Which AI economics and governance research most needs independent expert evaluation? (help The Unjournal prioritize it)

The Unjournal commissions public expert evaluations and ratings of research (and “pivotal questions”) with potential for global impact. We’re currently making a push to consider AI economics and governance research. We’re seeking input from people who use this: which themes, which papers/projects, which claims, who should evaluate them, what should evaluators focus on, and why?

(If you only have a moment, skip to “”.)

Why do this? (Why us, and why now?)

The usual story, which you’re ~probably on board with: AI is advancing rapidly, and becoming more clearly and deeply important to most major global priorities. It’s having large actual and predicted impact on the economy and the concentration of power. This is itself entangled with AI safety and AI global catastrophic risks. “How fast will capabilities grow and replace human work?” “How will gains flow to labor versus capital?” “How will revenue and compute accumulate in a few firms and countries?” “Will regulation change which models get released where?” Questions like these impact the nature of catastrophic AI-risk (resources, race dynamics, human empowerment, state capacity, etc.), as well as the strategic approaches to handling it. There are also direct implications for global poverty, global health systems, and for non-AI GCR (e.g., bioengineered pandemics). These questions (and ‘cruxes’) seem to directly impact a set of live decisions surrounding AI safety strategy, philanthropic funding, regulation and legal safeguards, public spending, international controls, advocacy, and more.

Economists, social scientists, and policy analysts have been moving into this area. Some of the work crosses disciplinary lines. Economists are now modeling AI scaling and recursive self-improvement. Growth theorists are co-authoring scenario models with funding and data from frontier labs. Forecasting researchers are eliciting and synthesizing beliefs and predictions from economists, AI-company staff, and superforecasters. Policy organizations are producing quantitative estimates (chip smuggling, regulation and model delays) that may directly inform policy arguments, legislation, and implementation.

This work is getting attention, and some of it looks influential. The Yale Budget Lab built fiscal scenarios and a tax microsimulation on one set of AI growth forecasts; two prominent monetary economists used its scenarios in framing commentary on Fed policy. A study of model release delays in the EU and UK was reported as showing that data-protection rules, not the AI Act, slowed releases, as the EU considers amending each of these. Epoch AI’s estimate of chip smuggling to China is cited in CSET’s policy analysis relevant to chip-tracking legislation now moving through the US Congress. A widely covered intelligence-explosion paper leans on an economics model in claiming that AI R&D is approaching the threshold for self-sustaining acceleration.

Little of this (as far as we know) has had careful, public, expert scrutiny. Critiques happen in elite academic seminars, but these are usually ~private and not entirely focused on practical impact. There’s some engagement in forums, podcasts, and social media threads, but in general these are not meaningfully vetted, thorough, nor with clear legible expertise. There are some strong ‘independent blog responses’ (e.g.) to some of this work, but even when these are systematic and seem rigorous, they are self-selected. The author chose to write about it and take a position: this makes it less reliable as a ‘neutral expert assessment’.

On the other extreme, academic journal review in economics arguably has the neutrality and legible expertise, but it can take years, and at the end it reveals only which journal (if any) accepted it. Even after a usual ~‘six month turnaround at a single journal’ much of this work is no longer relevant. A paper on AI workforce capabilities, scaling trends, or regulation may be out-of-date in months. Also, people are entering unfamiliar territory in a hurry, making mistakes more likely. This makes rapid peer review from experts in areas like growth, trade, labor economics, industrial economics, and psychometrics particularly valuable; along with plausibility checks from people on the technical side (computer science, ML researchers, engineers, and industry practitioners). We hope to provide more rigor without accepting the (far longer) timelines for traditional review.

How our process fits the problem

We’re neither an academic journal nor a public forum thread. A few things about our process seem to fit this problem:

1. Speed. We commission public evaluations, yielding a public credibility signal in a single shot (not a long submit-reject-resubmit loop). We have a rich evaluator pool, manage the process, and provide compensation, incentives for promptness and strong work, and evaluator prizes. Evaluators write a referee-style report with quantified ratings (and uncertainty) and claim-level assessment, and we publish it with the authors’ responses and a management synthesis.

