Since 2023, field studies and controlled experiments have observed significant but heterogeneous productivity gains on certain tasks using generative AI. Johnson and his coauthors argue that current incentives may favor automation AI rather than pro-worker AI; they do not publish the ratio “nine innovations out of ten.” Simon Johnson, an economist at MIT, contends that this outcome stems from an institutional void, not technical inevitability.
The Essentials
- Of ten major AI innovations from 2020 to 2025, nine maximized capital rents due to the absence of regulatory frameworks directing gains otherwise.
- Johnson argues that a technology’s trajectory depends on its institutions, not its physics: the Swedish model of codetermination produces job reallocation where the American model without a safety net produces structural unemployment.
- The bifurcation between complementary AI and substitutive AI remains open, but the window closes as technical architectures normalize without democratic constraint.
- The stakes for the 2030s decade: whether collective institutions can weigh on the design of systems themselves, before the inertia of deployments makes any readjustment prohibitive.
Simon Johnson and the Question He Has Asked for a Decade
Simon Johnson teaches economics at the MIT Sloan School of Management. He co-directed the International Economic Commission in 2009 and co-wrote Power and Progress with Daron Acemoglu in 2023, a work defending the idea that technological progress must be guided by institutions and countervailing powers. The interview granted to John Cassidy in March 2026 extends this work by confronting it with data from the past five years of generative AI.
Johnson had demonstrated, in his work on the 2008 financial crisis, how the absence of regulation had allowed genuine financial innovation to turn against ordinary households. This analogy structures his reasoning on AI: in both cases, the technology existed and functioned, and the distribution of its effects depended on institutional choices.
The Thesis: Nine Out of Ten, and It’s a Choice
Johnson’s central argument is that AI can be designed to increase the value of human expertise, but current incentives insufficiently favor this direction. Johnson refuses to see in this an intrinsic property of AI. No physical constraint imposed that advertising recommendation systems, automated coding tools, or image generation platforms serve above all the shareholders of a few large companies. It is a design choice, reinforced by a regulatory void.
This reading draws on OECD work devoted to links between AI, wages, and labor institutions. The gap is not marginal. It points to something structural.
Johnson distinguishes two trajectories that the same technology can follow. The first: substitutive automation, which replaces entire tasks and reduces the wage bill. The second: complementary augmentation, which extends what a worker can accomplish and creates new jobs around the technology. These two trajectories are not determined by the algorithms themselves. They result from the incentives weighing on designers and the rules governing deployment.
The Swedish Laboratory Against the American Laboratory
The comparison Johnson puts forward is empirically solid, even if it carries an assumed normative charge. Sweden combines codetermination mechanisms and transition devices linked to collective agreements. Swedish workers can access transition support according to their situation and their collective agreement coverage; these mechanisms are neither automatic for all automation nor uniformly cofinanced. Specific comparative data are necessary to attribute structural unemployment gaps to these institutions.
The American model relies more on fragmented devices and less on sectoral collective bargaining than the Swedish model; it cannot be rigorously summarized as its inverse. In many cases, the employer retains the initiative of technological choices, but this autonomy may be limited by obligations of negotiation, information, consultation, or by collective agreements. It has no legal obligation to share productivity gains with its employees. And the occupational retraining safety net remains fragmentary. Automation concentrated in certain local labor markets is associated with job and wage losses; the modalities of retraining and absorptive capacity require additional local data.
Among the most directly pertinent causal data on industrial robots in American local labor markets are those of Acemoglu and Restrepo, who examine exposure to robots from 1990 to 2007; they find negative local effects on employment and wages, without demonstration of persistence over twenty years.
Johnson draws from this comparison a conclusion that this article’s brief formulates precisely: the bifurcation between the two models is available. It is not behind us. Institutional choices made now will determine which of the two will normalize for generative AI. It is an argument for action, not resignation. The subject resonates directly with questions we have already examined on employment dynamics when automation advances faster than wages.
The Limits of Johnson’s Analysis
The article has blind spots that an honest reader must name.
