The Role of Humans in an Age of Intelligent Machines: Jobs, Inequality, Governance. Joanna Bryson

Joanna J. Bryson · 2020 · 1:31:52 · ▶ Watch

Summary

Central question. Can AI replace people? Bryson turns it into a different question: what do we, as a society, choose to do with the productivity AI gives us?

Main thesis, in five steps.

1. Jobs are not the real problem; wages, inequality and fragility are. AI makes workers better at their jobs, not obsolete. Her example is Norwegian radiologists. The danger is that productivity gains flow to few people, widen inequality, and push down pay.
2. Whether software costs jobs is a political choice. “Software that made teachers twice as good” could mean twice-as-good schools or half as many teachers; “there’s nothing in that sentence that tells us about employment”.
3. Inequality and polarization are linked, and she has a model of why. When people feel unsafe they cut risk, cooperate only with their own group, and polarize. Identity politics is read as a survival strategy.
4. Ethics is made by societies, so AI cannot be a moral agent in the way a person is. Accountability works by dissuasion, which depends on human (animal) vulnerability. Legal personhood for AI would be “the ultimate shell company”.
5. Fix it with governance and transparency. Wealth taxes (not robot taxes), transnational coordination, GDPR-style rules, and software that can be audited. “AI ethics is just ethics”.

Close. Humans keep their role because everything we value is a project of “a bunch of apes trying to survive on the planet”. The job of governance is to share what technology produces.

Speaker’s stance. Deflationary about AI (a tool, an artifact, “not magic”) and strongly institutionalist. She is neither doom-minded nor boosterish. Her worry is power and politics, not machine takeover. She also pushes back on several popular ideas: the robot tax, the Hippocratic oath for programmers, universal basic income as the main answer, and machine legal personhood.

Map: talking points

#	Point
1	Moderator’s intro: lockdown, expertise and uncertainty, the series, how Slido works
2	Bryson’s bio and her habit of warning against anthropomorphizing AI
3	Norwegian radiologists, and Hinton’s “stop training radiologists” clip. There are more radiologists now
4	Definitions: intelligence = doing the right thing at the right time; AI = intelligence in an artifact
5	Teachers thought experiment. Same software, two outcomes. Secretaries and PCs. Stripped-down organisations; fragility
6	Work becomes less fixed; unions and freelancers; platform work; surveillance
7	Wages: Autor’s “Why are there still so many jobs?“. More AI, more jobs, but possibly more inequality
8	Bank tellers and ATMs; truck drivers and power steering; pay halved since the 1970s
9	Why you want people: diversity, Fisher’s theorem, and variation as insurance
10	Accountability: why a robot cannot be punished; dissuasion not recompense; sheep and guppies
11	Modularity: suffering could be removed. Shell companies. AI as legal person = ultimate shell company
12	Computation is physical. Agent vs moral agent vs moral patient. Moral subjects
13	Ethics is created by a society; identity signals; responsibility and intention as co-constructed; public goods
14	”We with AI are the intelligence explosion”. 12,000 years of AI (writing). Bostrom’s paperclips
15	The real unintended consequences: sustainability (megafauna decline) and inequality
16	Inequality and political polarization over 150 years. Reduced distance → fewer winners. Great “coupling” and the New Deal
17	Bretton Woods; transnational wealth extraction; need for a coalition including some powerful people
18	Her mathematical model: diversity raises the mean and the variance; decline pushes in-group cooperation; “gamble for redemption”
19	Chimpanzees; human multi-layered identity; identity politics as survival strategy
20	Society is richer than ever. Basic income: Standing’s pilots, her doubts, hidden “basic income” of public goods
21	What government is for (pick problems, allocate resources, stabilize solutions); corporations and governments police each other
22	Norway (no resource curse), Germany, UK, Wales; financial-sector regulatory capture
23	Recommendations: better at jobs, lower entry barriers, so benefits depend on governance; propaganda against governments
24	Climate, war as dismantling infrastructure; “AI is not magic”
25	Close: social connection, diversity, security; wealth sharing; transparency and accountability
26	One-to-one: inequality, war, policy vs technology
27	Global poverty fall; OECD up; France; China; effective altruism; information and mobile phones
28	Wages and AI; radiologists earn more; German arts and small shops
29	Q&A: why a broad AI definition; thermostats; plants; software ethics and due diligence
30	Q&A: bias, language, adoption metaphor, higher standard for AI, HR and diversity
31	Q&A: diversity vs bias; documentation; no Hippocratic oath; architecture and bank models
32	Q&A: identity politics vs diversity; large single donors; “a test of loyalty”
33	Q&A: robot tax? Wealth tax, digital tax, GDPR as a model, coalition of the willing
34	Q&A: is AI cultural imperialism? Shifting GDP shares; China and WeChat
35	Last question: do we need AI ethics? “AI ethics is just ethics”

Putting it together

Mental model. Bryson treats AI as the newest layer of an old human practice, building artifacts that carry out our goals, and argues that every consequence people fear (job loss, inequality, runaway systems, “robot rights”) is really a consequence of how people and institutions choose to deploy it. Intelligence is defined functionally; AI is “intelligence in an artifact”, so a maker always stands behind it. Because ethics is built by societies around who is responsible and who is owed care, and because accountability works through vulnerability that designed artifacts need not have, the right object of regulation is the human and corporate chain behind the system. The economic danger is not a shortage of work but a concentration of gains: technologies that cut the cost of distance create fewer winners, inequality feeds risk-aversion and in-group closure, and polarization follows. The remedy she offers is governance: taxes on wealth, transnational coordination on the GDPR model, audit trails and professional liability, so that the benefits of what a society collectively builds are shared.

Open tensions.

• Tool vs agent. “It’s just a tool” is her anchor, but it was said before large language models. Does the tool framing survive systems that act with wide autonomy?
• Optimism about jobs vs her own inequality worry. If wages fall as entry barriers fall (truck drivers), “more jobs” is a weak comfort.
• Strong-state premise. Most remedies need effective, uncaptured governments; she also says those governments are under propaganda attack.
• Speculation vs evidence. The loyalty-test and identity-politics readings are hypotheses she labels as such; the polarization model has a published paper but a contested causal story.
• Date. The talk is from May 2020. Everything after (generative AI, the AI Act, the 2021 tax deal) is outside it.

Reading path.

• Start: Autor, “Why Are There Still So Many Jobs?” (2015); Bessen, “Toil and Technology” (2015); the Stewart-McCarty-Bryson paper (2020).
• Then: Bryson, Diamantis and Grant, “Of, for, and by the people” (2017); Caliskan, Bryson and Narayanan (2017); Piketty (2013); Scheidel (2017); Wilkinson and Pickett (2009).
• Deeper: Acemoglu and Restrepo on tasks; Gunkel, Robot Rights (2018) as the opposing view; Bostrom (2014) for what she is answering; Bradford, The Brussels Effect (2020); Mazzucato (2013); Standing on basic income.

Part 1: Before you read

Key terms

• Intelligence (Bryson’s definition). “the capacity to do the right thing at right time” is how she glosses it in the caption (the phrase runs across a segment break; check the clip). She calls it a form of computation that takes information now and generates the correct action. She says it comes from the behavioural-science tradition she learned at Chicago, built “so that we could talk about which animals were more intelligent than each other”. Watch out: this is a functional, behaviourist definition. It deliberately excludes “souls or life”.
• Artifact / artificial. “artificial intelligence is intelligence in an artifact”: something “deliberately built”. The key word is deliberately. It puts a designer behind every AI, which is what lets her talk about responsibility and due diligence.
• Agent vs moral agent vs moral patient. Three different things.

