Brief #12: How Trump and Sanders would take AI stocks
AI equity taxes, remote work-driven reductions in junior hiring, and how AI fails to ship code.
Welcome! This bi-weekly newsletter, published by the Windfall Trust, curates the most important developments in AI economics research and policy.
Need to Know
Senator Sanders has proposed an AI equity tax, and President Trump is exploring voluntary government stakes in AI companies. Both approaches would let US citizens share in the windfall from transformative AI scenarios, but they raise questions about international inequality, skewed government incentives, and policy design.
Remote work is a plausible culprit for reduced junior hiring. It’s highly correlated with AI exposure, and when the two are tested jointly, remote work predicts the decline much better than AI in their dataset. Meanwhile, other papers find at most limited attenuation of the AI coefficient when controlling for remote work — and adoption-level AI data, rather than exposure measures, may be needed to adjudicate between the two factors.
While synchronous AI agents increase the number of lines of code by up to 741%, they increase the number of released products by only 20%. This indicates that complementarities in the later stages of the task chain limit the productivity gains in final outputs.
In the news
Overall, the Overton window on both wealth redistribution and labor market policies appears to be shifting rapidly, even though AI’s current measured economic impacts are limited.
Donald Trump is considering receiving shares from leading AI companies. While some are worried about the incentives this would give governments regarding regulation, we discuss the original policy proposal and some design options below.
Meanwhile, Suchet Mittal and Sam Manning (GovAI) analyze potential legislative action regarding AI’s labor impact. Senator Banks’ (R-IN) AI Workforce PREPARE Act seems to offer the most comprehensive proposals for improving measurement and forecasting.
Dean Ball (FAI) and Anton Leicht (Carnegie) suggest more ambitious AI labor policy solutions, such as wage subsidies to help juniors gain work experience.
In detail
What would Sanders’ AI equity tax look like?
Erik Carter illustrated the proposal’s revolutionary potential for the New York Times.
Jeremy Bearer-Friend (George Washington) and Sarah Polcz (UC Davis) suggest a one-time tax on AI labs in the form of equity stakes in a law review article. Senator Sanders cited this article in an op-ed announcing the future introduction of the American A.I. Sovereign Wealth Fund Act.
The authors suggest that reducing AI’s harms, such as copyright appropriation, labor market disruption, algorithmic discrimination, and wealth concentration, is best addressed through in-kind taxation of equity.
The authors suggest that such an equity tax is preferable to higher income or robot taxes, because it reaches firms with novel governance structures, reaches foreign investors, and requires no liquidity from the firms. Additionally, the authors hope that public involvement improves corporate governance.
They discuss a list of policy design options. Regarding the form of equity to be taxed, remitting in proportion to the equity classes already outstanding is simple. Still, it can affect the control rights of some firms more than others, depending on their equity structures. Alternatively, firms could place their AI assets in a holding and grant the government a share. This would isolate the AI part of the holdings, but it could be hard to evaluate the value without external stakeholder pressure on the holding.
Regarding the tax rate, the authors acknowledge trade-offs between minimizing crowding out of private investment and compensating creators. Distributing shares directly to the US population would forfeit the governance influence the authors see as a benefit of public ownership. They therefore prefer a publicly managed sovereign wealth fund governed by a board: it would not trade its shares, but could borrow against them to maintain the stake.
Our analysis: This interesting and rather unprecedented proposal has rapidly entered the Overton Window, with support from Sanders, Bannon, and active interest from, Trump. These policymakers now need to make smart design choices to prevent the downsides.
An AI equity tax helps only in some transformative AI scenarios: What would happen if the rents from AI don’t accrue to frontier labs but to their suppliers, as AI hedge funds seem to bet on, or to adopters of the technology? Diversification to a broader range of stocks would reduce the risk of betting on a specific scenario. But even then, AI’s benefits may flow to consumers just like in the digital economy.