• With strong incentives for promptness and quality, we can generate useful public credibility signals within about a month of commissioning. We’re putting in a special push to make this timely work go faster, and enabling careful use of AI tools with full human oversight, ownership and scrutiny to help make this possible. Increased incentives would help speed this up further, both through direct incentives, and via their signaling value.

1. Impact focus. We focus on credibility and decision relevance rather than novelty or intellectual sparks. We don’t ask “is this clever, novel and ‘publishable’?” We ask whether it’s asking the important questions, using valid and appropriate methods, correctly describing the results and implications; providing and sharing useful tools; and which claims hold up, which assumptions drive the results, what a decision-maker should and shouldn’t rely on, and what would change the conclusion. (See our evaluator guidelines.)
2. Neutrality. We have a reasonable claim to being disinterested. We don’t produce AI research, we don’t have a specific policy agenda. We are not funded by labs. However, we do have some links to EA/AI-safety aligned funders. Essentially, we have no stake in which way a research result comes out. Labs, advocacy organizations, and bloggers with a thesis choose what to scrutinize and often have a view going in; we try to match each paper with balanced expertise, check conflicts, and pay for careful critique rather than a verdict.
3. A team and a track record. We believe we have a strong team, a systematic process, and set of tools and platforms. We have a pool of 200+ vetted evaluators, mostly academic economists and quantitative social scientists, and have (as of Oct. 19, 2026) published 114 evaluations (with over 1000 ratings; see recent ratings analysis) across 59 research papers and projects. Some of this work is relevant here: evaluations of Aghion et al. on AI and economic growth (where an evaluator found a flaw in a central proof and worked with the authors to fix it), of the Existential Risk Persuasion Tournament, of “Towards best practices in AGI safety and governance,” and of Matt Clancy’s “The Returns to science in the presence of technological risk”.

Where we are

We generated briefs on 34 candidate papers (overviews here): what each claims, why it matters, how well its claims could be checked by expert evaluators, and how much scrutiny it has already had. (You may also find the AI audio walkthroughs helpful; I did.) These were drafted with AI assistance. We checked these against sources, gave human feedback (still in progress; all listed work has at least one human rating), and allowed the model to iterate and improve. The linked page is also an interface for ‘voting on prioritization’, nominations, and discussion: your suggestions and corrections are very welcome.

Scope: AI governance and regulation, AI’s rate of progress, and AI’s primary effects on the economy, on growth, on labor, on the concentration of economic and political power, and the implications for catastrophic risk.

Our key criteria include:

1. Decision relevance: Would the optimal policy, funding, or organizational decisions meaningfully depend on the truth of the claims or the precise value of the estimates in the paper?
2. Prominence/apparent influence: Is it already being discussed or used in influential circles?
3. Checkable claims with legible methods: Models, estimates, or forecasts that evaluators can actually scrutinize.
4. Our scope: Economics/Social Science/Policy/Forecasting: Generally, quantitative or quant/model-adjacent work not focused on technical computer science/ML issues.
5. Timing: recently released or otherwise still relevant to the key debates and decisions.

Two papers prioritized for evaluation

• Forecasting the Economic Effects of AI (Karger et al., Forecasting Research Institute). Structured forecasts from economists, AI-company staff, policy researchers, superforecasters, and the public; the median expectation is 2.5% annual US growth; under a rapid-AI scenario experts see labor-force participation falling from 62% to 55% by 2050. The evaluation work could focus on the methods and applied theory: e.g., the participant selection, non-response and reweighting; question design (and potential anchoring on the historical series); classification of free responses; forecasts weighting and aggregation, the variance decomposition approaches (which drive their headline results), the translation from participation points into “lost jobs”, and using 2026 short-horizon results as an early calibration check.
• Economic Scenarios for Transformative AI (Korinek et al.; Anthropic Institute). This maps modest, substantial, and extreme AI-capability paths onto GDP, labor share, wages, and unemployment to 2030; in the extreme case labor’s share falls from 60% to 45%. This could become a quantitative anchor for the redistribution and “AI dividend” debate. Evaluation work could consider the model structure, sensitivity to the assumptions (exogenous scenario inputs, two-group cognitive/other labor structure, sticky relative wages, the assumed capital-supply elasticity, task-substitution elasticity of 0.5); the reliance on Claude usage data for exposure and automation anchors, the approach to calibration with real metrics, the task-based automation framework, the labor-market adjustment modeling, how scenarios were selected and interpreted, the reliability and usefulness of their scenario explorer, and their suggested policy implications.