  Term	Her sense	Where
  Agent (plain)	“something that changes the world”, so a mobile phone counts	chemistry-like use
  Moral agent	who is responsible for their actions, “determined by a society”	philosophy
  Moral patient	who we are obliged to take care of	philosophy
  Moral subject	the two together	philosophy
  English-language AI debate often slides between these. Much of the “robot rights” argument depends on which one is meant.

• Moral agency is society-relative. Different societies have different rules for when someone is an adult (age, or “whether your father is still alive and what gender you have”). She uses this to argue that ethics itself is “created by a society”.
• Dissuasion vs recompense. Justice, in her account, works to stop wrongdoing before it happens, not to repay victims after: “it’s actually more about dissuasion than recompense”. This is a standard idea in the deterrence tradition.
• Modularity. In engineered systems, parts can be swapped or removed. Her key use: any AI “suffering” you build in can be taken out again, “by a hacker” or by the owner.
• Shell company. A legal entity set up with no real stake in succeeding, often to dodge accountability. Her image for AI legal personhood.
• Fisher’s fundamental theorem. In evolutionary biology (R. A. Fisher, 1930), the rate of increase in a population’s fitness is proportional to its genetic variance. She generalizes it: “evolution goes faster the more variation you have”.
• Great decoupling / great coupling. Economists note US wages stopped following productivity from about 1978 (the “decoupling”). She points out the curve shows an earlier “coupling” under the New Deal.
• Resource curse. When a country with a valuable resource (oil) ends up more unequal and less democratic. Norway is the standard counter-example.
• Regulatory capture. An industry bends the rules that govern it. She says the UK’s version is by finance, not oil.
• Polarization. Bryson uses it for groups cooperating only inside their own group, not for disagreement over views.

People named (a reading list)

Person	Why they come up	Start with
Geoffrey Hinton	His “stop training radiologists” prediction is the foil	The 2016 talk (see Part 1 debate)
David Autor	”Why are there still so many jobs?”	Autor (2015), Journal of Economic Perspectives
James Bessen (not named by her)	The bank-teller/ATM evidence (a colleague at Boston University)	Bessen, “Toil and Technology”, Finance & Development (Mar 2015)
Donald Davidson	The philosopher’s idea of agent	Essays on Actions and Events (1980)
I. J. Good; Nick Bostrom	”Intelligence explosion”, paperclips	Good (1965); Bostrom, Superintelligence (2014)
Nolan McCarty and collaborators	The polarization/inequality model	Stewart, McCarty, Bryson, Science Advances (Dec 2020)
Jane Goodall	First documented chimpanzee “war”	The Chimpanzees of Gombe (1986)
Guy Standing	Argues for basic income, ran pilots	Basic Income: And How We Can Make It Happen (2017)
R. A. Fisher	Fundamental theorem	The Genetical Theory of Natural Selection (1930)
John McCarthy	Thermostats as AI	”Ascribing Mental Qualities to Machines” (1979)

Bryson’s own frame

Bryson is known for the position that robots should be treated as tools, not persons. Her early paper “Robots Should Be Slaves” (2010) argued that we should not build machines we owe moral duties to, and she has spent a decade warning against over-identifying with AI (Hertie School CV, June 2025, from memory for the paper title; not checked there). The 2017 paper with law professors that she cites, on “synthetic persons”, argued against legal personhood for AI. Her Science paper with Caliskan and Narayanan (2017) showed that word embeddings pick up human biases, which she cites. The HIIG event page describes the talk as arguing that humanity has “two problems to solve: sustainability and inequality” (HIIG event page; the exact phrase comes from a search summary, not read verbatim on the page). That matches the shape of the talk: names exactly these two.

The current debate (2025-26)

Nov 2026 is six years after the talk. What has happened to her main claims:

• Radiologists. Her lead example was that radiologist numbers rose after Hinton’s 2016 prediction. Later reporting supports this. Fortune (May 2026) quotes Hinton as saying the opposite has happened and there is a large shortage; the same coverage cites Mayo Clinic radiology headcount growing from about 260 in 2016 to over 400, and radiologist pay near $571k (Fortune, 4 May 2026; Fortune on salaries; Radiology Business). Caveat: many of these sources are trade or blog pieces. Check the original Medscape, Neiman HPI and Mayo data before quoting figures. Commentary also notes the task vs job point: AI automated parts of the job, not the whole job.
• Bank tellers. Her ATM story matches Bessen’s 2015 argument: tellers per branch fell (20 to 13 between 1988 and 2004) but urban branches rose 43%, so teller jobs held up (IMF F&D, Bessen). Bessen’s later point (not Bryson’s): teller employment finally fell in the 2010s because of mobile banking, a different technology.
• Her polarization model has been published. It appeared as Stewart, McCarty and Bryson, “Polarization under rising inequality and economic decline”, Science Advances, 11 Dec 2020 (vol 6, issue 50). The abstract says that under economic decline, in-group cooperation rises, and the effect can spread to the whole population and persist after the conditions end. The authors also found inequality and affective polarization correlated across US states (Bath research portal; PMC full text; Princeton release). So the “not peer-reviewed yet” status has changed. Whether the model is right is a separate question; correlation across states is not the same as the mechanism.
• GDPR as a model. Bryson’s praise of the GDPR is now tested by years of enforcement and by the EU’s later AI Act and digital-markets rules. I did not run a fresh search on current EU enforcement or on any rollback debate. Treat any claim about that as unverified. (Not checked.)
• The robot-tax and legal-personhood questions. The European Parliament’s 2017 “electronic personhood” idea she argued against was not adopted; since then, discussion has moved to model-welfare programmes at AI labs.†
• Hinton himself went from the 2016 radiologist forecast to leaving Google in 2023 to warn about AI risk, and won the 2024 Nobel Prize in Physics (shared with John Hopfield). From memory; the Radiology Business headline above refers to him as a Nobel winner.

The ideas

Bryson spoke for about 50 minutes to a general audience. She compresses seven bodies of thought. Here they are laid out. Where she does not name a thinker, I say so.

 LABOUR ECONOMICS ───────┐                       ┌──── PHILOSOPHY OF AGENCY & RESPONSIBILITY
 Keynes · Leontief       │                       │     Aristotle · Kant · Davidson
 Autor · Acemoglu        ▼                       ▼     Hart · Strawson · Latour
 Bessen              ┌────────────────────────────────┐
                     │ BRYSON: AI is an artifact; its │ ◄──── LAW OF PERSONS & LIABILITY
 EVOLUTION &         │ effects on jobs, inequality    │       corporate personhood · shell cos
 COOPERATION ──────► │ and ethics are decided by      │       deterrence theory · EU Parliament 2017
 Fisher · Hamilton   │ governance, not by the tech    │
 Axelrod · Boyd &    └────────────────────────────────┘
 Richerson · Henrich       ▲              ▲            ▲
                           │              │            │
   INEQUALITY & POLITICAL  │  AI SAFETY / │   STATE, TAX & PUBLIC GOODS
   ECONOMY                 │  SUPERINTELL.│   Mazzucato · Lamy? · Piketty
   Kuznets · Piketty ·     │  Good · Bost.│   Bretton Woods · Norway · GDPR
   Milanovic · Scheidel    │  Yudkowsky   │

Where each school of thought is covered:

• Labour economics: does technology destroy jobs?: idea 1
• Inequality and political economy: idea 5
• Evolution, diversity and cooperation: idea 6
• Polarization and identity politics: idea 5
• The philosophy of agency and moral status: idea 3
• Law: persons, corporations and shell companies: idea 3
• AI safety and the “intelligence explosion”: idea 4
• The state, tax and public goods: idea 6

1. Jobs, wages and who gets the productivity

In short. Bryson argues that AI does not decide whether jobs are lost. People do. Radiologists, bank tellers and secretaries show that more automation has often come with more jobs, but pay and power can still shift. Software that makes teachers twice as good could mean better schools or half as many teachers, and that is a political choice. Cutting staff can also leave an organisation more fragile.

