Stronger evidence: ForVerdict
The evidence supports government regulation of AI models, with the important caveat that the FORM of regulation matters enormously. The "for" side is anchored by the 2024 Science consensus paper from Bengio, Hinton, Russell, Kahneman, and over 20 other leading AI researchers, which argues that advanced AI systems pose risks of large-scale social harm, malicious use, and irreversible loss of human control — and that voluntary industry commitments are demonstrably insufficient. Brundage et al. (Cambridge, 2018), a 26-author report from OpenAI, FHI, EFF, and Cambridge CSER, surveyed concrete malicious-use threats across digital, physical, and political domains and explicitly called for regulatory action including export controls and licensing for high-risk capabilities. Tutt (Administrative Law Review, 2017) makes the structural legal argument that opaque dynamic algorithms cannot be effectively governed by tort liability alone — only a specialist FDA-style agency with pre-market review can prevent diffuse algorithmic harms before they scale. Acemoglu and Lensman (AER:I, 2024) provide formal economic backing: their model shows that without regulation, equilibrium AI adoption is inefficient because firms do not internalize social damages, and that faster capability progress paradoxically strengthens rather than weakens the case for regulatory intervention. The "against" side, however, raises legitimate methodological challenges. Marchant, Allenby, and Herkert (Springer, 2011) document the "pacing problem" — formal government regulation is structurally too slow for fast-moving technologies, leading to rules that either lock in obsolete approaches or arrive too late. Chan, Papyshev, and Yarime (Technology in Society, 2024) show in a game-theoretic model that light-touch regulation can be welfare-maximizing under intermediate foreign competition, because heavy-handed regulation may stifle domestic innovation while foreign competitors race ahead. Yeung (Regulation & Governance, 2018) argues that the conceptual instability of "AI" and severe regulator-developer informational asymmetries make ex ante licensing approaches risky, and that hybrid governance combining limited statutes, professional standards, transparency, and judicial review is more effective than prescriptive command-and-control rules. The crucial point is that the "against" papers do not actually argue against ALL government regulation — they argue against the WRONG kind. Marchant favors soft law alongside legal frameworks; Chan finds heavy regulation optimal in some competition regimes; Yeung accepts statutes, transparency rules, and judicial review. None recommends abdication of government oversight. Combined with the strong evidence from Bengio, Brundage, Tutt, and Acemoglu that AI systems impose externalities firms will not internalize voluntarily, the verdict tilts toward "for" — government regulation is necessary, but should be designed to be adaptive (regulatory sandboxes, performance-based standards), competition-aware, and complemented by industry standards and transparency requirements rather than relying on slow-moving prescriptive rules alone.
Ranked by citation count.
The debate
Lisa: The case for government regulation of AI is now overwhelming. Bengio, Hinton, Russell, Kahneman, Yao, Song — basically every senior figure in the field signed a 2024 Science paper saying voluntary commitments by AI labs are not enough. They argue for mandatory pre-deployment risk assessments, technical safety standards, and oversight institutions with the authority to delay or restrict frontier models. When the people who built modern AI are calling for governance, that is a serious signal.
James: A serious signal, yes. But Marchant, Allenby, and Herkert documented in 2011 — and it is even more true today — that government regulation has a pacing problem. By the time legislation is drafted, debated, passed, and enforced, the technology has moved three generations ahead. The EU AI Act took five years from proposal to entry into force; in that time we went from GPT-3 to multimodal frontier models. Prescriptive command-and-control rules lock in obsolete approaches or arrive too late to matter.
Lisa: That is an argument for adaptive regulation, not against regulation. Tutt's Administrative Law Review paper from 2017 specifically proposes an FDA-style agency with expert judgment, pre-market review, and the authority to update standards as the technology evolves. The FDA does not write a new statute every time a drug class changes — it has standing authority to evaluate. Algorithmic governance can work the same way.
James: An FDA for algorithms sounds good in principle, but Yeung's 2018 paper in Regulation & Governance points out that "AI" itself is a moving conceptual target. What counts as a regulated AI system? Recommendation algorithms? Spreadsheets with optimization? Spam filters? The regulatory category is unstable in ways that pharmaceuticals are not. And regulators face severe informational asymmetries with developers, so an FDA for AI is likely to either over-block or rubber-stamp.
Lisa: Categorical fuzziness is a design problem, not a fundamental objection. Risk-based tiering — like the EU AI Act's prohibited, high-risk, limited-risk, minimal-risk categories — handles this. And Acemoglu and Lensman in 2024 in American Economic Review: Insights provide the economic backbone: without regulation, AI adoption is inefficient because firms do not internalize the social damages. They formally show that faster AI growth strengthens rather than weakens the case for regulation. That is striking.
James: It is striking, but Chan, Papyshev, and Yarime in Technology in Society 2024 model the international dimension. In a world of heterogeneous regulatory regimes, light-touch regulation can be welfare-maximizing because heavy domestic rules push frontier development to less-regulated jurisdictions. The EU is on track to become a consumer of AI rather than a producer. That is not an obvious win for European citizens.
Lisa: That is the international coordination problem, which Bengio's Science paper directly addresses. The argument is not "every country should regulate unilaterally"; it is "the major AI-producing jurisdictions should coordinate on minimum standards." The G7 Hiroshima Process and the international AI Safety Institutes are early forms of this. Going first does not have to mean going alone.
James: Brundage and the 26 coauthors in 2018 also recommended export controls and licensing for high-risk capabilities — but a decade later, the US-China dynamic shows how brittle that approach is. Export controls on advanced chips have created retaliation, smuggling networks, and arguably accelerated Chinese domestic capability development. Regulation that ignores geopolitics is regulation that fails.
Lisa: True, and that is why the right answer is not no regulation, but smarter regulation. Combine pre-deployment safety testing, transparency requirements, third-party red-teaming, incident reporting, and liability frameworks — but design them adaptively. Marchant's soft-law approach has a place ALONGSIDE statutory rules, not instead of them.
James: I agree the strict dichotomy of "regulate" vs "don't regulate" is wrong. The real disagreement is between heavy ex ante licensing and light ex post oversight. The literature does not actually support the heavy version. It supports targeted high-risk rules plus transparency and standards.
Lisa: That is still government regulation, just well-designed. Which is exactly what the question is asking — and the answer is yes.
Frequently asked questions
What are the main arguments for regulating AI?
Supporters cite risks such as large-scale social harm, malicious uses like cyberattacks and deepfakes, and possible loss of human control. They also argue that voluntary commitments by developers are not enough and that firms do not absorb the full social costs of AI.
What are the main arguments against government regulation of AI?
Critics point to the pacing problem, where formal regulation moves too slowly for fast-changing technology. They also note that regulators know less than developers, that "AI" is hard to define, and that heavy rules may hurt domestic innovation while foreign competitors move ahead.
Do critics of AI regulation want no government oversight at all?
Not according to the studies reviewed. They argue against prescriptive command-and-control rules and favor alternatives such as soft law, transparency requirements, limited statutes, professional standards and judicial review.
What kind of AI regulation does the evidence support?
The verdict favors adaptive regulation, such as regulatory sandboxes and performance-based standards, that takes international competition into account. It suggests complementing government rules with industry standards and transparency requirements instead of relying only on slow-moving prescriptive rules.