Apocaloptimist: Why the Barriers to AI Abundance Are Legal, Not Technical

Image from Pinterest

I. The Chasm

In 2024, the Nobel Prize in Chemistry went to the team behind AlphaFold—an AI system that predicts protein structures with near-experimental accuracy, solving a problem that had resisted fifty years of scientific effort.¹ What previously took years of laboratory work now takes seconds at negligible marginal cost. Drug candidates that once required billions of dollars and more than a decade of human research are being generated in months at a fraction of that cost. The AI pharmaceuticals market, valued at $1.8 billion in 2023, is projected to reach $13.1 billion by 2030, with AI workflows compressing discovery timelines from five years to as little as twelve months and reducing costs by up to forty percent.² 

And yet, against this backdrop of accelerating capability, we have failed at something far simpler: treating diseases we have known how to cure for half a century. Tuberculosis has been curable since the 1950s. Since then, it has killed more than 150 million people. It still ranks among the top ten causes of death globally, claiming over a million lives every year.³ The bacterium that causes it can be killed. The drugs are cheap. AI is now generating new drug candidates at negligible cost. At this point in the twenty-first century, it is no longer accurate to say that tuberculosis is caused by a bacterium, because we have known how to kill that bacterium for generations. In the twenty-first century, tuberculosis is ultimately caused by us. By human choices, and legal and economic systems that have decided, in effect, that some lives are worth protecting with monopoly pricing and some are not. 

The case of TB is one illustration, not the whole argument. This article is concerned with something larger: wherever AI and other forms of technologies of abundance collapse the cost of production (e.g., in manufacturing, logistics, knowledge work, scientific discovery) the productivity gains from that collapse, under the status quo, will be captured by large corporations and capital owners, exacerbating existing wealth disparities rather than producing a more equitable distribution for the people in need and whose labour, data, and publicly funded science made those systems possible. As technology inevitably continues to develop, one statement becomes increasingly true: the barriers in crossing the chasm from a world of scarcity to the promised land of post-scarcity are not technological in nature but socio-legal. The barriers that stand in the way between the world we inhabit today and the world that AI and technologies of abundance make possible for us are inherently political in nature. Given this premise, the central question that this article considers is how we can navigate the rise of AI to produce more equitable outcomes and how we can harness these technologies of abundance to actually translate to a broadly shared economic abundance—one in which access to basic goods, healthcare, and the material preconditions for a dignified human life are not contingent on geographic accident, market position, or the legal architecture governing who owns the productive outputs of automated systems.

II. The AI Race 

One of the most pressing issues surrounding AI is that most of the AI being built is not of the same character as AlphaFold. The overwhelming majority of the trillions of dollars currently flowing into AI development is directed toward a race to artificial general intelligence (AGI)—AI systems capable of performing any cognitive task at superhuman speed and scale. In 2024 alone, investment commitments to frontier AI labs reached figures that would have seemed implausible a decade prior: Microsoft committed $13 billion to OpenAI, and the Stargate project announced in January 2025 projected $500 billion in AI infrastructure investment over four years.⁴ Large AI labs are operating on maximum incentives to reach recursive self-improvement first—the point at which an AI system performs the research and analysis necessary to develop by teaching itself, potentially triggering an intelligence explosion that surpasses human understanding across every domain simultaneously.⁵ The commercial logic is not difficult to understand. The firm that reaches that threshold first captures a prize of almost unimaginable magnitude. But that logic produces a particular kind of AI—powerful, general, enormously expensive, and optimised for commercial return rather than human need. In 2023, approximately 70% of AI investment was directed toward general-purpose foundation models, with a fraction flowing toward domain-specific applications in areas like healthcare and drug discovery.⁶ 

AlphaFold represents a fundamentally different approach. It is a narrowly scoped, domain-specific system with a concentrated application: predicting protein structures. Its training data was publicly available, and its outputs were released openly to the global research community. Precisely because of those constraints—its limited scope, its public inputs, and its open outputs—it has generated more targeted, demonstrable benefit per dollar of development cost than almost any AI system in history, while carrying substantially less systemic risk than general-purpose frontier models. The AI attracting the most capital right now is animated by different values and different incentives entirely. 