🎙 In the talk.

• She opens with Norwegian radiologists, invited to an annual conference and expecting doom. Hinton “said there was no reason to train any more radiologists ever” yet “there’s more radiologists now than there were four years ago”, and she reads this as “AI has increased the value of radiologists”.
• The teacher thought experiment: software that makes teachers “twice as good” could mean twice as good schools or half as many teachers. “there’s nothing in that sentence that tells us about employment… this is a political decision or a normative decision”.
• The PC era: firms bought a PC, fired two-thirds of their secretaries and kept the best one, who could do more.
• She cites Autor’s “why are there still so many jobs” and says “more artificial intelligence than ever and we also have more jobs than ever”, but warns of inequality.
• Bank tellers: “there’s actually more bank tellers now than there were before automated teller machines” because cheaper branches meant more branches; tellers did harder work for slightly more money; the managers were the ones who lost out.
• Truck driving: “power steering is already translating the situation into action”, so driving was partly automated long ago, making the job easier to enter and cutting pay: “being paid half as much as they were in the 1970s”.
• A trade-off she names: fewer people can mean “a stripped-down organisation” with an easier management problem, but less diversity and more fragility.

📚 Deeper theory.

1. Task vs job. Autor’s framework (Roots) separates automating a task from eliminating a job. Radiologists read images, but also consult, supervise procedures, and write reports. This explains why the headline prediction failed. Later reporting repeats the point (Part 1).
2. Elasticity matters. Bessen’s finding depends on demand for the output being elastic. Cheaper bank branches raised demand for branches. If demand is saturated (food, perhaps), productivity gains do cut jobs. That is why bank tellers are a poor guide to every sector. (From Bessen’s argument; see the IMF article.)
3. Entry barriers and pay. The truck-driver story is a deskilling story: technology lowers the skill needed, widens the labour pool, and pay falls. It is close to Harry Braverman’s Labor and Monopoly Capital (1974) on deskilling, though she does not cite him.
4. Fragility from interfaces. Her point that a computer form as the only way into an organization stops humans solving problems together echoes the sociology of “street-level bureaucracy” (Lipsky, 1980) and later work on automated decision systems (Eubanks, Automating Inequality, 2018). She names neither.
5. The ATM example is old news. It was widely cited by 2015, so it works as a rebuttal to a naive claim more than as evidence for the future.

📚 Roots: Labour economics: does technology destroy jobs?

The problem. Every wave of machines raises the same fear: mass unemployment. The question is whether the fear has ever been right, and what actually changes.

The point. Bryson’s argument sits in the Autor-Bessen line: the number of jobs is rarely the issue, the distribution of pay and power is. Her teacher example is a plain statement of the idea that productivity gains can be taken as better output or fewer workers.

The chain of thinkers

1. Luddites (1811-16) and Ricardo’s “On Machinery” (1821). The early English textile workers broke looms. Ricardo, in the third edition of his Principles, changed his mind and said machinery could hurt workers. Key idea: gains from technology do not automatically reach labour.
2. Keynes, “Economic Possibilities for our Grandchildren” (1930). He predicted that by 2030 people would work around fifteen hours a week. Bryson refers to this prediction (“15 weeks a year or something by the year 2000” in her words). Key idea: technological unemployment is a “temporary phase of maladjustment”, and the long-run question is how to spend the leisure. He turned out right about wealth, wrong about hours.
3. Wassily Leontief (1983). He warned that humans could become like horses: displaced as labour inputs. Key idea: a technology can replace a whole factor, not just a task.
4. David Autor, “Why Are There Still So Many Jobs?” (2015). The paper Bryson recommends. It explains the puzzle with two ideas: complementarity (automation raises the value of tasks it does not do) and Polanyi’s paradox (we cannot automate tasks whose rules we cannot state). Autor also documents job polarization: growth at the top and bottom of the skill ladder, hollowing in the middle. This is the economics behind her worry that AI “may be making it harder to have a range of possible wages”.
5. James Bessen, ATMs and tellers. She credits “colleagues at Boston University”. Bessen’s mechanism: automation cut the cost per branch, so banks opened more branches, and teller jobs shifted to relationship work. It is a case of demand elasticity: when a product gets cheaper and demand is elastic, employment can rise.
6. Acemoglu and Restrepo. Task-based models (2018-19) split automation into displacement (machines take tasks) and reinstatement (new tasks for labour). Their warning, from memory, is that recent technology has been displacing faster than it reinstates. This is the strongest academic counterweight to Bryson’s optimism. Not checked against a specific paper.

Where her version differs. She adds two things economists mention less: diversity as insurance (idea 6) and fragility from single computer interfaces.

Where others push back.

• Acemoglu and Restrepo: automation can displace faster than new tasks appear (Roots). The 2010s mobile-banking decline in tellers shows a later technology can finish what the ATM did not.
• Critics say radiologists are a lucky case: a licensed, regulated, capital-protected profession. Many other jobs lack that cover.
• Frey and Osborne (2013) estimated about 47% of US jobs at high risk. Their method was criticized, but the point is that her optimism is one side of an open debate.

What they leave out.

• Generative AI (2022 onward) hit white-collar writing, coding and translation, which her 2020 examples did not cover. Her translator mention was about freelancing, not machine translation capacity. I do not make claims about 2025-26 labour data here†.
• Her story is about OECD labour markets. Informal work, which dominates India and much of the Global South, is outside it.

Check yourself.

1. In the teacher example, what exactly is the decision that determines employment, and who makes it?
2. Why did ATMs lead to more branches, and what would have to be different for tellers to lose jobs?
3. Is “wages not jobs” a stronger or weaker claim than “no job losses”? What evidence would separate them?

2. Definitions, and why she keeps AI so broad

In short. She keeps her definition of AI broad on purpose. Intelligence is the capacity to do the right thing at the right time, and AI is intelligence in an artifact: something deliberately built. Computation takes time, space and energy, so AI is physical. A broad definition makes it easier to talk about moral obligations. Her conclusion is that “AI ethics is just ethics”, and that makers should do due diligence as with any other product.

🎙 In the talk.

• Intelligence is “the capacity to do the right thing at right time” (as the caption splits it) and a form of computation. She chose it because it was made to compare animals, and it lets us ask what enhances human value.
• “artificial intelligence is intelligence in an artifact”, where an artifact is “something that’s been deliberately built”.
• “computation takes time space and energy”, so AI is physical, not abstract.
• In Q&A: asked why so broad, she says McCarthy said “thermostats are AI”, and on her definition “plants are intelligent” but “not very”. Reason: it is “the easiest way to talk about the moral obligations”.
• Digital technology will soon not be separable: it is going to be “harder and harder to differentiate” where AI stops and computing starts, so ethics should focus on software development and who decided what.
• She wants due diligence like other manufactured products; the car industry follows “pretty good DevOps”; search and social media should be treated “like a critical system”.
• On the moderator’s last question she says “AI ethics is just ethics” and “the fundamental problems are the fundamental problems”.

