Can a regulator promise both a risk ceiling and a development deadline when safety productivity is unknown? A ceiling below the final model's unprotected hazard requires a minimum stock of safety knowledge, so both promises hold only if weak research can be ruled out. Learning first and then replaying development, with spare compute in safety, comes close to that minimum. In a calibration with only state risk, constant safety yield, and full knowledge transfer, it needs only 3.3 percent more productivity than the necessary bound. Customer services pay for the guarantee. Rules that fix compute allocation fix dates, not risk.
David Almog, Romain Gauriot, Lionel Page, Daniel Martin
We study how algorithmic review changes error composition using Hawk-Eye records from professional tennis. Within 20 mm of the line, mistaken "in" call rates increase by 8.6 percentage points, while mistaken "out" call rates decrease by 4.1 points. Within matches, calls shift toward avoiding previously overturned errors. We find no evidence that initial accuracy deteriorates within 100 mm of the line, despite concerns that workers may rely on correction as a safety net. We derive an investment framework incorporating human responses alongside correction benefits and remedy costs, showing that holding human decisions fixed can misstate the organizational returns to review.
Dull, Dirty, Dangerous? Working Conditions and the Limits of Occupational AI Exposure
Siyi Yu
AI 速览：作者用21个O*NET条目和754个美国职业，考察“枯燥、肮脏、危险”工作条件与职业AI暴露度之间的关系，并指出条目分组和职业特征会影响暴露估计的解读。
Automation is often justified as relief from dull, dirty, and dangerous work. I examine how these conditions relate to occupational AI exposure using 21 O*NET items across 754 US occupations. Higher exposure is associated with lower condition scores under a large language model (LLM) measure and AI Occupational Exposure (AIOE), with equal occupation weights and employment weights. Adjustment for physical demands, required education, training and experience, and wages reduces the composite association. The adjusted confidence intervals include zero for LLM exposure and remain below zero for AIOE. Separating repetition from restricted discretion also changes what the combined Dull score conveys. These comparisons show how item groupings and occupational characteristics shape the interpretation of exposure estimates. Occupational profiles offer a starting point for selecting contrasting research settings. Establishing desired automation or realized relief requires evidence about workers' preferences and specific interventions.
Recommender systems usually learn what people want from what they have done. Platforms could also ask people directly, but asking is costly: 98.8% of the prompts in our data went unanswered. Theories of statistical prediction imply that a person's own forecast adds little to a recommender algorithm built from many people's behavior. Theories of dispersed knowledge imply that people hold information about their own taste that the data lack. We test the two with the MovieLens Beliefs dataset. In 2023-2024, a homepage prompt asked members how they expected to rate movies they said they had not seen. Of 26,020 such expectations, 551 from 277 members were followed by a rating at least a day later. We compare each expectation with four benchmarks: the platform's own predicted rating, two collaborative-filtering models (matrix factorization and an item neighborhood), and a stack trained on other members' outcomes. The trained models use each member's latest ratings. People predict their own ratings more accurately than all four benchmarks (correlation 0.69, against 0.34 for MovieLens and 0.54 to 0.59 for the trained recommenders). Added to all the machine forecasts and features, the expectation lowers out-of-sample mean squared error by 26%. The gain holds when models are trained on earlier expectations and tested on later ones. A member's earlier expectations about other movies do not replace the current one. In ordering a member's own movies, the trained recommenders are statistically indistinguishable from the person. On old, heavily rated movies, the machines do as well or better. The results reconcile the two theories: the crowd's data suffice where they are rich, and the person adds information about the specific item. We derive four design principles for preference elicitation and recommender evaluation.
Defenders and attackers race to find the same software flaws. I use the launch of Project Glasswing, which gave a small group of large US tech firms early access to Claude Mythos Preview, a model for finding software vulnerabilities, to study how access to frontier AI affects defenders, external researchers, and attackers. Using the universe of publicly disclosed vulnerabilities, I find that firms with access more than doubled their disclosures relative to Chinese and Russian tech firms without access, disclosed about 150 percent more high-impact vulnerabilities, and fixed reported flaws as quickly as before. As treated firms found more flaws themselves, external researchers, who sell their findings through bug bounties, reported about half as many vulnerabilities in treated firms' code and more in other firms' code, without leaving the market. There is no evidence that attackers exploited the additional vulnerabilities: the share exploited within a month of disclosure fell by about two thirds, mainly because a smaller share was already exploited when disclosed, while the number exploited did not change. Frontier AI helped defenders find and fix more flaws before attackers exploited them, but its restricted release creates an asymmetry in software security between firms in different countries.
