?
• What does the chess story suggest about augmentation as a long-term state?

---

Putting it together

Mental model. Abbott treats law as a set of incentive systems that were tuned for a world where humans do the tasks. When a machine does the same task, the old categories (employee vs machine, driver vs product, author vs none, natural vs artificial person) tax, punish or protect the two differently, and often in the wrong direction: they push firms to automate for tax reasons, slow safer machines down with stricter liability, leave AI output unprotected so firms avoid using AI, and fit poorly with criminal doctrine. The remedy is to judge behaviour and outcomes without regard to who or what acted, and to grant rights to non-humans only when that helps humans. The result is a principled, consequence-led position that deliberately avoids the questions of machine consciousness and rights. Six years on, courts and legislatures have mostly kept the human-centred categories (inventor, author, person), while the safety evidence for self-driving has strengthened his empirical premise.

Open tensions

1. Neutrality vs asymmetry. He applies the same principle to lower tax and liability burdens, extend IP and refuse criminal punishment. The test for when to differ is outcomes, which is flexible enough to justify nearly any result.
2. Negligence for machines vs the law’s design. Negligence presupposes foresight and care; machines have design and behaviour. The proposal depends on courts being able to judge machine “behaviour” without opening the box.
3. The ratchet on humans. The reasonable-robot standard makes the human duty depend on a market that may be unequal. His affordability condition is a patch, not a full answer.
4. Predictions vs outcomes. His five-year forecast on AI patents did not come true in the major jurisdictions, though his tort and criminal positions still look like live proposals rather than refuted ones.
5. Advocacy and analysis mixed. He leads the test case he discusses.

Reading path

• Start: the talk itself, [03:07] (principle), [10:50]–[15:28] (tort), [27:54]–[32:38] (criminal law).
• Then: Calabresi, The Costs of Accidents (1970); Solum, “Legal Personhood for Artificial Intelligences” (1992); Abbott’s own book.
• Deeper: Hallevy on AI criminal liability; Matthias on the responsibility gap; Acemoglu, Manera and Restrepo on the tax wedge; Rudin on interpretable models; the sibling pages abbott-future-of-ai-human-vs-machine and abbott-patent-law-ai-inventors.
• Items marked “from memory; not checked” above should be verified before citing.