📚 Deeper theory.

1. Behaviourist and functional roots. Her definition is close to Turing (1950, behaviour as test), to McCarthy’s “ascribing mental qualities”, and to the comparative-psychology tradition (Thorndike; Morgan’s canon) that avoided talk of inner states. This is deliberate: it avoids the “soul” question.
2. Contrast: Legg and Hutter (2007). Their formal definition: intelligence measures an agent’s ability to achieve goals in a wide range of environments. Bryson’s is similar but stresses appropriate action in context.
3. Contrast: Searle’s Chinese Room (1980) and Dreyfus. Critics say action-matching is not understanding. Her definition sidesteps that: she does not claim understanding.
4. Why breadth helps her ethics. A broad definition covers everything from spell-checkers to learning systems, so no regulation can be dodged by arguing “that is not really AI”. It shifts the weight from the label to the artifact and its maker, which is an accountability move. A cost: a definition that includes thermostats says little about what is new.
5. The “AI ethics is just ethics” line has an old pedigree: it echoes the view that ethics of technology is the ethics of design and use, from the engineering-ethics tradition (Winner’s “Do artifacts have politics?” 1980, for the opposite emphasis on built-in values). From memory.

Where others push back.

• Many researchers say a definition so wide is useless for regulation: the EU AI Act spent years on its definition. Others argue that learning systems have different risk profiles from rule-based software (opacity, drift, emergent behaviour).
• The “just a tool” line (she uses “remember it’s just a tool”) is contested by people who think modern models act with enough autonomy that tool talk misleads.

What they leave out.

• Anything about how modern large models differ in kind from earlier software. In 2020 (before ChatGPT) this was less pressing.
• She does not discuss consciousness except to set it aside.

Check yourself.

1. Under her definition, is a thermostat intelligent? Why is that not a problem for her?
2. What practical work does “deliberately built” do in her argument?
3. Does “AI ethics is just ethics” imply that no new laws are needed? Why or why not?

3. Ethics, accountability and why AI is not a legal person

In short. She says an AI should not be a legal person. Law punishes mainly to dissuade, and what dissuades is what it is like to be a human who can lose time, status, wealth or liberty. A built system need not suffer, and a safe, modular one could have the suffering taken back out. Shell companies show the risk: AI as a legal person would be the ultimate shell company. Responsibility and intention are things a society builds together, not things science discovered.

🎙 In the talk.

• She starts with a puzzle: why not make the robot pay tax or be a legal person? Her answer: “it’s actually more about dissuasion than recompense”, and what dissuades is “what it’s like to be a human”, e.g. losing time, status, wealth, liberty.
• That vulnerability is “systemic”: “it’s systemic and it’s not just us”, extending to dogs, sheep and guppies (a separated guppy “can die of stress”).
• A well-designed artifact will not have that: “any system that we buy is built and designed”, and if it is safe and secure it is modular, so “if you put the suffering you could take the suffering back out”. She offers to discuss a human-brain upload in Q&A; the slides were cut.
• Corporate personhood is a warning: limited liability was over-extended after the 1990s, and shell companies are set up to fail. “AI as a legal person would be the ultimate shell company”.
• Agent, moral agent, moral patient definitions. Societies differ on adulthood, so “ethics itself is also created by a society”: “there’s nothing more fundamental to ethics than who are your moral agents who are your moral patients”.
• Identity signals (clothes, music, language) sustain a society and carry a moral component.
• “responsibility and intention are things that allow us to co-construct who we are”, and not things that science discovered. Government enforcement of them is a public good.

📚 Deeper theory.

1. Deterrence is the load-bearing idea. If you believe punishment works mainly by cost to the offender (Beccaria, Bentham, modern law-and-economics), then an entity that cannot suffer cannot be deterred. If instead you think law mainly compensates victims (corrective justice, Weinrib), then an AI with assets could be sued as a corporation is. Bryson takes the first view. The second is the usual route for those who favour personhood.
2. The “up-loaded brain” caveat. Her concession (“if you could magically upload a human brain then yes you would have something that suffers”) matters. It shows her argument rests on design contingency, not on a claim that suffering is impossible in silicon. This is a more careful position than “machines cannot be moral patients”.
3. Contrast with the relational view. Gunkel and Coeckelbergh say moral status comes from social relations, not inner properties. That is partly compatible with Bryson’s “created by a society”, but they reach opposite conclusions about machines: if society decides who counts, society could decide that robots count. She would say we should not make that choice, because it benefits the owners who offload liability. Her argument therefore turns on political risk, not on metaphysics.
4. Corporate personhood as precedent. The history (from Santa Clara County v. Southern Pacific, 1886 in US lore, to modern limited liability) is that legal personhood is a tool of convenience. That fits her use of it as a warning: tools of convenience can be abused.
5. Why she mentions Washington. She says she has been told “get rid of the people you can’t rely on the people they’re dangerous”. The ethical point behind it is that removing humans removes accountability points, not only error points.
6. Moral crumple zones. Madeleine Elish’s term (2019) for how humans in automated systems absorb blame for failures they could not control. It is the reverse risk of her argument: if humans are always the accountable party, system designers can hide behind a nominal operator. From memory; not in the talk.

📚 Roots: The philosophy of agency and moral status

The problem. Who can be held responsible, and who is owed care? Can a machine be either?

1. Aristotle, Nicomachean Ethics III. Voluntary action requires knowledge of what one does and not being forced. The root of “responsibility and intention”.
2. Kant. Moral agency needs rational autonomy; persons are ends in themselves. Mostly assumed, not argued, by AI-rights debates.
3. Donald Davidson (1963, “Actions, Reasons, and Causes”). An action is something done for a reason; reasons can be causes. Bryson mentions him to dismiss the sense of agent that “you care about” only if you read him. Her chosen sense is the chemistry-like, minimal one.
4. H. L. A. Hart, Punishment and Responsibility (1968), and Strawson, “Freedom and Resentment” (1962). Responsibility is a social practice built on reactive attitudes (blame, gratitude). This is close to Bryson’s claim that “responsibility and intention are things that allow us to co-construct who we are”, “not things that science discovered”.
5. Moral patients: Bentham and Singer. Bentham: “the question is not, Can they reason? nor, Can they talk? but, Can they suffer?” Singer extends moral standing to animals by sentience. Bryson’s moral patient is in this tradition, and her argument is that AI is not made to suffer in a non-removable way.
6. Moral status of AI: Bryson vs others. Her 2010 position (tools, not persons) contrasts with David Gunkel’s Robot Rights (2018), which argues the question of moral status should be asked relationally, and with Schwitzgebel and Garza (2015) on “a defense of the rights of artificial intelligences”.†
7. Latour’s actor-network theory. Non-humans as actants. Bryson does not use it, but the “mobile phone is a kind of robot” line has the same flatness.

The core of her argument. There are two claims, and they are separate.

• Descriptive: what counts as a moral agent or patient differs by society, so ethics is made by societies.
• Design: AI is designed. If the suffering that would make punishment bite were built in, it could be removed. So AI cannot be the locus of deterrence the way a person is.