The allocation of resources toward general-purpose AI rather than targeted, public-interest systems is not simply a market outcome. It is a function of the same legal architecture this article examines: the R&D tax offset subsidises AI development without distinguishing between AGI optimised for commercial surplus and narrowly scoped AI that serves public health; the patent monopoly rewards whoever reaches the market first regardless of what problem they were solving; and geopolitical competition between US and Chinese AI developers creates maximum incentive to build the most powerful general system as quickly as possible, regardless of what it is actually for.⁷ 

This is what it means to be an apocaloptimist. Not to believe AI will inevitably produce a better world—it will not, absent deliberate choices to make it so. Not to believe the dystopian outcome is inevitable either—it is not. The difference between the world AI makes possible and the world AI is likely to produce, will largely be determined by human institutional choices, which means they are amenable to reform. The task this article sets itself is to identify the specific mechanisms through which that architecture could be redesigned, so that 1) the race dynamics may be eliminated, allowing in turn for targeted, public-interest AI systems to be better prioritised in development, and 2) the productive surplus generated by AI and other technologies of abundance translates into broadly shared economic welfare rather than further concentrating wealth.

III. Potential Limitations to Government Regulation

The standard institutional response to concerns about AI's distributional consequences is regulatory in character. Proposals to tax automation, mandate disclosure, legislate reduced patent terms, or expand compulsory licensing are regularly advanced in policy and academic literature. These proposals have theoretical merit; in practice, however, three structural features of the current environment substantially undermine their viability. 

A. Regulatory capture 

Regulatory capture occurs when the industries that regulation is designed to constrain instead come to shape the regulatory frameworks governing them, redirecting those frameworks toward the protection of incumbent interests rather than the public purposes that originally justified intervention. Stigler's foundational 1971 analysis established this as a structural feature of the relationship between technically complex industries and the regulatory agencies that depend on those industries for expertise, personnel, and information.⁸ The phenomenon is not a product of bad faith—it emerges from the basic informational asymmetry between regulator and regulated, which is more pronounced in AI than in almost any previous technology. 

In the AI context, regulatory capture operates through several mutually reinforcing channels. First, the revolving door between frontier AI labs and government agencies means that the officials responsible for designing AI regulation frequently come from, and return to, the industry they are regulating.⁹ Second, AI companies have invested heavily in shaping the normative vocabulary of AI governance—framing debates around safety, alignment, and existential risk in ways that position frontier labs as responsible stewards rather than subjects of external constraint, and that direct regulatory attention toward long-run speculative harms rather than the immediate distributional consequences of deployment.¹⁰ Third, and most structurally, the technical complexity of frontier AI systems means that regulators lack the independent capacity to evaluate compliance without relying on self-reporting by the regulated entities themselves—a dependence that systematically advantages those entities in any regulatory negotiation. The consequence is that the question of who captures the productive surplus AI generates rarely enters the frame within which governments receive advice on AI law reform. That absence is not accidental; the entities with the institutional resources to shape that consultative frame possess no structural interest in inserting it. 

B. Geopolitical competition 

Any regulatory framework that meaningfully constrains AI development in democratic jurisdictions while autocratic adversaries operate under no equivalent constraint produces a predictable consequence: it accelerates the relative position of the unconstrained actor. This is the core structural objection to over-regulating so as to undermine innovation. At the international level, unilateral constraint in a geopolitically competitive environment generates strategic disadvantage without producing the coordination benefits that would justify it.¹⁰ The semiconductor export control problem illustrates the deeper difficulty. NVIDIA chips have been reaching private actors from foreign countries through third-party intermediaries, circumventing US export controls.¹¹ The fundamental issue is that compute capacity is being distributed through private commercial channels rather than government-monitored ones, making verification and enforcement structurally intractable. This dynamic compounds at the software and regulatory level: meaningful international coordination requires verification mechanisms that do not currently exist, and unilateral domestic action in their absence arguably relocates rather than resolves the underlying problem. A further point of consideration is that unilateral regulation in one country may lead to corporations in that country fleeing and trying to operate in foreign jurisdictions to escape regulation, producing a ‘race to the bottom’ scenario. 