The main issue the problem addresses is wealth redistribution, which may be necessary in the future. A sovereign wealth fund wouldn’t target creators, automated workers, or victims of discrimination directly. And even for wealth redistribution, it’s only helpful for the small share of the world population with a US passport. It may be counterproductive for other regulations if the government has an incentive to protect the profits of AI labs, especially when the government only gets a share in, e.g., OpenAI, rather than in its competitors. The same dynamic may protect incumbent AI labs from upstarts.
The debate illustrates that some policymakers are taking this proposal to extremes by demanding 50% stakes (Sanders, Bannon) or, potentially, direct government shareholding outside a special fund, as happened with Intel. This would drastically affect the labs or the government’s decision-making, respectively. Furthermore, once a government has appropriated shares, it might do so again, prompting investors to preemptively flee.
Does remote work, not AI, reduce junior hiring?
Peter Lambert (LSE) and Yannick Schindler (Oxford) challenge the converging literature that AI caused the reduction in junior hiring. Remote work is strongly correlated with AI exposure, since both are most strongly associated with white-collar, computer-intensive knowledge work.
Junior Hiring started to fall during the pandemic, but fell off a cliff in 2022. When you look at how AI exposure or working from home (WFH) affects the fall in the junior share over time, they can explain similar magnitudes even after controlling for all fixed differences between firms and occupations, as well as for common shocks over time. Entered jointly, the WFH hardly changes, while the coefficient for AI exposure drops to zero or even flips positive. This holds for both WFH exposure and actual WFH adoption from job ads.
The increasing use of remote work during the pandemic might have benefited commuters but had adverse effects on junior employees, according to the authors. WFH increases supervision costs for juniors and reduces their learning when they can’t look over seniors’ shoulders.
They use hundreds of millions of hiring and job-posting data points from the US, the UK, Canada, and Australia, covering 2017-2025. Extensive robustness checks, including measurement-error simulations and various AI exposure measures, confirm this.
The authors suggest that a shared reduction in junior hiring across multiple economies indicates a common shock. If WFH were the culprit, changes in managerial practice, such as a return to hybrid work or more mentoring, could be sufficient to limit the impact of this one-time effect. According to the authors, this should make policymakers cautious regarding taking AI-specific steps early, such as wage subsidies.
Our analysis: This paper challenges the previous literature that converged on AI as the primary driver of reduced junior hiring. Remote work is strongly correlated with AI exposure, since both are associated with white-collar jobs. Whether one mediates or biases the other is the core question.
It depends on the dataset how much WFH attenuates the effect of generative AI. In the ADP payroll data of the Canaries paper, an AI effect survives a WFH control, though a noisy remote-work measure would leave residual bias in the same direction. The sharper test comes from Hosseini and Lichtinger (2026), who respond directly: the junior decline at AI-adopting firms appears equally in positions unsuited to remote work. While studies based on AI exposure might not survive this challenge, adoption-based measures of generative AI largely do.
They deliver robust evidence that classical measurement error in AI exposure is unlikely to drive the results and that remote work is the main driver in their dataset. But if AI exposure scores only differentiate well between carpenters and analysts but not between different kinds of analysts, then a better-measured WFH score might absorb an AI effect.
Overall, this paper should advance understanding of AI’s impact on junior hiring, but AI adoption data is needed to make further progress. Both AI and remote work probably reduce junior hiring, but neglecting remote work leads to an overestimation of AI’s effect.
Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools
While Lambert and Schindler are skeptical about AI’s effect on junior hiring, Mert Demirer (MIT), Leon Musolff (Wharton), and Liyuan Yang (MIT) are skeptical that gains have reached shipped output. AI is much more useful for writing code than for reviewing, integrating, and shipping it, according to the authors. They track over 100,000 GitHub developers across three generations of AI coding tools and find large but sharply bottlenecked productivity gains. In matched event studies, autocomplete raises commits by roughly 40%, interactive agents (“sync”) like Claude Code by 140%, and autonomous (”async”) agents like Codex by 180%. But the gains shrink at every step toward shipped software: sync agents raise lines of code by 741%, pull requests by 65%, but actual releases by only 20%. While AI has increased the number of new apps, total usage of new apps is flat. Even more surprisingly, the share reaching even a modest audience is declining. This is consistent with a decline in the quality of new apps or consumers being unable to discover new products.