Hui Chen, Antoine Didisheim, Mo Pourmohammadi, Luciano Somoza, Hanqing Tian
We use sparse autoencoders (SAEs) to "brain scan" large language models, mapping internal activations to interpretable concepts and enabling targeted "steering" of specific features. We show that prompting models to adopt specific preferences activates many unintended concepts, complicating causal inference. In contrast, SAEs permit smooth, more isolated steering of target concepts. Furthermore, our scans reveal a stark disconnect between the model's internal feature contributions and its self-generated explanations, showing that these explanations are unreliable guides to its reasoning. Finally, SAEs provide a lightweight intervention for existing models, producing interpretable embeddings and allowing researchers to attenuate specific economic biases.
Maxime Saxena, Fabio Caccioli, R. Maria del Rio-Chanona
Decision Models (DMs), a new class of AI models, give fast responses and return a probability distribution over a fixed set of options. These properties make them candidate human surrogates in economic simulations and surveys, a role increasingly filled by Large Language Models (LLMs), which are costly at scale and under-represent the variation in human behaviour. We present a first evaluation of DMs as human surrogates in economics, comparing 10 DMs with 8 LLMs in three economic settings: a household spending survey on consumption, several classic economic games and a labour-market field experiment. Frontier DMs, those that score highest on general decision-making benchmarks, reproduce the variation in human responses more closely than LLMs in all three settings. They capture the modal human response as well as LLMs in the spending survey, but less well in the economic games. Frontier DMs are also roughly ten times faster than the fastest LLM and can be over a hundred times cheaper. Together, these results suggest that DMs capture the heterogeneity of human behaviour better than LLMs and make large-scale economic simulations more feasible.
Stefano Balietti, Pietro Saggese, Markus Strohmaier
Decentralized Autonomous Organizations (DAOs) replace formal managerial authority with collective decision-making, by handing both proposal rights and token-weighted votes directly to community members. Yet, as in corporate votes, where management sets the agenda and its proposals rarely fail, DAO votes are strikingly one-sided. Using 127 281 proposals and about 51 million votes cast on Snapshot, the leading off-chain DAO voting platform, between 2020 and 2023, we ask which features of the ballot are associated with this support. We relate each choice's voting-power share to three observable features: its position in the list, whether its stance expresses approval, and whether the proposal author selected it. Each identifies one voting bias: position bias, approval bias, and author-selected choice bias. Author-selected choices show the strongest and most robust association, receiving a voting-power share 58.8 percentage points higher than non-author-selected ones, compared with 27.1 percentage points for approving choices and 7.7 percentage points for first-listed ones. About one third of the author bias comes from the author's own voting power; that advantage is smaller in the largest protocols and in proposals with higher participation. The results are robust across alternative specifications and outcomes. Our findings suggest that the agenda power that management holds in corporate voting persists even without a formal hierarchy, and that ballot design and the visibility of proposers' preferences are governance choices rather than neutral implementation details.
Applying Reinforcement Learning (RL) to economic modeling reveals a fundamental tension: standard RL agents do not behave as competitive price-takers. While macroeconomic models typically assume agents take aggregate conditions as given, a naive single-agent RL algorithm learns to internalize its impact on the environment, effectively exerting market power. We illustrate this divergence in a search-and-matching model with concave production, showing that a standard RL agent converges to a monopsonistic allocation rather than the competitive equilibrium. Furthermore, we show that RL's approach to time discounting distorts standard intertemporal economic trade-offs. To resolve these issues, we propose a Calibrated Mean-Field Reinforcement Learning framework that restricts the agent to optimize against an exogenous aggregate state to restore price-taking behavior, and recalibrates the RL reward structure to reflect true economic opportunity costs. By iteratively updating the aggregate state, our algorithm converges to a stationary fixed point that recovers the competitive equilibrium. This approach provides a method for embedding learning algorithms into dynamic general equilibrium models.