C. Legislative lag 

The rate of AI development systematically outpaces the institutional timelines of legislative processes. The EU AI Act required three years of development before entering into force in August 2024, by which point the regulatory categories it employed did not map coherently onto large language models and generative AI systems that emerged after the drafting process had commenced.¹² The Australian experience is equally instructive. The Full Court in Commissioner of Patents v Thaler correctly applied s 15(1) of the Patents Act 1990 (Cth) to a set of facts the statute's drafters could not have anticipated, and simultaneously flagged that Parliament should address the resulting policy questions as a matter of urgency.¹³ Three years have elapsed without legislative action. The Law Council's own submission to the government's AI consultation acknowledged that the standard five-year statutory review cycle will be inadequate for AI governance purposes.¹⁴ That acknowledgment from the peak body advising Australian governments on law reform amounts to an institutional admission that the legislative machinery, as currently configured, is perhaps structurally unsuited to the pace of the technology it is being asked to govern.

IV. Private Governance: A More Promising Mechanism, But Currently Incomplete Proposal 

The structural failures of government regulation do not support the conclusion that governance is unnecessary in any way. They support a different conclusion: that effective governance of AI's distributional consequences perhaps requires institutional mechanisms that are faster, more technically adaptive, and less susceptible to the capture and obsolescence that afflict traditional legislative processes. 

The private governance model conceived by Dean Ball offers a more promising institutional starting point.²⁸ Under this framework, a government body licenses private AI standards-setting organisations that possess demonstrated technical and legal competence and verifiable independence from industry. AI developers may opt in to certification from those bodies, submitting to audits verifying compliance with published safety and security standards. In exchange for certification, developers receive safe harbour protection from tort liability. The private body conducts annual audits of each certified developer. The authorising government body periodically reviews and re-licenses each private regulatory body. Safe harbour protection is withdrawn in cases of reckless, deceitful, or grossly negligent conduct. The structural advantages of this model are meaningful: non-profit standards bodies can operate across jurisdictions without requiring treaty-level coordination, generating effective international standards through voluntary adoption rather than intergovernmental negotiation. The safe harbour creates a genuine competitive incentive for compliance rather than relying on the deterrent effect of uncertain regulatory enforcement. 

The critical limitation of the Ball model—and of every analogous proposal currently advanced in the AI governance literature—is that it addresses safety and accountability while leaving the distributional question largely unresolved. Certification standards under existing proposals govern security protocols, transparency obligations, and harm avoidance. They impose no requirements relating to the allocation of productive surplus. An AI system may be fully certified, technically safe, and operationally transparent while simultaneously concentrating the economic returns from its deployment in corporate hands through the mechanisms described in Part III. Private governance as currently conceived addresses one dimension of the AI governance problem while the structural dimension that is at least equally significant remains unaddressed. 

The appropriate response is not to reject the private governance model but to extend its normative scope so that distributional standards operate as conditions of certification alongside safety standards. Three reform mechanisms warrant consideration in this regard. 

A. Open Access Requirements 

Certification eligibility could be conditioned on a requirement that AI systems trained on publicly funded data—including protein structure databases, government-funded clinical trial records, and publicly funded scientific literature—make their outputs available under compulsory licensing arrangements. This reform would have direct institutional precedent: the WIPO Development Agenda, adopted in 2007, already establishes a multilateral framework for incorporating development and access considerations into IP standard-setting processes.²⁹ The normative principle—that publicly funded inputs generate public obligations with respect to outputs—is neither novel nor legally radical. 

B. Revenue Sharing Obligations 

Safe harbour eligibility could be made conditional on contributions to a public access fund, structured as a pooled mechanism for subsidising generic production or financing price controls on AI-generated pharmaceutical and other discoveries. The structural logic of this proposal is analogous to the unemployment insurance experience-rating system that Abbott and Bogenschneider identified as a model for automation taxation: firms that generate identifiable social costs bear a proportionate contribution to the mechanisms designed to manage those costs.³⁰ Bogenschneider's subsequent analysis of AI-specific tax implications documents the consistent political failure of government-imposed automation levies and argues that private governance certification frameworks may offer a more tractable pathway to functionally equivalent redistributive outcomes.³¹ 

C. Mandatory Disclosure of AI Inventive Contribution. 

Certification standards could require developers to disclose, as a condition of safe harbour eligibility, the degree to which AI systems rather than human researchers contributed to any claimed invention. Such disclosure would serve two functions simultaneously: it would generate the evidentiary foundation necessary for a future tiered patent term regime calibrated to actual discovery costs, and it would create public accountability for the gap between formal legal inventorship requirements and the substantive reality of AI-driven discovery. Transparency of this kind is not itself redistributive—but it is the institutional precondition for informed redistributive reform. 