The authors formalize software production as a task chain that spans code, files, commits, pull requests, and releases. Each layer combines AI-abundant upstream output with scarce human effort: review, integration, judgment. If these are complements, flooding the bottom of the hierarchy with AI-written code yields bounded gains, the vertical analog of the “weak links“ logic. AI outputs and human judgments are strong complements in their data. At the estimated substitution elasticity of 0.25 (where 0 is perfect complements like shoes and 1 would imply constant expenditure shares for AI and humans), even infinite automation of code-writing would raise final output by at most 60% under their estimated parameters, as long as humans still review and ship.
Treated developers are matched to similarly active controls observed one year earlier, avoiding contamination from silent AI adoption among contemporaneous non-adopters who would serve as controls. While the most motivated developers might adopt new tools even if they have no productivity effect, other tools, such as GitHub Pro, show no such effect, and even long-time Claude Code users speed up after new model releases. Selection can’t explain either pattern.
Task-level productivity estimates overstate near-term output effects wherever production has human-bottlenecked stages. The binding constraint is migrating up the chain: from writing code to reviewing, integrating, and distributing it.
Our analysis: Comparing coders to similar coders from last year is a clever move, and the main vulnerability at once: any platform-wide change between control-year and treatment-year loads onto adoption, so the estimates bundle tool adoption with frontier-model improvement.
The calibrated σ of 0.25 is the most interesting and indicates how complementarities lessen AI’s impact even in a heavily affected workstream like software engineering. But it’s not a universal parameter; it describes today’s organizational technology, in which humans review AI code by hand. AI-assisted review, or firms redesigning workflows around agent output, would raise it, which is why the authors’ own finding that asynchronous agents like Codex already automate the pull-request layer runs counter to the extrapolation of the bound. Some of the attenuation from writing to shipping code might be due to composition, not solely to AI’s limited usefulness in later stages.
Since the paper includes only coding developers who shipped before the testing period used to estimate the uplift on releases, these coefficients focus on more senior developers. These established coders might now work on more mature, longer-term projects, where review and integration are more important than for the general population of coders who see high uplift on coding-specific files. This might inflate the gap in AI’s usefulness in writing and shipping code. If you interpret writing code as junior and review and shipping as senior, this would also be consistent with AI contributing to fewer junior software developers.
The fact that users don’t download new, potentially AI-generated apps is the most policy-relevant fact. Flat downloads confirm that task-level AI productivity gains have not yet translated into more software shipped and used, consistent with the presence of bottlenecks in the task chain. This should make policymakers wary of extrapolating a huge uplift on some tasks to a macroeconomically useful uplift in the same domain.
In Other News
Labor Market & Employment
Joshua Gans (Toronto) demonstrated that rebundling tasks can determine the wage effect of AI exposure.
Seyed Hosseini and Guy Lichtinger (Harvard) also find that AI-exposed tasks vanish from junior task bundles, besides the remote work test discussed above.
The Stanford Digital Economy Lab published dashboards to track youth unemployment, economic takeoff indicators, and AI adoption.
AI Capabilities & Infrastructure
Tom Cunningham (METR) suggests that agents’ marginal productivity diminishes much more quickly than humans’, implying that agents’ expenditure share is low even when they produce a lot of value.
Philip Trammell (Stanford) and Alex Imas (DeepMind) discussed future bottlenecks, the redistribution of AI’s windfall, and the future of developing countries on the Dwarkesh Podcast.
Market Power & Risk
Alicia Garcia Herrero (NATIXIS, Bruegel) argues that Europe’s third way on AI is easier said than done.
Research Opportunities
The OpenAI Foundation commits $250M to “work aimed at building secure and abundant economic futures.”
OpenAI introduces its Economic Research Exchange and welcomes empirical research proposals.
Thanks to Deric Cheng and Joel Christoph for contributing to this week’s edition of the newsletter.