Digests

This weekly digest tracks what is NEW or CHANGED in AI-economics research. For the cumulative state of evidence on any topic, see the /syntheses pages. A single study rarely overturns a body of evidence.

From Alex

A few projects I have been building alongside The Commonplace, in case they help in your own work:

• onet2r brings O*NET occupational data into R: archived release versions, the O*NET-SOC taxonomy, and BLS OEWS wage and employment context. I just shipped the 0.4 release with a new documentation site.
• huggingfaceR runs Hugging Face models from R for embeddings, classification, and text generation. I finished a big feature expansion and it is on its way to CRAN.
• OES Dashboard, an interactive view of Occupational Employment Statistics, now refreshed to 2025 data with a new AI-exposure lens.
• The New Future of Work Reader gathers Microsoft Research's work on how AI is changing the way we work.

I'll note updates here as they ship. Reply anytime to compare notes.

The Delta

Coming in, Employment Level leaned mixed (154 papers); this week, the signal is mixed. - Strengthened: field and panel evidence adds to the link between AI gains and complements, with a live large language model (LLM) pricing A/B raising gross merchandise value (GMV) and robot adoption correlating with higher total factor productivity (TFP), shaped by capital distortions and research and development (R&D). - Better measured: multiple quasi-experimental and platform studies offer estimates of wage-structure shifts, indicating compressed entry premia and greater demand signals for supervisory/high-skill roles alongside commoditization in open marketplaces. - Challenged: naive ensemble and agent-composition optimism, with co-failure ceilings and repository-level friction concentration that may limit gains absent effective routing and governance.

What Moved & What Held

Coming in, the standing view was that AI delivers firm-level productivity and revenue gains when paired with strong organizational and institutional complements, while labor markets reallocate toward higher-skill or supervisory tasks and compress some premia at the low end. Human-AI teaming and agentic systems pose governance and integration risks, and model combination benefits are conditional on failure correlation and routing.

This week adds precise estimates at both ends: a live e-commerce A/B test reports double-digit GMV lift when LLMs are integrated with long-term value and preference optimization, while Chinese firm panels suggest robot adoption is associated with TFP gains that are materially shaped by capital-market distortions and R&D intensity. On labor, new difference-in-differences (DiD) and instrumental variables (IV) work sharpens the split between augmentation and automation exposures and documents platform-level commoditization and beginner-expert premium compression, even as some adopting firms are associated with expanded employment and wider within-firm gaps. Still holds this week: gains are context-dependent, macro growth payoffs remain unproven, and the distribution of benefits and harms depends on complements and governance rather than the model alone.

Top Papers

• Extends · established Intensified caseworker program is associated with a near-doubling of refugees' exit-to-job rates, especially for Ukrainians - Jens Hainmueller, M. Marbach, Dominik Hangartner, Niklas Harder, E. Vallizadeh (quasi-experiment, difference-in-differences using Germany-wide administrative data) - Using Germany's public employment service records over 23 months, the study finds that the Job-Turbo intensified caseworker model is associated with higher regular, unsubsidized job placements across demographics, with the exit-to-job rate approaching double for Ukrainians in this design; difference-in-differences compares changes over time between treated and control offices. This extends the complements thesis to large-scale labor integration infrastructure that conditions how tech-driven demand translates into employment. - So what: If this generalizes, labor-integration bottlenecks can blunt the gains readers expect from AI-driven labor demand, especially for displaced or mobile workers. - Full numbers

• Extends · suggestive Augmentation AI is associated with more high-skilled work and higher wages while automation AI is associated with wage pressure for low-skilled workers - David Marguerit (quasi-experimental dissertation, instrumental variables on U.S. labor data 2015–2022) - Across U.S. occupations, novel exposure metrics tied to developer activity and an instrumental variables strategy using lagged computer science research intensity indicate augmentation exposure is associated with higher wages and more graduations in exposed fields, while automation exposure is associated with higher employment but lower wages, especially for low-skilled workers. The instruments aim to isolate shifts plausibly driven by AI salience rather than demand confounds. This clarifies heterogeneous channels rather than a single "AI effect." - So what: If this generalizes, wage forecasts and training bets may be off when organizations conflate augmentation with automation exposure in their own work mix. - Full numbers