V. The Imperative of Organised Civil Society 

None of the reform mechanisms identified above will be realised without sustained and organised civil society pressure. The historical record on this point is unambiguous. The legal infrastructure of employment protection—minimum wage obligations, occupational health and safety regulation, collective bargaining rights—was not the product of governmental beneficence. It was the product of organised social movements that identified a structural inequality, articulated a principled claim against it, and made the political cost of inaction greater than the political cost of reform. The environmental movement generated environmental law through the same mechanism. This pattern reflects the basic political economy of legal change in democratic systems: diffuse public interests require organised advocacy from the ground up to compete with the concentrated private interests that benefit from the status quo. 

The movement this article calls for must make three specific demands of private governance bodies. First, that distributional standards be incorporated into certification frameworks as auditable conditions of safe harbour eligibility—not as aspirational principles appended to governance documents, but as enforceable requirements subject to the same rigour as safety and security standards. Second, that AI developers be subject to mandatory disclosure obligations regarding AI's contribution to productive output, making legible the currently invisible relationship between publicly funded scientific infrastructure and privately captured commercial value. Third, that the governance structures of private standards bodies include meaningful civil society representation, providing an institutional guarantee of independence from the industry interests those bodies are charged with regulating. 

The empirical stakes of inaction are well documented. Acemoglu and Restrepo's research established that automation consistently suppresses labour's share of national income, and that the dominant pattern of AI deployment to date corresponds to what they term "so-so technology"—substitution of capital for labour that generates productivity gains accruing primarily to capital owners rather than to the broader workforce or to consumers.³² Piketty's analysis of long-run capital dynamics provides the structural backdrop: in conditions where the rate of return on capital persistently exceeds the rate of economic growth, wealth concentration is the expected equilibrium outcome absent deliberate policy intervention.³³ AI, by simultaneously expanding capital returns and contracting labour's bargaining position, accelerates this dynamic rather than disrupting it. These findings do not represent contested claims at the margins of economic debate—they constitute the empirical baseline against which the adequacy of any governance response must be assessed. 

The technology generates the material conditions for abundance. The legal architecture, as currently configured, converts that potential into concentration. The distance between those two outcomes is not fixed by technical necessity. It is determined by the institutional choices societies make about the legal frameworks within which AI operates—and those choices are ours to make. 

VI. References

Primary Legal Instruments 

Patents Act 1990 (Cth) ss 15(1), 67 

Income Tax Assessment Act 1997 (Cth) s 355-100 

Superannuation Guarantee (Administration) Act 1992 (Cth) 

Commissioner of Patents v Thaler [2022] FCAFC 62 

IceTV Pty Ltd v Nine Network Australia Pty Ltd [2009] HCA 14 

Telstra Corporation Ltd v Phone Directories Company Pty Ltd [2010] FCAFC 149 

Agreement on Trade-Related Aspects of Intellectual Property Rights, opened for signature 15 April 1994, 1869 UNTS 299 (entered into force 1 January 1995) 

Australia–United States Free Trade Agreement, signed 18 May 2004, [2005] ATS 1 (entered into force 1 January 2005) 

Declaration on the TRIPS Agreement and Public Health, WTO Doc WT/MIN(01)/DEC/2 (14 November 2001) 

Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence [2024] OJ L 1689 

Books 

Bostrom, Nick, Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014) Hao, Karen, Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI (Penguin Press, 2025) 

Piketty, Thomas, Capital in the Twenty-First Century (Arthur Goldhammer trans,

Harvard University Press, 2014) (first published 2013) 

Journal Articles 

Abbott, Ryan, 'I Think Therefore I Invent: Creative Computers and the Future of Patent Law' (2016) 57 Boston College Law Review 1079 

Abbott, Ryan and Bret Bogenschneider, 'Should Robots Pay Taxes? Tax Policy in the Age of Automation' (2018) 12 Harvard Law and Policy Review 145 

Acemoglu, Daron and Pascual Restrepo, 'Robots and Jobs: Evidence from US Labor Markets' (2020) 128(6) Journal of Political Economy 2188

Acemoglu, Daron and Pascual Restrepo, 'The Wrong Kind of AI? Artificial Intelligence and the Future of Labour Demand' (2020) 28 Cambridge Journal of Regions, Economy and Society 25 