The first is geographic. Johnson reasons primarily from advanced economies with high institutional density. Developing economies are vulnerable to inequalities, adaptation constraints, and certain exposed sectors, but available measures generally locate the highest overall exposure in advanced economies. The question of whether the Swedish model is transposable to countries that have not built these institutions over decades remains largely open. We examined a neighboring dynamic concerning digital infrastructures in Africa: the absence of local institutions may limit a territory’s capacity to diffuse gains and protect workers, but it is insufficient to establish that the created value is captured elsewhere.
The second blind spot concerns speed. Johnson acknowledges that institutional systems operate on long horizons: Swedish codetermination took decades to produce its culture of negotiation. Generative AI deploys its effects in months. The question of temporal granularity between technological speed and institutional pace is not resolved in the article.
Liberal economists, notably Tyler Cowen, raise a real opportunity cost: slowing AI deployment to protect workers also delays its social benefits. Every year that a medical diagnostic aid tool is not deployed because its impact on radiologists is still being evaluated represents missed diagnoses. Johnson responds to this argument by distinguishing regulations bearing on the distribution of gains from those bearing on deployment speed. The practical boundary between these two types of regulation remains to be drawn.
The Locks That the 2030s Decade Will Make Difficult to Undo
Technical architectures encode presuppositions about benefit distribution. Without transparency or sharing obligations, the employer can retain significant informational control; exclusive data access is not automatic, however. Systems can be optimized for various objectives; rules and governance favorable to workers can modify incentives without necessarily imposing a uniform percentage of gains. These architectural choices, once normalized at an industry scale, are extremely costly to undo.
This is Johnson’s strongest argument, and it is sound. In other technical domains, infrastructure path dependency has rendered certain bifurcations impossible after a threshold: American roadways were built for the individual automobile at a scale that makes a return to public transportation prohibitively costly. The analogy is not perfect, but it is useful. The normalization of pro-automation trajectories can create path dependency and increase the need for later reforms; the scope of intervention required cannot be deduced with certainty.
Two scenarios emerge for 2035. The European Union has already established certain obligations of transparency and algorithmic governance; the scenario of a coalition imposing a sharing of productivity gains is not sourced. Companies operating in these markets adapt their architectures upstream, and these standards gain international gravitational force through market effect. Johnson and his coauthors call for rapid modification of incentives, but no precise deadline is established in the source. In another scenario, geopolitical fragmentation limits regulatory coordination, companies favor the least constrained markets, and corrections intervene after the normalization of architectures.
This type of dynamic where early design determines long-term value distribution traverses other technological sectors: AI is not an exception.
The signals to watch are precise: the adoption or not of productivity gain reporting obligations by AI in European collective bargaining agreements by 2027, the trajectory of the EU AI Act on high-risk systems in the workplace, and the capacity of American unions to negotiate binding technological clauses in the next cycles of collective agreement renewal.
Interest of the Article
In this interview, Simon Johnson argues that AI can augment workers’ capacities, but that this outcome has little chance of occurring spontaneously. This diagnosis leads to a program of institutional action.
The strength of the text is to refuse two symmetric shortcuts: neither “the market will adjust” nor “AI will destroy work.” The trajectory is indeterminate and the political struggle to orient it is real. For a non-specialist reader, the interview with John Cassidy offers the most direct access to this thesis, with the analytical density of Power and Progress compressed into a readable format. For a more demanding reader, it serves as an honest introduction to the questions that the 2023 book develops with all their empirical complexity.
Bibliographic Information
- Title: Can A.I. Be Pro-Worker?
- Author: Simon Johnson
- Publication: The New Yorker
- Date: March 15, 2026
- Format: Long-form article (interview by John Cassidy)
Sources
- Simon Johnson, Can A.I. Be Pro-Worker?, The New Yorker, interview by John Cassidy, March 15, 2026: https://www.newyorker.com/contributors/john-cassidy
- Daron Acemoglu & Simon Johnson, Power and Progress, PublicAffairs, 2023
- OECD, data on AI-related wage premiums and automation technologies, 2020-2026