📚 Roots: Law: persons, corporations and shell companies

The problem. Can the law hold a non-human to account? It already does for corporations. What breaks if AI is next?

1. Corporate personhood. Limited liability and legal personality were created so groups could pool capital and sue and be sued. Bryson: “perfectly good reasons” but “since the 1990s or so it’s been over extended”.
2. Shell companies. Entities with no stake in success, used for money laundering and evasion. Her phrase: AI as legal person “would be the ultimate shell company”.
3. Deterrence theory in criminal law (Beccaria 1764; Bentham). Punishment works because of its cost to the person. This is why she says justice is “more about dissuasion than recompense”.
4. European Parliament, 2017: “electronic personhood”. A resolution on civil-law rules for robotics floated a status for autonomous robots. Bryson and others signed an open letter against it†. Her 2017 paper with law professors (Bryson, Diamantis, Grant, “Of, for, and by the people”) is the academic form of this argument.
5. Product-liability and professional-liability models. She rejects the “Hippocratic oath” idea for programmers because “we don’t have a one-to-one relationship with them” as with doctors or lawyers; she prefers the bank-and-rogue-trader and architecture models: layers of licensing, inspection and institutional liability.
6. Due diligence and audit. She argues for documentation: who wrote the code, why, and whether an unfair result was intended, hacked, or an insider’s act. This is the logic behind software supply-chain audits and, later, the EU AI Act’s documentation duties†.

Where others push back.

• Philosophers such as Schwitzgebel/Garza and Gunkel say a future AI might deserve moral standing and refusing in advance is a moral risk.
• Legal scholars (e.g. Chopra and White, 2011) argued for limited legal personality for autonomous agents on practical grounds. The 2017 EU Parliament debate shows the idea had traction.
• Some say deterrence of developers and firms is the real target, which she agrees with; so the dispute is smaller than it looks, centred on whether to add an AI entity in the middle.

What they leave out.

• Non-Western views of moral status. She does show that adulthood differs by society, but she does not discuss traditions where non-humans (rivers, ancestors, spirits) have standing. The Whanganui river and Indian court rulings granting personhood to the Ganges (2017) are counterexamples to “only humans”. I add them; they are not in the talk, and the Indian ruling was stayed by the Supreme Court†.
• How to handle AI systems that operate across jurisdictions.

Check yourself.

1. Why does she care that the AI’s suffering can be removed?
2. What does the corporate “shell company” tell us about what could go wrong with AI personhood?
3. If you held the corrective-justice view of law instead, would her argument still go through?

4. The intelligence explosion, and the harms already here

In short. She answers the “intelligence explosion” worry by saying that we, with AI, are the explosion, and have been for about 12,000 years if writing counts as AI. The paperclip story is about unintended consequences, and she says those have already happened. Human growth came with a drop in every other large animal, and with problems of sustainability and inequality. Her close: “we author AI”.

🎙 In the talk.

• A “final invention” worry from I. J. Good, then Bostrom “made it sound like a bad thing”.
• Her answer: “we with AI are the intelligence explosion”, with a double-exponential curve of human population and technology. Writing is counted as AI: “something like 12,000 years of AI”.
• Bostrom’s paperclip story: unintended consequences, e.g. “a planet might be turned into paperclips” from a request to organize offices.
• Her counter-claim: unintended consequences have already happened. Human growth has gone with “a decrease in the number of every other large animal”. Biomass is higher now but is “almost entirely things we eat and play with”.
• Two such consequences: sustainability and inequality.
• Close: “AI is not magic it’s something that people make”, “we author AI”.

📚 Deeper theory.

1. The evidence she is drawing on. The megafauna decline is documented in work on the Quaternary extinctions (e.g. Barnosky et al.; the human-arrival drops she describes). The “biomass of mammals” chart is from Bar-On, Phillips and Milo (2018), popularized in a well-known xkcd-style graphic (she says “the xkcd version”).†
2. The Anthropocene frame. She is close to the argument that humans plus technology are already a planetary force; the question is whether the system is steered.
3. Is it a good analogy? Bostrom’s concern is a single optimizing agent with misaligned goals. Hers is a distributed socio-technical system with no single goal. Both can cause harm, but the remedies differ: alignment research vs institutional reform. Her argument shifts the burden to the second.
4. Her “AI = writing” move has an ancestor in Andy Clark’s natural-born cyborgs (2003) and in Leroi-Gourhan and Stiegler’s idea of externalized memory: tools extend cognition. It makes the explosion look continuous rather than a threshold.
5. Later events. After 2022, many researchers argued that frontier models change the picture. I make no claim here about those debates†.

📚 Roots: AI safety and the “intelligence explosion”

The problem. Will machines that improve themselves run away from human control?

1. I. J. Good (1965). The “ultraintelligent machine” would be “the last invention that man need ever make”, provided it is docile. Bryson quotes the gist.
2. Nick Bostrom, Superintelligence (2014). Orthogonality (goals independent of intelligence) and instrumental convergence; the paperclip maximizer thought experiment.
3. Yudkowsky and the alignment community. The worry of misaligned objectives at scale. Bryson does not engage them by name.
4. Bryson’s reframe. “we with AI are the intelligence explosion”: the explosion is the last ~12,000 years of human artifact-aided growth (writing counts as AI in her definition). Bostrom’s feared “unintended consequences” have already occurred: biodiversity loss and inequality.

Evaluating the reframe. It is rhetorically strong and moves attention to present harms. A critic would say it flattens a real difference: past tools did not design their own successors or act autonomously at speed. She would reply that the harms come from human institutions deciding how to use them. Both views are still live.

Where others push back.

• Alignment researchers say present harms and future risk are not exclusive; collapsing them lets us ignore a qualitatively new risk (systems that outthink their overseers).
• Some say human population growth plus technology is not an “intelligence explosion”; no recursive self-improvement is shown by the curve.

What they leave out.

• A treatment of the arguments for fast take-off (e.g. Bostrom’s, Yudkowsky’s), which she dismisses with the reframe rather than rebuts.
• Any estimate of the probability of misalignment risk.

Check yourself.

1. What does it mean to say “we with AI are the intelligence explosion”?
2. What are the two “unintended consequences” she names, and why are they the same kind of thing Bostrom fears?
3. What would a critic say the double-exponential curve does not show?

5. Inequality, polarization and her model

In short. She shows US data in which political polarization follows the share of income going to the top 1%. Her hypothesis is that a technology which cuts the cost of distance leaves fewer winners. Her model says diversity widens the range of outcomes, so in bad times people cut risk or gamble. Identity politics may be a survival strategy for hard times. She calls this informed speculation, and says the model is not yet peer-reviewed.

🎙 In the talk.

• She shows US data in which political polarization tracks the share of income going to the top 1%, high before WWI, low mid-century, high again.
• Hypothesis: “when you introduce a new technology that reduces the cost of distance then you naturally get fewer winners”. Illustration: beer on bicycles across Berlin means fewer pubs win.
• The curve falls after the New Deal and Bretton Woods: “the political polarization of the Congress was brought down first” and then inequality. Enough elites “signed up with the proletariat”; transnational wealth extraction was banned “for a few decades”.
• Model: diversity raises the expected outcome but “the variation in the outcome is wider”, so when things go badly you cut risk, even at lower expected return. Under bad conditions, “you gamble for redemption”. She says the result is not peer-reviewed yet.
• Chimpanzees split and fight when food is short. Humans have layered identities (family, neighbours, coworkers), so identity politics may be “a survival strategy that we had for dealing with hard times”, the opposite of public goods.
• On world data: inequality fell globally while OECD countries became more unequal; China has rising inequality but stability, which she explains in terms of the bottom still rising.
• Q&A: single large donors and extreme politicians; extreme statements are “a test of loyalty”. She calls it “informed speculation”.