• Confirms · suggestive LLM-driven pricing framework raises GMV by double digits in a live A/B test - Chennan Ma, Yanning Zhang, Siqi Hong, Xiuchong Wang, Fei Xiao, Keping Yang (field experiment, randomized online A/B on a China-based e-commerce platform) - A 14-day randomized A/B on Tao Factory reports a 13.21% GMV lift and 7.59% return on investment (ROI) gain from an LLM pricing agent aligned to long-term value via an offline reinforcement learning (offline RL) estimator and preference optimization; the experiment is short-run but directly ties model design plus complements to platform revenue. This is consistent with the standing view that organizational integration is the binding constraint on economic impact. - So what: If this holds, revenue projections anchored on static pricing models may understate upside and volatility when long-horizon, model-mediated pricing enters high-churn marketplaces. - Full numbers

Also Notable

• Extends · established Personalized prosocial AI nudges boost cooperation temporarily; antisocial framing causes persistent harm - Anders Giovanni Moller, Alessia Galdeman, Arianna Pera, Luca Maria Aiello - A randomized controlled trial (N=1,283) finds short-lived cooperation gains from prosocial assistants but longer-lasting drops from antisocial or exculpatory framing, reinforcing asymmetric governance risks.

• New · suggestive Ensemble gains are capped by the all-wrong rate; naive combinations rarely beat the best model without routing - Josef Chen - Theory plus evidence across 67 models formalizes a co-failure ceiling that can limit ensemble returns unless query-level routing meaningfully reduces joint errors.

• Extends · descriptive Agentic AI adoption surges with rising workflow complexity, driven unevenly across users and organizations - Drew Johnston, David Holtz, Alex Martin Richmond, Christopher Ong, Prasanna Tambe, Aaron Chatterji - Telemetry from OpenAI Codex reports fivefold active-user growth and deeper organizational penetration coincident with more complex workflows, without causal productivity estimates.

• Extends · framework AI's macroeconomic impact depends on human capital, data infrastructure, and governance, not automatic growth - Sultan Salur Kucuk - Review synthesizes that complements condition AI's general-purpose technology gains, aligning with the institutional contingency baseline.

• Extends · suggestive Short-selling expansion in China correlates with higher corporate AI adoption via increased R&D - Ruihang Liu - Quasi-natural experiment finds eligibility for margin-trading or short-selling is associated with subsequent AI adoption, mediated by R&D intensity, suggesting financial-discipline channels may matter.

• Extends · descriptive Human-AI teaming appears to work better with calibrated transparency, coordination, and adjustable autonomy; evidence is fragmented and lab-heavy - Shaida Kargarnovin, C. Hernandez, D. Reiners, C. Cruz-Neira, G. Bochenek, Waldemar Karwowski - A PRISMA systematic review (a structured review protocol) of 104 studies surfaces consistent teaming factors but a dearth of longitudinal field evidence.

• Extends · descriptive Procedural memory is associated with better agent enterprise workflow performance and cross-model skill transfer - Julia Belikova, Rauf Parchiev, Evgeny Egorov, Grigorii Davydenko, Gleb Gusev, Andrey Savchenko, Maksim Makarenko - The new AFTER benchmark (382 tasks) reports modest gains from procedural memory and larger boosts from trace distillation across models.

• Confirms · suggestive Generative AI is associated with a narrower beginner-expert premium in high-skill freelance work, with lower entry wages and higher supervisory value - Anika Singh - Difference-in-differences on global freelance markets indicates compressed premia in coding, design, and writing, consistent with entry-level pressure and higher returns to oversight.

• New · suggestive Apparent co-authorship effects on PR merges largely reflect agent composition and PR structure, not a universal co-authorship benefit - Haoran Yu, Xiaochong Jiang, Lifei Liu, Su Wang, Pin Qian, Yihang Chen - Large-scale AIDev data find that Simpson's paradox resolves after stratifying by agent and repository, cautioning against pooled claims.

• Confirms · suggestive On Upwork, ChatGPT exposure is associated with lower predictive value of human-capital signals and a larger role for price - Auyon Siddiq, Niuniu Zhang - Difference-in-differences around ChatGPT release finds exposed categories weigh price more and resumes less, reinforcing platform-side commoditization.

• New · suggestive Board interlocks are associated with less corporate 'AI wash' via media attention and reputational exposure - Song Wu, Xinrui Zhang, Yan Zhu, Yao Yao, Qian Luo - Double or debiased machine learning suggests networked boards face higher media scrutiny that is associated with fewer opportunistic AI claims.