Robin Na, Kehang Zhu, Logan Cross, Michiel A. Bakker
Large language models (LLMs) are increasingly used to simulate human behavior, but which descriptions of people improve prediction when experimental designs change remains unclear. We develop a framework that draws candidate behavioral features from the literature and constructs profiles from earlier human experiments. An LLM-based profile generator assigns feature values from observed actions, messages, and surveys; a separate LLM simulates behavior using these profiles. By comparing simulated and observed outcomes in the training data, we select profile contents through forward feature selection, adding features while they improve prediction. We then evaluate the selected contents without modification against outcomes from new participants under new designs. Across 616 public-goods games, 144 multi-party bargaining games, and 1,480 generalized 11--20 money-request designs, leading one-feature profiles predict human outcome distributions better than no profile or general-purpose personas. Profiles with more features provide no clear additional benefit, and features' relative predictive value largely persists across learning and validation. Using profiles from one game family to simulate another produces uneven gains and can make predictions less accurate than simulations without profiles. The framework makes the choice of behavioral information a testable part of simulation design, helping researchers assess which descriptions remain useful as experimental conditions change.
How fast AI inference prices have fallen depends on how an index holds quality constant. We build quality-adjusted price indices for text and code inference sold through APIs, joining 21,024 posted prices to a capability scale fitted to 4,605 benchmark scores. From January 2025 to July 2026, matched-model methods record an annual fall of 0.07 in log price per token. Our pre-specified hedonic index falls at 0.51 a year in logs (95% interval 0.11 to 0.89), entirely through the rising capability of the models on sale. Across three treatments of measurement error and two anchor sets it runs from 0.34 to 0.89. Excluding six anchors flagged for possible contamination leaves model rankings intact (rank correlation 0.998) yet lowers the rate by 0.40 to 0.51 under two treatments. Agreement on rankings therefore does not license claims about levels or increments. Per completed task, the data do not establish a quality-adjusted decline.
Amirreza Zeinali, Jon Saad-Falcon, Avanika Narayan, Mersad Abbasi, Jonathan S. Hartley, John Hennessy, Christopher Re, Azalia Mirhoseini
AI systems are increasingly used to complete economically valuable tasks in various professions, such as software engineering and finance. This motivates the question: how much economic value does an AI system produce per unit of energy? To answer it, we introduce Joules-to-GDP, a framework that measures output dollar value minus AI execution cost per unit of energy, which we call Net Value per Joule (NVJ), and expert time saved per unit of energy, or Time Saving per Joule (TSJ). We then apply the framework to 220 GDPval tasks across 44 occupations. Four findings stand out. First, the most valuable work is not always the most efficient. For example, engineering work such as CAD design yields about 10% more net value per task than software development such as building websites, yet the latter yields about 27% more per joule. Second, defaulting to the most capable model can mean less total value from the same energy budget. Of the five tested models, GLM 5.2 delivers the most net value, $72.27 per task on average, using the energy of about four hours of active laptop use, while Gemma 4 31B achieves 56% of that with 23% of the energy, for 2.5x the NVJ and 2.8x the TSJ. Third, for the same model, NVJ and TSJ vary widely across different accelerators, and adding GPUs can raise energy use for modest gains: eight H100s use 3.2x the energy of one to cut completion time by only 22%, at similar net value. More capable models and more hardware must therefore earn their extra energy with enough gains in output value. Finally, the work of experts who use AI is harder for AI to match than that of experts working alone. Measured against it, GLM 5.2's NVJ falls by 70% and its TSJ by 65%. Taken together, these findings justify economic return per joule as an objective for selecting professional tasks and improving language models and accelerators, and as a basis for examining AI's effect on the economy
Platforms ask customers for online reviews to reduce self-selection, but theory conflicts on whether review solicitation makes online ratings more or less informative. Selection theory implies that asking helps, voice theory that it dilutes the news, and research on rating inflation that compression destroys information. We reconcile these views with signal detection theory, which separates what a rating knows (its sensitivity) from what it shows (its criterion). We test the framework on Steam, a leading PC video game store, whose client began in October 2019 to prompt players for a review after about three hours of play. Using a difference-in-discontinuities design at that inferred threshold in 21.6 million Steam reviews of 315 games, we find that review counts just above the threshold become 4.25 times those just below, and the jump in the positive share grows by 6.6 percentage points. Agreement with critics changes little near launch and erodes later. The rating's movement changes far more: pass-through from reviews written in the first two hours falls from 0.73 to 0.18 just above the threshold, against 0.81 to 0.64 just below, and a fixed rating tier responds to 12% of downturns in those reviews above the threshold, down from 42%, against 41% and 47% below it. What the stream knows changes less. At equal review counts its sensitivity falls more above the threshold, because added reviews respond about half as much and scores near the ceiling are noisier per review, but at natural counts added volume offsets this loss, leaving a change of -0.03 relative to the stream below (95% CI -0.13 to 0.06). Failures that developers acknowledge, which run through complaints, still move the score after launch. Review prompts therefore mainly change how a rating must be read, and platforms that solicit reviews should recalibrate their rating labels.