📚 Deeper theory.

1. What the model actually claims. (See Roots.) It is a formal argument that risk-sensitivity can generate in-group closure without any hatred or ideology. That is a strong and interesting result; whether it explains real polarization is empirical. The authors link it to state-level US data, a correlation across states (PMC).
2. Alternative explanations of the same U-curve. Party-system and media changes (Prior; Gentzkow-Shapiro), geographic sorting (Bishop’s The Big Sort), race and immigration politics, and the decline of unions and cross-cutting institutions (Putnam’s Bowling Alone). Her story is an economic one; many of these are not.
3. Piketty vs her technology story. Piketty attributes the post-war fall largely to shocks and to progressive taxation; she wants to show policy matters beyond war. The “great coupling” point (wages tracking productivity) is her own reading of the chart.
4. “Reducing the cost of distance.” Compare “superstar” economics (Rosen 1981) and later work on globalization. It predicts concentration among suppliers (pubs, platforms), not across households, so it is an incomplete account of personal-income inequality.
5. China and Germany as stable-despite-inequality. She says stability holds while the bottom is rising; if distribution stops, it could fail. This is consistent with the “J-curve” idea in the revolution literature (Davies 1962): unrest comes when rising expectations are frustrated. Not named by her.

📚 Roots: Inequality and political economy

The problem. Why does inequality rise and fall in waves, and what does technology have to do with it?

1. Simon Kuznets (1955). The “Kuznets curve”: inequality rises in early industrialization, then falls. Later data broke the neat shape.
2. Thomas Piketty, Capital in the Twenty-First Century (2013). The U-shaped top-1% share in rich countries: high before WWI, low mid-century, high again. This is the curve Bryson shows. Piketty’s core claim is that when returns to capital exceed growth (r > g), wealth concentrates.
3. Walter Scheidel, The Great Leveler (2017). Over the long run, large reductions in inequality came from violent shocks: mass-mobilization war, revolution, state collapse, plague. Bryson addresses this directly: “you only drop inequality when there’s a war” is the view she pushes against, pointing at the US New Deal and Bretton Woods as policy-led.
4. Branko Milanovic, Global Inequality (2016) and the “elephant curve”. Global inequality between countries fell (China, India) while inequality within rich countries rose. Bryson says exactly this: world inequality falling, OECD inequality rising.
5. Goldin and Katz, The Race between Education and Technology (2008). Technology raises the demand for skill; whether wage inequality widens depends on whether education keeps up. This is a rival mechanism to Bryson’s “reducing the cost of distance”.
6. Her own mechanism, reducing the cost of distance. When transport or communication lets one supplier reach everyone, a few winners take most. The beer-on-bicycles example is winner-take-most logic, close to Sherwin Rosen’s “economics of superstars” (1981) and to network-effect arguments. She presents this as a hypothesis (“we’re not sure this is working progress”).

The thread. Piketty and Scheidel describe the pattern. Autor and Goldin-Katz describe labour-market mechanisms. Bryson adds polarization as the political consequence and offers a cooperation-based explanation (Roots). Her policy implication: since the wars-only story would be hopeless, find the cases where policy did the work (New Deal, Bretton Woods, Norway).

📚 Roots: Polarization and identity politics

The problem. Why does inequality go with political polarization?

1. The correlation. McCarty, Poole and Rosenthal’s work on US congressional polarization tracks the same U shape as top income shares. Bryson’s slide overlays these.
2. Her model (Stewart, McCarty, Bryson). Agents choose whom to cooperate with. Out-group cooperation has a higher expected payoff but higher variance; in-group has lower mean, lower variance. If the environment worsens, agents shift in-group. At extreme decline they gamble. The published version also shows the effect can persist after conditions improve (PMC).
3. Identity politics as survival strategy. Rival readings to compare: Francis Fukuyama’s Identity (2018) treats identity politics as a demand for recognition (thymos), not for resources; Robert Putnam’s work on diversity and trust (E Pluribus Unum, 2007) finds short-run trust falls; Katherine Cramer’s rural-resentment work stresses perceived neglect. Bryson’s resource-competition story is one of several and she calls it speculative.
4. Large donors. She cites papers on single big donors pulling politicians to extremes, then adds her own “theoretical biology” extension: extreme statements are a costly loyalty test that tells a donor who owns the politician. Her words: “informed speculation”. Treat it as a hypothesis.

Two things not to conflate. Polarization in her model = group closure in cooperation. In US political science, “affective polarization” = dislike of the other party. The paper links them with state-level data, but they are not the same measure.

Where others push back.

• Scheidel: war and catastrophe have been the main equalizers; policy alone rarely did it. Bryson says Germany and the UK both implemented a German model and that welfare ideas pre-dated the war.
• Methodological: the claim that inequality causes polarization rests on correlation, and the causal arrow might run the other way (polarized legislatures make redistributive policy harder).
• Her chimpanzee analogy: inference from animals to modern politics is looser than a model result. She treats it as an illustration.

What they leave out.

• The role of race, religion and colonial history in identity politics.
• Evidence from the Global South, where inequality and conflict follow different patterns.
• Any test of her donor/loyalty speculation beyond what she says.

Check yourself.

1. In her model, why does a worse economy reduce cooperation with out-groups even when that lowers the expected payoff?
2. What is “reducing the cost of distance”, and how does it make fewer winners?
3. Why is “identity politics” read here as an economic behaviour, and what other readings exist?

6. Governance: what government is for, and what to do

In short. She says government picks problems to solve, allocates resources and stabilises solutions. How AI’s benefits are shared is a question of governance, not of AI. Norway invested its oil wealth for everyone, while the UK shows hollowed-out towns. She argues that without effective governments it is hard to solve other problems, and that the crisis of democracy and governance is one of the biggest.

🎙 In the talk.

• A French official’s three-part definition: “the first thing government does is pick problems to solve”, then allocate resources, then stabilize solutions.
• Corporations and countries both have governments, and “we want our corporations to help us police our governments” and vice versa. (Exact words at the clause level: “to help us police our governments”.)
• “we already are getting a massive basic income” in the form of streets, water, security, though unevenly.
• Norway “how not to experience a resource curse” because oil wealth was invested for all. The UK shows hollowed-out towns; “it’s not regulatory capture by oil but apparently by the financial sector”. “the real tragedy of brexit” is school and service cuts in places like rural Wales.
• Recommendation: AI will make many of us better at our jobs, which can lower entry barriers and salaries, so how benefits are distributed “isn’t anything to do with AI it’s about governance”. She cites propaganda against successful governments: “the crisis of democracy and governance is one of the biggest ones” and “without effective governments it’s hard to solve the other problems”.
• Hypothesis: some powers dislike regulation of tobacco, money laundering, NATO, the EU, climate action. She quotes “one definition of war is dismantling your enemy’s infrastructures” and says people may later see this period as already a war.
• Final recommendations: use “technology including the law… to increase transparency and accountability not confuse those”; help people understand the system; think about “how we share wealth through the system”.
• Q&A: she rejects the robot tax because “robots aren’t countable” and a programmer can change the robot to evade the tax; she mentions the window tax as a precedent. She favours “the wealth taxes”, a transnational digital or general tax, and mentions work with the European Commission.
• GDPR: firms said it “could be such a disaster” but then found “a single API to the whole of Europe”. Chinese colleagues like it because “it tells you exactly how much it costs if you do the wrong thing”. Like arms negotiations, “it’s not to prevent all war it’s to clearly communicate how much people care about things”.
• She suggests a coalition of the willing outside the EU if Ireland and Luxembourg block a digital tax.
• On professional ethics: no Hippocratic oath; prefer bank/architecture models with licensing and inspection; “that level of professionalism”; and “we need to be able to see the documentation”.