The two loosely complement each other: FRI elicits forecasts conditional on capability scenarios, Anthropic Institute’s work models outcomes given specific paths. Together they address key economic questions: what will happen to labor markets, to the labor share of income, and to where the gains from AI accumulate.

Others in the current top tier

• The Economics of Recursive Self-Improvement (Cunningham et al.). This paper builds economic models of recursive self-improvement, represents AI progress as feedback loops, and argues that net acceleration depends on the product of elasticities around each loop. Its back-of-the-envelope calculation suggests loops are not yet self-sustaining but are strengthening. The GovAI intelligence-explosion paper cites it for the claim that AI R&D is near the threshold for self-sustaining acceleration. An evaluation can offer perspective on whether economists’ models of AI acceleration processes are reliable and useful.
• Epoch AI’s estimate of compute smuggling to China. Estimates median 660,000 H100-equivalents through 2025, roughly a third of China’s AI compute. Seems relevant to export-control and chip-tracking proposals.
• Game-theoretic modeling of AI race dynamics and safety (Bueno de Mesquita et al., “The AGI Race and Existential Risk”; Fudenberg & Koh, “Cooperating against Catastrophe”).
• AI liability and insurance papers (Weil; Trout) seem among the most directly relevant to current action and legislation.

What we’re unsure about

Some on our team are skeptical of parts of this; let me share some of the objections. In fast-moving AI work it’s hard to pick papers that will still matter in a month. Forecast and scenario models have many degrees of freedom. A lot of scrutiny already happens informally, reducing the incremental value of our process and feedback. And “papers” may not be the unit through which these debates move. Our response involves setting a high bar: favoring papers where an (~impartial) outside expert view is clearly missing, evaluating methods for domains where predictions can’t be checked, looking for work with clear signs of current impact, and organizing around high-value questions (e.g., the labor share and where resources flow; the pace of capability growth) rather than isolated high-status papers. But we may be getting this wrong; let us know your take.

What would help most

1. Which claims and what research do you rely on? Do you or your organization rely on a number or result from any research on (or outside) the list, e.g., a growth forecast, a smuggling estimate, a release-delay rate, a labor-share path? Does it influence your decisionmaking, funding choices, a policy argument or a research agenda? Tell us which one, and what you’d do differently if it were wrong/if the true answer were different.
2. Evaluator suggestions. We have a strong evaluator pool, but we’re looking for particular types of experts willing to do strong and careful work on a short timeline. E.g., for the two selected papers mentioned above we need people with expertise in:
   • expert elicitation and survey methodology (sampling and nonresponse, weighting, aggregating judgments);
   • macroeconomics and growth theory applied to AI;
   • labor economics of AI.

• We currently offer a modest honorarium and will offer further evaluator prizes, depending on funding (we’re also cautiously considering a sponsored evaluation route, if we can resolve bias and conflict of interest issues). Evaluators can choose anonymity or public recognition. Please make suggestions privately to contact@unjournal.org. (Self-nominations are also welcome, and you can sign up to be part of our evaluator pool or broader team here; again ping contact@unjournal.org if you think you’re particularly relevant for the above.)

1. Your help prioritizing further research for evaluation. You can rate and discuss papers on the page (try to form a view before looking at the existing ratings). You can also nominate further research that’s not in our database, or submit your own research.

How we use prioritization ratings and discussion

These rankings are preliminary. We’ll weigh public input among several considerations. Reasoning matters a lot, not just the vote total, and the decision rests with our management team, advisory board and field specialists.