There is growing experimental evidence that $Q$-learning agents can learn to charge supracompetitive prices. We ask when such behavior persists, whether it emerges, and which strategies support it. In a repeated pricing game in which firms condition on the previous period's price profile, we give conditions on the learned action values under which supracompetitive prices become an absorbing state once experimentation ends. We then show that under softmax exploration these conditions are reached with positive and explicitly bounded probability, both from a random and from an all-zero initial $Q$-table. In the latter case firms lock in supracompetitive prices because they never learn that undercutting pays; the same mechanism can lock in the competitive price, so the long-run outcome is selected by the exploration path. Finally, we give sufficient conditions on the learned table under which the induced strategy is naive collusion, a grim trigger strategy or an increasing-price strategy; grim trigger strategies are subgame perfect under a standard discounting condition, whereas naive collusion is generally not an equilibrium.
Locating political parties on environmental and climate policy is central to the study of party competition, yet the established instruments, expert surveys and the hand coding of manifestos, are costly, periodic, and slow to extend to new documents. We ask whether a panel of large language models can serve as a cheaper and reproducible alternative. Treating each manifesto as a set of sentences, three models from independent providers score every sentence on two dimensions, environmental protection and climate change, separating a party's position on an issue from the salience it gives it, and the judgments are averaged to the manifesto level. Agreement across the models is retained as a built-in reliability signal. We apply the panel to 41 manifestos from the 2024 European Parliament election, covering Austria, France, Germany, Italy, Spain, and the European party families, and validate it against the 2024 Chapel Hill Expert Survey for the 25 parties with a like-for-like expert benchmark, using the other manifestos as diagnostic cases. The panel recovers the expert ordering of parties with strong rank agreement on both dimensions, while reading them as systematically greener than experts do, a gap we interpret as the distance between manifesto promise and governing record. Inter-model disagreement modestly flags where the instrument errs, and position and salience prove related for environmental protection but independent for climate change. The result is a transparent, low-cost, and multilingual instrument that complements rather than replaces expert and hand-coded approaches; the prompt, code, and manifesto-level results are released openly.
Disruptions spread through supply networks when customers cannot replace the products they lose. We ask how portfolio-overlap and product-specific rules for selecting candidate suppliers change cascades when customers search for new suppliers and form new supply edges. From the MarkLines database, we construct a product-annotated network of about 77,000 firms and 197,000 supplier-customer edges in the global automotive industry. We extend an agent-based disruption model with edge-level products, a product-specific substitution rule based on the semantic similarity of product names, and calibrated capacities and inventories. Only 6% of the edges carry observed products. We impute the others and test the imputation against the annotated edges. We run 119,000 simulations. With all other mechanics fixed, the product-specific rule makes the cascades after a hub failure four to five times larger than the portfolio-overlap rule of the original model. Under alternative imputations and similarity thresholds, the factor ranges from 1.2 to 18. Within the simulations, substitutability also shapes which suppliers are critical. Network topology largely determines which firms a shock reaches, and substitutability determines how many of them fail. Of the 50 most damaging suppliers under random disruptions, topology alone identifies 20 and an index that adds product substitutability identifies 32. The index carries over without refitting to a regional scenario modelled on the 2011 Tohoku Earthquake. Supply network risk assessment should account for product-level substitutability and therefore needs product data on supply relationships.