📚 Deeper theory.

1. Landau/Lamy and state functions. The definition she attributes to a French official is a minimal functional theory of the state: problem selection, resource allocation, stabilization. I could not identify him (see warnings). It resembles Charles Lindblom’s “science of muddling through” in spirit, but that is my comparison.
2. Hidden basic income. Her claim that public goods are already a basic income recasts the debate. The counter-argument is that services in kind and cash transfers differ in autonomy: the first constrains, the second frees (a standard point in the UBI literature, including Standing).
3. UBI evidence. Pilots (Kenya GiveDirectly, Finland, Stockton) have reported modest, mostly positive effects on wellbeing, with disputed effects on work. Her objection that single-community pilots can create cohesion by identity is a methodological point about external validity. I did not check recent pilot results.
4. Regulatory capture is Stigler’s 1971 concept. Her Norway/UK contrast suggests it is not oil per se but weak rules that decide outcomes.
5. The GDPR and the “Brussels effect”. Bradford (2020) argues EU rules spread because large markets can set standards. Bryson’s version adds a business argument: clarity of rules helps firms. That framing, regulation as predictable pricing of wrongdoing, is close to Pigouvian thinking (penalty = cost). Her arms-control analogy uses the same signalling logic as the Roots in this idea.
6. Taxing robots vs taxing wealth. Economists disagree. Opponents of a robot tax note the definition problem she raises; supporters (e.g. Gates; and some work by Guerreiro, Rebelo and Teles, 2017, from memory) argue the optimal tax code should tax automation if labour is displaced. Her line that tax is easier on observable wealth parallels the “tax people, not things” principle.
7. Transnational tax coordination. The practical history of the OECD’s global minimum tax came after this talk (2021). Not checked here.
8. Architecture/finance as models for software. Both are cases of layered liability: firms, individuals, regulators, licensing. The point is that no one person is “the” responsible party. Compare the later EU AI Act, which assigns duties to providers, deployers and importers†.

📚 Roots: Evolution, diversity and cooperation

The problem. Why would a society (or a firm, or a species) want people who think differently? Her answer comes from biology, not from fairness.

The point. Diversity is not decoration; it is a hedge against a changing environment. That is why she opposes “get rid of the people” and why she warns that a stripped-down organisation is fragile.

1. R. A. Fisher (1930). The fundamental theorem: the rate of fitness gain tracks the variance available. Bryson: “evolution goes faster the more variation you have”. She stresses that machine learning works much like evolution in this respect.
2. Scott Page, The Difference (2007). The “diversity trumps ability” theorem for problem-solving groups, under some conditions. Not named by her; she cites board-diversity findings. The empirical evidence on board diversity is mixed.†
3. William Hamilton (1964), Robert Trivers (1971), Axelrod (1984). Kin selection, reciprocal altruism, and the iterated prisoner’s dilemma: how cooperation survives. Her in-group/out-group model draws on this family (evolutionary game theory).
4. Boyd and Richerson; Joseph Henrich. Cultural evolution: norms and identity markers spread because they help groups cooperate. Her point that “defining your identity… with your music and your culture and your language” sustains a society is cultural-evolution reasoning. Costly signals of loyalty (her “test of loyalty” idea) come from the signalling theory of Zahavi and Spence.
5. Risk-sensitive foraging and prospect theory. Her model’s key move: when you are poor, you cut risk even at lower expected return, and when things are desperate, you “gamble for redemption”. That is the biological risk-sensitivity result (Stephens, Caraco) and Kahneman-Tversky’s finding that people turn risk-seeking when facing losses. She does not name them; I add them as the closest relatives.
6. Chimpanzees and Goodall. Conflict between groups under resource scarcity is documented in other primates. This is her bridge from biology to politics.

A caution she raises herself. She says it is hard to discuss biology in political science departments (she reports being told psychology cannot be included “because psychology involves biology and biology and false genetic determinism”) and adds “I’m not a determinist”.

📚 Roots: The state, tax and public goods

The problem. Who pays for the shared systems technology depends on, and how can a single state tax a company that sells everywhere?

1. Public goods (Samuelson 1954; Olson 1965). Non-excludable benefits are under-provided without coordination. Her list of “streets”, “water”, “security” as a hidden basic income uses this idea.
2. Bretton Woods (1944). Capital controls limited “transnational wealth extraction”. Dismantled in the 1970s-80s. (Her date of 1945 is a slip.)
3. Resource curse and Norway (Sachs and Warner; Norway’s sovereign fund).
4. Wilkinson and Pickett, The Spirit Level (2009). The graph she shows: inequality vs health and social problems. Her caption gives “2011” for the data, which fits the book’s later edition. The causal reading is contested by critics. From memory.
5. Basic income: Guy Standing and Philippe Van Parijs. Standing’s pilots (India, Namibia) reported community benefits. Bryson’s doubts: pilots on one sub-community can create identity and cohesion by themselves, long-run effects are unknown, and employment creates “interactive links”.
6. Mariana Mazzucato, The Entrepreneurial State (2013). The state funds much of the research that private firms later monetize. Bryson’s closing point, “how the system provides the materials that produce the AI”, is Mazzucato’s logic. She does not name her.
7. Taxes. Robot tax (rejected): Bill Gates floated it in 2017; Bryson argues robots are not countable so the tax is evadable, with the window-tax analogy. Wealth tax (supported): see Piketty and Saez-Zucman. Digital services tax (supported but blocked by Ireland and Luxembourg); she proposes a general transnational tax and a coalition of the willing. The OECD’s later two-pillar global minimum-tax deal (2021) is the real-world descendant.†
8. GDPR as proof of concept. Her claim: firms predicted disaster, then found “a single API to the whole of Europe”. Compare Anu Bradford’s “Brussels effect” (2020), the idea that EU rules become global standards because firms prefer one compliance regime. Not named by Bryson.

The thread. Idea 6 is where her ethics turns into a program: since ethics is made by societies and AI is an artifact, the lever is governance, and governance needs money and coalitions.

---

Where others push back.

• Public-choice theorists doubt that governance improves just because we need it; capture by large firms is exactly the risk.
• Industry voices argue GDPR raised compliance costs and entrenched large incumbents; there is empirical work on this†.
• Others argue her faith in “coalitions of the willing” underestimates how hard tax coordination is.
• Left critics want stronger measures (public ownership of data infrastructure); libertarians want less.

What they leave out.

• India and other large democracies with weak enforcement capacity. A strong-state premise underlies her policy list.
• How to govern AI inside firms that are stateless in practice.
• Power asymmetries between governments and platforms that she mentions only briefly.

Check yourself.