Remember, the prioritization for evaluation isn’t a judgment of research quality. Some excellent papers may be poor candidates (not relevant to this agenda, hard to evaluate, etc.), and weaker papers may be good candidates for evaluation if they seem to be prominent and impactful.

Co-written by David Reinstein and David Manheim. First-person passages (“I”) tend to refer to David Reinstein’s own views. The co-authors do not necessarily agree with every statement made in this post. Bob Kubinec (Unjournal Co-director) has also approved this post.

1. This area is not precisely defined. We aim to include papers/research focused on the types of actionable and current questions discussed below. (Research doesn’t need to be presented in ‘paper’ format; I actually prefer more dynamic and interactive formats. But below I’ll just say ‘papers’.) These involve the pace and nature of AI progress, its impact on the economy and power concentration, direct and indirect catastrophic and existential risks, and the potential for regulation and legal solutions. We’re considering work using methods from economics, decision theory, social science, applied statistics, and forecasting. We’ll avoid focusing on short-term or backwards-looking policy measures or economic analyses; we’re aiming to provide value of information for decisionmakers and applied researchers.
2. See, e.g., Coefficient Giving’s request for proposals on global health and wellbeing in an era of transformative AI.
3. And it often focuses on issues of academic field and intellectual interest, rather than just credibility and impact.
4. We often reach outside this pool as well, using standard approaches to finding the most relevant expertise, and leveraging our team’s networks.
5. On PubPub, also mirrored elsewhere, with a DOI for each object, and robust integration into traditional information standards, aiming for visibility. We’re also building more interactive/integrated presentations of the research evaluation conversation.
6. Not all authors respond. When they do, we ask evaluators to consider if the authors found errors or omissions in their evaluations, and adjust accordingly.
7. We often take longer, but still far quicker and more informative than the typical journey across economics journals.
8. We haven’t run a controlled experiment on the impact of compensation on effort and promptness, but our anecdotal evidence generally supports this, as ~does some previous formal academic work. Here is an AI recap of this evidence; I’m familiar with the Chetty et al. paper. The evidence is limited, but generally points in this direction.
9. Whatever the latter means; that’s an annoyingly circular term, IMO.
10. We have sought and obtained funding from organizations that seem aligned with AI safety, such as the Survival and Flourishing Fund, and others that endorse ideas like “making AGI go well.” Still, we believe we are clear in all our applications and materials that our processes aim for neutrality and engaging with a range of viewpoints, and our work is about truth-seeking and not advocacy.
11. Some of our team and advisors have ties to organizations whose work is listed; we work to avoid COIs, and acknowledge these where they are hard to avoid. E.g., if someone on a team has a tie to the research or a possible interest-based COI, we work to ensure their vote is not decisive in prioritizing a paper, and to keep them at arm’s length from the evaluation process. Otherwise, the possible COI connection is explained in the public materials.
12. Papers involving the ~beneficial use of AI in a particular market or setting, e.g., automated health care advisors in LMICs, may be relevant to our broader agenda, but they’re outside this stream. And once again, we’re not aiming to evaluate technical AI safety or AI interpretability work.
13. Our approach tends to work better with quantitative work and ~original research making specific claims, rather than synthesis and broad think-pieces, but we’re somewhat flexible.
14. E.g., (A) a paper on two AI labs competing on capabilities and safety, presenting clear arguments backed by formal modeling, may be relevant for a longer period than, e.g., (B) a paper focused on measuring the current labor-market capabilities of frontier models.
15. As mentioned above, this is already being used in some fiscal scenarios and in commentary on Fed policy. Also note a Brookings Hutchins Roundup summary and a favorable Marginal Revolution note.
16. This work has had an independent computational reproduction; that checked the ~arithmetic, not the assumptions nor the interpretation.
17. David Reinstein: my personal take is that these microeconomic decision-theoretic models are likely to generate insights and organize discussion, but will not yield results or predictions that we should rely on in a literal sense.
18. However, these are not standard quantitative/economics papers, so this would extend our field and methods scope.
19. We’re also open to suggestions for “pivotal questions” and crucial parameters/cruxes (see our crux explorer) that inform your work where you are looking for relevant research.