The Costs of Early-career Disciplinary Pivots: Evidence from Ph.D. Admissions
Sidney Xiang, Nicholas David, Dallas Card, Wenhao Sun, Daniel M Romero, Misha Teplitskiy
AI 速览：分析17万份博士申请发现，申请者从先前学位转向语义距离更远的目标项目时，录取概率更低，且入学后毕业更慢，但发表和引用表现无显著差异。
Science often advances when researchers pivot from one discipline to another and import new ideas. However, for established researchers, such pivots are associated with fewer publications and citations. Here, we investigate pivoting at the beginning of a research career--doctoral admissions--when the benefits of importing ideas might outweigh the switching costs. Analyzing 171,347 PhD applications at a large research university from 2013 to 2023, we measured the semantic distance between each applicant's prior degree and target program. Applicants who pivot further had lower undergraduate grades and, controlling for academic and demographic predictors, lower admission chances: the average pivot corresponds to a 5.3% lower admission probability (0.80 percentage points). Among enrollees, the average pivoter takes longer to graduate than non-pivoters, with a 10.7% lower graduation hazard, and similar publication and citation counts on average and at the tails. Our results suggest that pivoting at the outset of research careers, even if it benefits research communities, comes with significant costs to individuals.
Crowds Believe in You: Online Prediction Markets Overestimate How Often People Achieve Their Own Goals
Zihan Chen, Lei Nico Zheng, Di Zhu
AI 速览：基于 Manifold 上 6,150 个自我目标预测市场的订单重建，作者报告自我市场比外部事件市场更常被高估，且纠偏信息主要掌握在当事人手中。
People increasingly share personal goals online and invite others to forecast them. Research on the planning fallacy finds that observers predict a person's behavior less optimistically than the person does, and prediction markets aggregate such judgments, harnessing the wisdom of crowds. We study 6,150 prediction markets on Manifold in which 1,660 users asked about their own futures ("Will I ...?"), reconstruct every order and execution, and compare these self-markets with markets about external events. Self-markets resolved YES 4.4 percentage points less often than their final prices implied, whereas external markets were nearly calibrated (0.9 points); the difference was 2.2 points at the same price level and 1.9 points within creators who made both kinds of market. Absent when markets opened, the overstatement emerged within a day of trading, even where the person never traded, and peaked in the week before close. In a market's first day, observers moved external-market prices down toward the base rate, but moved self-market prices 3.6 points further toward YES, except in markets naming an outcome the person would not want, which were not overpriced. The correction came late and largely from the person, whose orders supplied 38% of the eventual movement toward the outcome in self-markets, against 9% in external markets. Because self-market prices mainly failed to separate successes from failures, a recalibration learned from earlier markets improved later forecasts by only 1.5%. Crowds believe in the futures people pose for themselves, and the information that would correct them lies with the person.
The integration of Artificial Intelligence (AI) into corporate strategy has become critical for organizations seeking to maintain competitive advantage in the digital age. Although organizations increasingly rely on AI as a strategic and organizational resource, existing C-suite roles remain only partially equipped to govern, integrate, and leverage it coherently at the enterprise level. Organizations vary in their responses. Some create a dedicated Chief AI Officer (CAIO), others extend existing mandates into hybrid roles, and still others coordinate AI through federated structures. This paper develops a role-design theory to explain this variation. I identify three properties that distinguish AI systems from earlier cross-cutting enterprise technologies, distributed accountability for judgment, upstream governance, and non-stationarity, and three configurations through which organizations respond: concentrated extension, distributed extension, and role creation. Four propositions specify when a dedicated CAIO emerges, what form an organization's response takes, when the role is effective, and how configurations evolve. Drawing on the first archival census of executive AI leadership, I find that the role has diffused sharply since 2023, that role creation and mandate extension are advancing in near parity, that adoption concentrates where digital leadership structures already exist, and that executive tenures contradict practitioner accounts of a transient role. The paper contributes to research on executive leadership, AI governance, and the design of AI-driven organizations.