1. What three jobs does she say government does, and which one does she think was disrupted after 2008?
2. Why does she reject a robot tax but support a wealth tax?
3. What does her arms-control analogy say about what regulation is for?

7. Bias, diversity and what AI can fix

In short. In Q&A she says all technology expresses power structures. Her 2017 paper showed that sexism and racism are uploaded with language, because machines learn words as people do. Since AI is built on purpose, it can be held to a higher standard, like adopting a child rather than having one. AI can reduce or increase bias because it is just a tool, so we need to see the documentation.

🎙 In the talk (Q&A).

• Asked why AI discourse treats AI as neutral, she says all technology expresses power structures, though the mobile phone overturned some.
• She cites her 2017 paper, which “showed that sexism and racism are uploaded with language”, listed under cognitive science, not AI, because the machine learns words as people do.
• Metaphor: “having your own children the biologically versus adoption”. Because AI is built deliberately, “of course they can be held to a higher standard”.
• Anecdote: HR people say AI is helping them find diversity they had missed, but will not say so on record for fear of being sued. She says she cannot document it.
• AI can reduce or increase bias: “remember it’s just a tool”. So “we need to be able to see the documentation”.

📚 Deeper theory.

1. The 2017 paper. Caliskan, Bryson and Narayanan, Science 356 (2017): word embeddings reproduce human implicit associations measured by the Implicit Association Test (e.g. flowers/insects with pleasant/unpleasant, names with gender and race). From memory of the paper; not re-checked.
2. Human vs machine bias. The adoption metaphor is a claim about auditability: a machine’s rules can be inspected and revised in ways that a person’s habits cannot. Sendhil Mullainathan has argued something similar: algorithms are easier to audit than people (from memory).
3. Counter-evidence. Amazon’s scrapped hiring tool (2018) and the COMPAS debate (ProPublica 2016) show systems that scaled bias. Fairness research (Kleinberg et al., 2016; Chouldechova, 2017) proved that common fairness criteria cannot all hold at once, so “fixing” bias means choosing which one. Bryson’s “if those are your priorities” fits this: it is a values choice.
4. The higher standard. Her point that self-driving cars should be “a hundred times safer” is a risk-ethics question: people accept human error but not designed error. It parallels the debate over autonomous weapons and vehicle regulation.
5. Documentation as a mechanism. Model cards (Mitchell et al., 2019) and datasheets for datasets (Gebru et al., 2018) are the documentation she asks for; later laws require similar records. She does not name them.

Where others push back.

• Critics say “fixing bias” with technology treats a structural problem as a technical one (Benjamin, Race After Technology, 2019).
• The HR anecdote is unsourced by her own admission; treat it as a lead, not a finding.

What they leave out.

• Who audits and who has standing to challenge.
• Intersection with caste, a central axis of bias in India (not discussed; my addition).

Check yourself.

1. Why does she say the machine learning words is “not that different” from people hearing them?
2. What does the adoption metaphor claim about the difference between engineered and inherited bias?
3. Why does documentation make the “tool” claim testable?

---

About the source and how this page was made

• Speaker: Joanna J. Bryson
• Event: “Making Sense of the Digital Society” (HIIG + bpb lecture series, Berlin). Talk title on the HIIG page: “The role of humans in an age of intelligent machines”. Held in an empty theatre as a livestream (COVID-19 lockdown).
• Date: 2020-05-25 (from the HIIG event page, 19:00-20:30; the talk itself refers to “the last two months of a rather loose lockdown”)
• Uploaded: 2020-06-03
• Length: 1:31:52 (talk about 50 min, then a short one-to-one, then Q&A from Slido and Twitter)
• Channel: Alexander von Humboldt Institut for internet and society (HIIG)
• Video: Joanna Bryson: The role of humans in an age of intelligent machines

A note on tone. This is a plain-spoken talk by a working scientist who mixes three jobs in one hour: a labour-economics argument, a moral-philosophy argument, and a political proposal. The page keeps them apart because Bryson’s own claims are of different strength. The model of polarization is her own research (and she says it is “not peer-reviewed yet”). The ethics definitions are philosophy. The tax proposals are opinion.

Speakers

Label	Who	Role / confidence
Bryson	Joanna J. Bryson	Professor of Ethics and Technology, Hertie School, Berlin. Named in the introduction. High confidence (also matches the HIIG page).
Moderator	”Toby Mueller” (as captioned; the real spelling is probably Müller)	Freelance journalist of Swiss background in Berlin. Says “I am the moderator”. High confidence it is one person; spelling unverified.
Slido reader	”Christian call focal” (caption garble; probably Christian Katzenbach of HIIG, not verified)	Reads audience questions from Slido and Twitter during the one-to-one and Q&A. Medium-low confidence on the name.
Audience	Remote. Questions come via Slido (Martin Schmidt; “Salman”; others) and Twitter	Not on camera

Bryson’s bio, as read out by the moderator: degrees in psychology and behavioural science (Chicago), AI (Edinburgh), and a doctorate at MIT; 18 years in the computer science department at Bath; AI consultant to Lego “in 95”; adviser to legislators on AI and robotics.

Caption warnings. Auto-generated captions with many errors. Corrections used on this page:

• “Hurdy school” / “Murari” = Hertie School
• “AU and other values” = HIIG series context, garbled
• “David Otter / David Archer” = David Autor; “Mel and McCarty” = Nolan McCarty (co-author, see Part 2)
• “Landau” is an unclear name. The speaker means a French official who was in the room “when the EU was like seven guys”. Unidentified; do not rely on the name. (Pascal Lamy is one guess; not checked.)
• “guy standing” = Guy Standing, the basic-income economist
• “Boström” = Nick Bostrom; “I J good” = I. J. Good
• “slide oh” = Slido; “Alex TV” = the Berlin broadcaster Alex
• “Donald Davidson” is correct (the philosopher of action). “Fartknocker” is a garble of the German Datenschutz-Grundverordnung (the GDPR).
• “weight loss” = wage loss (moderator). “Chimpanzee… wipes out the other”: fine.
• “Matthew/Mike we pay tax” is a garble for a joke aside about companies paying tax.
• The captions hold a Hinton video clip that is part of the slides, not live speech.

Factual slips in the talk (checked or flagged):

• “five billion people… moved out of extreme poverty in the last 20 years”. This is too high. World Bank extreme-poverty counts fell from roughly 1.9 billion (1990) to roughly 650-700 million (2018); the drop since 2000 is closer to one billion than five. Probably she means the number of people lifted above some poverty line, or it is a slip. From memory of World Bank headline numbers; not re-checked here.
• Bretton Woods “in 1945”. The agreement was reached in July 1944. Minor.
• “At the end of World War two the GDP of the United States was greater than the rest of the world combined”. Contested. US share of world GDP in 1945 is often put near one-half in rough accounts, but Maddison’s series gives about 27% for 1950. Treat as an overstatement of a real fact (the US was by far the largest economy).†
• “Geoff Hinton said four years ago there was no reason to train any more radiologists ever”. Hinton’s remark was in 2016 and was more hedged (“people should stop training radiologists now”). Close enough. Her claim “there are more radiologists now” is supported by later reporting (see Part 1).
• “John McCarthy was the first person who said that thermostats are AI”. McCarthy’s 1979 paper “Ascribing mental qualities to machines” does use thermostats. “First” is not checked.

---

🎙 marks what the speaker says. 📚 marks added background and the wider debate, not the speaker’s claim. † marks background from general knowledge that was not checked against a source.