Bogenschneider, Bret N, 'Will Robots Agree to Pay Taxes? Further Tax Implications of Advanced AI' (2021) 22 North Carolina Journal of Law and Technology 1 

Bouchard, Lindsey S et al, 'What is the Impact of Intellectual Property Rules on Access to Medicines? A Systematic Review' (2022) 18 Globalization and Health 40 

Dal Bó, Ernesto, 'Regulatory Capture: A Review' (2006) 22 Oxford Review of Economic Policy 203 

Gebru, Timnit and Émile Torres, 'The TESCREAL Bundle: Eugenics and the Promise of Utopia Through Artificial General Intelligence' (2024) 38 Science, Technology, and Human Values 1 

Good, I J, 'Speculations Concerning the First Ultraintelligent Machine' (1965) 6 Advances in Computers 31 

Jumper, John et al, 'Highly Accurate Protein Structure Prediction with AlphaFold' (2021) 596 Nature 583 

Solovy, Eric M, 'The Doha Declaration at Twenty: Interpretation, Implementation, and Lessons Learned on the Relationship Between the TRIPS Agreement and Global Health' (2022) 42 Northwestern Journal of International Law and Business 253 

Stigler, George J, 'The Theory of Economic Regulation' (1971) 2 Bell Journal of Economics and Management Science 3 

Veale, Michael and Frederik Zuiderveen Borgesius, 'Demystifying the Draft EU Artificial Intelligence Act' (2021) 22 Computer Law Review International 97 

Whittaker, Meredith, 'The Steep Cost of Capture' (2021) Proceedings of the ACM on Human-Computer Interaction 

'AI-Enabled Drug and Molecular Discovery: Computational Methods, Platforms, and Translational Horizons' (2025) Discover Molecules (Springer Nature) 

Government and Institutional Reports 

Australian Taxation Office, Research and Development Tax Incentive Transparency Report 2021–22 (ATO, October 2024) 

Australian Taxation Office, 'How Much Super to Pay' (ato.gov.au) 

Boston University Global Development Policy Center, 'Reigniting the Spirit of the Doha Declaration' (Policy Brief No 027, February 2024) 

Law Council of Australia, Safe and Responsible AI in Australia (Submission to Department of Industry, Science and Resources, 17 August 2023)

National Security Commission on Artificial Intelligence, Final Report (2021) 

The White House, Stargate: A Joint Venture to Build AI Infrastructure in America (Press Release, 21 January 2025) 

US Congressional Research Service, Semiconductors and the CHIPS Act: Technology, Economics, and Policy (CRS Report, 2023) 

WIPO, The 45 Adopted Recommendations under the WIPO Development Agenda (WIPO, 2007) 

WTO, Special Compulsory Licences for Export of Medicines: Key Features of WTO Members' Implementing Legislation (WTO Staff Working Paper, 2011) 

World Health Organization, Global Tuberculosis Report 2024 (WHO, 2024) 

Footnotes 

1. John Jumper et al, 'Highly Accurate Protein Structure Prediction with AlphaFold' (2021) 596 Nature 583. The 2024 Nobel Prize in Chemistry was awarded to Demis Hassabis and John Jumper of Google DeepMind and to David Baker. 

2. 'AI-Enabled Drug and Molecular Discovery: Computational Methods, Platforms, and Translational Horizons' (2025) Discover Molecules (Springer Nature). 

3. World Health Organization, Global Tuberculosis Report 2024 (WHO, 2024). 

4. The White House, Stargate: A Joint Venture to Build AI Infrastructure in America (Press Release, 21 January 2025). The $500 billion figure represents a projected four-year commitment, not a single-year expenditure. 

5. I J Good, 'Speculations Concerning the First Ultraintelligent Machine' (1965) 6 Advances in Computers 31; Nick Bostrom, Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014) ch 2. 

6. Stanford University Human-Centered Artificial Intelligence, AI Index Report 2024 (Stanford HAI, 2024). 

7. US Congressional Research Service, Semiconductors and the CHIPS Act: Technology, Economics, and Policy (CRS Report, 2023); National Security Commission on Artificial Intelligence, Final Report (2021) ch 1. 