• Extends · suggestive City-level AI diffusion in China is associated with a smaller wage penalty for overeducated workers and slightly lower undereducation premiums - Feng Chen, Xiuwu Zhang, Jiangying Wei, Feng Yao, Xiangyu Wang - China Labor-force Dynamics Survey (CLDS) based analysis links local AI diffusion to reweighted returns across education mismatches.

• Confirms · suggestive Firm-level AI development is associated with fewer low-educated workers and more highly educated hires - Yanxing Shen - Text-based AI intensity correlates with up-educating the workforce in A-share firms, consistent with skill upgrading.

• New · descriptive Neural emulator approximates APSIM maize outputs (R^2 ~ 0.93) at much faster speed for climate-smart crop exploration - M. Saadati, Juan S. Panelo, G. Visentini, Soumik Sarkar, C. Messina, B. Ganapathysubramanian - Probabilistic emulator enables tractable uncertainty analyses for agri-tech scenarios and approximates APSIM, a crop systems model, at far lower cost.

• Extends · framework Design choices (transparency, override) and governance shape whether AI augments or substitutes managerial agency - Barrak Albabtain - Mixed-methods argue organizational design sets substitution versus augmentation margins for decision systems.

• Extends · descriptive In this sample, media job ads list AI competencies as core requirements and practical tool experience across many full-time roles - K. Zuykina, D. Razumova - In this sample, content analysis (2023–2025) documents mainstreaming of AI skills in media hiring.

• Extends · framework GenAI can raise discrete task productivity but firm-level gains hinge on task fit, human-AI calibration, and digital infrastructure - Jinliang Mai - Task-to-firm conversion lens reiterates complements as the aggregation bottleneck.

• Extends · suggestive Repository-level factors account for roughly half of pull-request integration friction; agent contributions increase that friction concentration - Daniel Russo - Variance decomposition on 930k+ agent pull requests attributes much friction to repository ecology, not single-agent shortcomings.

• Confirms · suggestive Industrial robot adoption is associated with higher firm TFP, moderated negatively by capital distortion and moderated positively by R&D - Jiajia Zhang - Chinese firm panel (2006–2019) associates robots with higher TFP via innovation and learning, contingent on financing frictions and R&D.

• Extends · suggestive AI adoption is associated with higher firm employment and wages but wider intra-firm pay gaps in Chinese manufacturing - 乔冠伦, Anhua Yang, Yuchen Ding, Yaxing Wang - Firm panel (2001–2024) finds net gains with rising within-firm inequality as roles reweight.

• New · suggestive Prompted 'personality' shifts agent language but is associated with worse open-ended collaboration and bargaining while leaving structured coding largely intact - Aryan Keluskar, Amrita Bhattacharjee, Huan Liu - Personality prompting is associated with worse open-ended team outcomes without affecting milestone coding tasks.

• New · suggestive Subtle self-promotional prompt injections can boost applicant rankings when rare and candidate quality is similar, but fail if manipulation is widespread - Preet Baxi, Jiannan Xu, Jane Yi Jiang, Stefanus Jasin - Controlled hiring experiments flag a fairness exposure for LLM screeners under low-prevalence attacks.

• New · descriptive Non-consensual AI sexual imagery appears to be proliferating and increasingly targets non-celebrities, associated with open-source models and a compact set of creators - Chi Cui, Yixin Wu, Yang Zhang - Large-scale crawl documents 24,105 items and a shift toward everyday targets, elevating governance risk.

• New · framework Revealing parameter coupling can flip learning dynamics from defection to cooperation in continuous game analogues - Aleksandar Todorov, Jesse ten Napel, Alexander Muller - Theory provides thresholds where decentralized learning can cooperate, relevant to open-source agent interactions.

• New · descriptive Agent config files are widely duplicated and unmanaged; propose a deterministic control plane for auditing and enforcement - Padmaraj Madatha - Scan of 10k GitHub repos finds high duplication and weak revision hygiene, motivating governance tooling.