Consumers increasingly delegate product search to AI shopping agents powered either by generative large language models (LLMs) that deliberate in natural language or by non-generative decision models that output calibrated choice probabilities in a single forward pass. Because retail platforms monetize through sponsored placements that shoppers typically discount, we examine whether an AI agent's evaluation of paid listings depends on who delegates the agent across both model families. Across three studies with no instructions to favor advertisers, naming a booking platform (vs. a traveler) as the agent's principal significantly attenuates the penalty applied to sponsored listings. Study 1 establishes this sponsorship bias across generative LLMs, prompt wordings, and reasoning efforts, showing that platform delegation weakens pre-choice skepticism and that naming the consumer as the beneficiary eliminates the bias. Study 2 establishes the bias across six commercial and open-weight decision models on calibrated choice probabilities and shows that consumer-aligned fine-tuning of an open-weight decision model (Kev 9B) across non-hotel retail categories eliminates the bias zero-shot on hotel choices. Study 3 decomposes the disclosure label across both model families, showing that attributing the placement to the platform triggers the sponsorship bias, whereas stricter disclosure wording ("Sponsored" vs. "Promoted") lowers paid-listing evaluation without eliminating this bias.
Artificial intelligence (AI) has made it easier than ever to produce research outputs. Adding an agentic tool to a research team's work potentially extends the team's capabilities but calls for an active approach to organizing the research process. The pressing question, therefore, is how research teams should organize their work with this technology. We contend that the question is not new: software engineering and security research have long studied how teams of people working with machines keep control of projects with many contributors. We distill principles from the practices these literatures report and transpose them to AI-assisted research. The principles form TRACE: Track records, Review against sources, Accept independently, Control transitions, and Evaluate corrections. Together, the five principles give researchers a common basis for remaining in control while integrating AI into their workflows. By bounding units of work and attending to how each is shaped, they can help researchers pace AI-assisted work to human attention rather than to the speed at which AI produces content. We also introduce TRACE-OS, an operating system for AI-assisted research projects. It puts the principles into practice through shared instructions, tools, and records that fit researchers' existing workflows. With TRACE, we aim to join a conversation about the governing principles through which the research community can benefit from AI. We contend that a simple enough and useful standard can contribute to harnessing AI capabilities, allowing teams to delegate parts of their work to AI, while also building trust within teams and the broader research community.
When individuals in a social network learn about an unknown state from private signals and neighbors' actions, the network structure often causes information loss. We consider rational agents and Gaussian signals in the canonical sequential social-learning problem and ask how the network changes the efficiency of signal aggregation. Rational actions in our model are log-linear functions of observations and admit a signal-counting interpretation of accuracy. Networks where agents observe multiple neighbors but not their common predecessors confound information, and even a small amount of confounding can lead to much lower accuracy. In a class of networks where agents move in generations and observe the previous generations, we quantify the information loss with an aggregative efficiency index. Aggregative efficiency is a simple function of network parameters: increasing in observations and decreasing in confounding. Later generations contribute little additional information, even when generations are arbitrarily large and agents observe arbitrarily far into the past.
Eduardo Levy Yeyati, Darío Judzik, Juan Ignacio Barbiero
Generative AI changes the scarcity of different forms of human capital. It makes codified knowledge cheaply available at the point of production, while judgment acquired through practice remains tied more closely to the worker. We ask what this does to two familiar prices in the wage structure: the return to schooling and the return to experience. In a short-run framework, machine-supplied codified knowledge substitutes directly for the codified component of worker human capital, lowering the schooling gradient. If codified and practice-based knowledge are complements, the return to experience rises where learning through practice is fastest; technology-specific experience can also become obsolete, rotating the experience profile rather than shifting it uniformly. Using US Current Population Survey earnings data from 2018 through June 2026, we compare occupations in the top and bottom quartiles of the pre-ChatGPT Felten-Raj-Seamans exposure index. After the public release of ChatGPT, the return to a year of schooling falls by 0.55 log points in exposed occupations. The return to a year of experience rises by 0.26 log points at three years of potential experience, is unchanged at fifteen, and falls by 0.12 log points at twenty-five; under the baseline quadratic, the effect crosses zero at 18.1 years. These breaks appear with the release, disappear under a pandemic placebo, survive a common censoring rule across the full sample, and remain when each occupation is allowed its own pre-shock schooling and experience gradients. The pattern is a repricing of human capital inside exposed occupations: codifiable knowledge becomes less scarce, while the value of experience depends increasingly on whether it transfers to the new technology.