8. George J Stigler, 'The Theory of Economic Regulation' (1971) 2 Bell Journal of Economics and Management Science 3.

9. Karen Hao, Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI (Penguin Press, 2025); Ernesto Dal Bó, 'Regulatory Capture: A Review' (2006) 22 Oxford Review of Economic Policy 203. 

10. Timnit Gebru and Émile Torres, 'The TESCREAL Bundle: Eugenics and the Promise of Utopia Through Artificial General Intelligence' (2024) 38 Science, Technology, and Human Values 1; Meredith Whittaker, 'The Steep Cost of Capture' (2021) 5(CSCW1) Proceedings of the ACM on Human-Computer Interaction 1. 

11. Congressional Research Service (n 7). 

12. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence [2024] OJ L 1689 ('EU AI Act'), entered into force 1 August 2024; Michael Veale and Frederik Zuiderveen Borgesius, 'Demystifying the Draft EU Artificial Intelligence Act' (2021) 22 Computer Law Review International 97. 

13. Commissioner of Patents v Thaler [2022] FCAFC 62, [119]. The precise language at [119] should be verified on AustLII before submission. 

14. Law Council of Australia, Safe and Responsible AI in Australia (Submission to Department of Industry, Science and Resources, 17 August 2023) [53]. 

15. IceTV Pty Ltd v Nine Network Australia Pty Ltd [2009] HCA 14; Telstra Corporation Ltd v Phone Directories Company Pty Ltd [2010] FCAFC 149. 

16. Ryan Abbott, 'I Think Therefore I Invent: Creative Computers and the Future of Patent Law' (2016) 57 Boston College Law Review 1079, 1100. 

17. Bouchard, Lindsey S et al, 'What is the Impact of Intellectual Property Rules on Access to Medicines? A Systematic Review' (2022) 18 Globalization and Health 40. 

18. Australian Taxation Office, Research and Development Tax Incentive Transparency Report 2021–22 (ATO, October 2024). 

19. Superannuation Guarantee (Administration) Act 1992 (Cth); Australian Taxation Office, 'How Much Super to Pay' (ato.gov.au). 

20. Ryan Abbott and Bret Bogenschneider, 'Should Robots Pay Taxes? Tax Policy in the Age of Automation' (2018) 12 Harvard Law and Policy Review 145, 150. 

21. Agreement on Trade-Related Aspects of Intellectual Property Rights, opened for signature 15 April 1994, 1869 UNTS 299 (entered into force 1 January 1995) arts 27, 33.

22. Australia–United States Free Trade Agreement, signed 18 May 2004, [2005] ATS 1 (entered into force 1 January 2005) art 17.9.7(b). 

23. Declaration on the TRIPS Agreement and Public Health, WTO Doc WT/MIN(01)/DEC/2 (14 November 2001) [4]–[5]. 

24. WTO, Special Compulsory Licences for Export of Medicines: Key Features of WTO Members' Implementing Legislation (WTO Staff Working Paper, 2011). 

25. Boston University Global Development Policy Center, 'Reigniting the Spirit of the Doha Declaration' (Policy Brief No 027, February 2024) 12. 

26. Eric M Solovy, 'The Doha Declaration at Twenty: Interpretation, Implementation, and Lessons Learned on the Relationship Between the TRIPS Agreement and Global Health' (2022) 42 Northwestern Journal of International Law and Business 253. 

27. Boston University Global Development Policy Center (n 25). 

28. Dean Ball, 'AI Governance and Private Regulation' Hyperdimensional (Substack). Verify the exact post title, date, and URL before submission. 

29. WIPO, The 45 Adopted Recommendations under the WIPO Development Agenda (WIPO, 2007), Recommendation 45. 

30. Abbott and Bogenschneider (n 20) 165–167. 

31. Bret N Bogenschneider, 'Will Robots Agree to Pay Taxes? Further Tax Implications of Advanced AI' (2021) 22 North Carolina Journal of Law and Technology 1. 

32. Daron Acemoglu and Pascual Restrepo, 'The Wrong Kind of AI? Artificial Intelligence and the Future of Labour Demand' (2020) 28 Cambridge Journal of Regions, Economy and Society 25; Daron Acemoglu and Pascual Restrepo, 'Robots and Jobs: Evidence from US Labor Markets' (2020) 128(6) Journal of Political Economy 2188. 

33. Thomas Piketty, Capital in the Twenty-First Century (Arthur Goldhammer trans, Harvard University Press, 2014).