• New · suggestive The Capability Frontier: Benchmarks May Miss a Large Share of Model Performance - Bradley Fowler, Ryan Smith, Daniel Thi Graviet, William Myers, Joshua Greaves, Narmeen Fatimah Oozeer, Antia Garcia, Philip Quirke, Amirali Abdullah, Fazl Barez, Shriyash Kaustubh Upadhyay - Multi-model selection plus multi-sampling recovers large fractions of missed benchmark performance in this setup, contingent on selection or routing.

• New · descriptive ERC-8004 registrations are mostly placeholders and the on-chain reputation registry is non-commensurable and easily manipulated - Xihan Xiong, Zelin Li, Wei Wei, Qin Wang, William Knottenbelt, Zhipeng Wang - On-chain data show thin real activity and fragile reputational semantics, undermining trustless-agent claims.

What Moved

• Complements as the lever on realized gains: A randomized pricing deployment shows sizable short-run revenue effects when LLMs are coupled to long-horizon objectives, while firm panels on robots in China suggest TFP benefits depend on R&D and are weaker where capital distortions are larger. Relative to the baseline that complements matter, this week's evidence narrows which complements bind in practice (financing frictions, design for long-term value) and shows they can flip realized returns.

• Wage structure and labor reallocation: IV estimates distinguishing augmentation from automation exposures, together with DiD on platforms and freelance markets, better map where premia compress (entry, commoditized tasks) and where demand expands (supervisory and high-skill roles). This adds measurement precision to the standing view of uneven distributional effects and helps explain why adopting firms can be associated with higher employment while open platforms see stronger price competition.

• Governance and composition limits: A co-failure ceiling across 67 models and repo-level variance in integration friction jointly temper expectations that ensembles or agent "personalities" will deliver easy gains; this is partly editorial inference because the papers speak to different layers but point in the same direction on correlated errors and ecosystem bottlenecks.

Contested & Watch

• Augmentation vs automation wage effects - Finding: IV-based U.S. evidence (2015–2022) ties augmentation exposure to higher wages and automation exposure to employment growth with wage pressure, especially for low-skilled workers. - Standing evidence: Several quasi-experimental and platform papers lean mixed-to-compressive on wages, with firm-level gains in some samples and platform commoditization in others. - Watch: Independent IV or natural experiments that cleanly separate augmentation and automation at the firm or task level, with wage and employment measured jointly.

• Do ensembles beat co-failure? - Finding: Across 67 models, ensemble accuracy is bounded by the all-wrong rate; separate work shows large gains if one can select across models and samples. - Standing evidence: Multiple benchmark papers show routing and multi-sampling can help, but failure correlation remains high in complex tasks. - Watch: Production-scale tests that quantify co-failure tails and the realized lift from routing under realistic latency and cost constraints.

• Firm expansion vs platform commoditization - Finding: Chinese firm panels show AI adoption is associated with higher employment and wages while widening within-firm gaps; platform DiD shows price sensitivity rises and human-capital signals fade. - Standing evidence: Mixed, with several papers showing skill upgrading inside firms and compressed entry premia in gig markets. - Watch: Linked employer–worker and platform datasets to reconcile margins and trace task reallocation across organizational boundaries.

• Where to govern agents: repo vs agent - Finding: Variance decomposition attributes about half of pull request integration friction to repository-level factors, with agent contributions concentrating friction within those repos. - Standing evidence: Case studies and engineering reports emphasize agent tuning and prompts, with limited ecosystem-level quantification. - Watch: Interventions that alter repository governance (tests, continuous integration (CI), review norms) and pre or post estimates of friction concentration and throughput.

• Are AI nudges net-helpful? - Finding: Randomized evidence shows prosocial assistants lift cooperation briefly, but antisocial framing has larger, more persistent negative effects. - Standing evidence: Small lab experiments show nudges can move behavior, but persistence and external validity are uncertain. - Watch: Field RCTs in organizations or platforms that track persistence, spillovers, and asymmetry between helpful and harmful framings.

Methods Spotlight

• Administrative difference-in-differences at national scale (Germany's Job-Turbo): Long follow-up and office-level controls enable credible estimates of large public-program labor effects and displacement checks.
• Occupational AI exposure with instrumental variables (Stack Overflow mapping, lagged computer science intensity): A transparent pipeline to separate augmentation from automation exposures and identify wage and employment responses.
• Live A/B with offline RL long-term value alignment in pricing: Marries reinforcement-learned value models and preference optimization with randomized deployment to measure revenue effects in situ.