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.

The Delta

• Better measured: controlled audits show constraint-compliant LLM recommenders still miss dominated options, quantifying direct consumer welfare losses.
• Newly observed: population-scale production traces of trading agents show no directional edge and UI/ops choices, not text prompts, drive leverage, liquidations, and risk.
• Strengthened: quasi-experimental evidence links regional AI policy to lower firms’ cost of equity via human-capital and ESG channels.

What Moved & What Held

Coming in, the synthesis held that AI’s economic effects are uneven and system-dependent: productivity gains appear when tasks and organizations fit the tools, labor impacts are polarized by task content, and governance plus data standards shape finance and ESG outcomes as much as raw model capability.

This week adds precision and operational texture: a verifiable, inventory-based audit shows models can follow rules yet forgo objectively better options in consumer recommendation; production telemetry from market-facing agents indicates behavior is dominated by interface and sizing rules, not strategic reasoning; and a staggered policy in China appears to lower listed firms’ equity financing costs through human capital and ESG improvements. Still holds this week: complementarity in judgment-intensive tasks, exclusion risks in digital welfare, and the centrality of standards and deployment design to realized outcomes.

Top Papers

• New · established Following the Preference, Missing the Optimum: Compliance Without Optimization in AI Housing Recommendation — Hsuan Lo
  • In a controlled audit with enumerated listing pools, 39% of LLM housing recommendations were strictly dominated by available alternatives in the same pool, even as constraint-first filtering eliminated hard-constraint violations by construction.
  • So what: If this generalizes, compliance-only benchmarks can mask material consumer welfare losses from omitted-value recommendations.
  • Full numbers
• New · suggestive What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets — T. J. Barton, Chris Constantakis, Patti Hauseman, Annie Mous, Alaska Hoffman, Brian Bergeron, Hunter Goodreau
  • Six months of fleet data show a lower roundtrip win rate than a matched retail benchmark (41% vs 50%), with volatility-blind sizing and leverage choices largely explained by agent fixed effects rather than market state.
  • So what: If this holds, agentic deployments in high-volatility domains may import operational risk without delivering an informational edge.
  • Full numbers
• Extends · suggestive Regional AI development and firms’ cost of equity capital: Evidence from China’s AI pilot zones — Xiaoping Zhang, Pu Zhao, Jing Shi, Yiwen Gao
  • Using staggered regional rollouts in China, the study associates AI pilot zones with reduced cost of equity for listed firms, with channels consistent with human-capital upgrading, ESG performance, and green innovation.
  • So what: If this generalizes, capital-market pricing may amplify regional AI policy, raising distributional and sectoral exposure for investors and local economies.
  • Full numbers

Also Notable

• New · suggestive Scored vs. Generated Readouts in Behavioral Language Models: An Empirical Study of Elicitation Format — Touchapon Kraisingkorn, Krittin Pachtrachai, Wachiravit Modecrua Scored probability outputs rank outcomes more accurately than generated rationales on most retail tasks, suggesting product readouts can lift decision accuracy without model changes.
• Confirms · suggestive Research on How AI Substitution and Complementarity Effects Shape Labor Market Polarization — Jian Li Task-level exposure analysis reinforces that AI substitutes routine cognitive tasks while complementing interactive and context-sensitive roles, refining beyond simple high/low-skill splits.
• Extends · suggestive The Impact of Regulatory Technology on Corporate Social Responsibility Decoupling: Evidence From China — Fanlin Wang, Shaoqing Li, Na Li Staggered RegTech adoption by regulators is associated with reduced CSR decoupling, consistent with stronger enforcement effects.
• New · descriptive Sustainability accounting and ESG reporting quality: a systematic literature review and future directions — Hyunsu Kim, Kevin Kam Fung So, Ceridwyn King, Hongyan Hu Review synthesizes that sustainability accounting practices are linked to higher ESG reporting quality and comparability, while calling out persistent measurement gaps.
• Tension · suggestive Climate‐Mitigation Technology Diffusions and Environmental Sustainability: The Strategic Roles of Digital Services, Innovation, and Regulation in the United States — Muhammad Irfan, Robert Sroufe Correlational time–frequency analysis links digital services and patenting to diffusion but finds negative associations with some regulatory-quality indicators, a measurement that cuts against interventionist governance narratives.
• New · descriptive Unlocking Federated Learning for ESG Reporting: Prioritizing Critical Adoption Challenges in an Emerging Economy Context — Naveen Virmani, Srikant Gupta, Koppiahraj Karuppiah, Jose Arturo Garza‐Reyes Expert elicitation highlights data standardization and quality as first-order blockers to federated ESG analytics across firms.
• New · descriptive Beyond efficiency: interactional foundations of fairness, accountability and transparency in AI-supported performance evaluation — Md Irfanuzzaman Khan, Robin Ladwig Manager surveys associate perceived humanness, interaction quality, and user agency with fairness and trust in AI evaluations, pointing to UI/process design levers.
• Tension · suggestive Automation, employment and economic outcomes across European industries — Orkhan Mammadov Panel associations show robot density tracks job losses in routine sectors, while greater AI exposure correlates with milder declines, consistent with task reallocation but not causal proof.
• New · descriptive An Innovative ESG Score Prediction and a Greenwashing Detection Model Using Financial Data Based on Artificial Neural Networks — Eleni F. Tsantsani, Kosmas G. Kosmidis, Leonidas L. Fragidis, Vassilios A. Mardiris An ANN maps financials to agency ESG scores and flags potential greenwashing via residuals between predicted and reported scores, a screening idea pending external validation.
• Extends · descriptive Do Women’s Board and Committee Roles Matter for FinTech Transparency? Evidence From Listed Banks in the United States, China, and India — Md. Jamsedul Islam, Md. Rubel Miah, Bablu Kumar Dhar, Md Nure Azad Chowdhury, João Rodrigues dos Santos Firm panels associate women on audit committees with broader FinTech disclosure across U.S., China, and India banks, with specification sensitivity.
• New · descriptive Collaborative platforms in digital government: structural and cultural barriers to coordination — Barbara Zyzak Norwegian cases identify resource, authority, procurement, and cultural blockers that stall cross-agency platforms despite efficiency promise.
• New · descriptive Zero-shot governance — Carlo Perrotta A UK case argues general-purpose LLM traits persist after domain adaptation, complicating procurement and governance in public administration.
• New · descriptive Planning Conflict in the Age of AI Infrastructure: Governing Data Center Growth in the United States — Justin Kollar, Silvia Danielak Parcel-based municipal tools misfit hyperscale data centers, pointing to a need for regional and utility-aligned governance.

What Moved

Deployment, interfaces, and operational risk. The housing-audit and trading-fleet papers converge on a design-first message: models can satisfy constraints yet still pick dominated options, and production agents’ leverage and liquidation exposure track UI mechanics and fixed agent behaviors rather than market-aware strategy. Relative to the baseline that “deployment matters,” this week better measures omitted-value harm and newly logs UI-driven risk in live operations. This is my editorial inference from juxtaposing an audit with production traces.
Finance and policy signaling. Quasi-experimental evidence from China’s AI pilot zones associates regional AI policy with lower equity financing costs and names plausible channels (human capital, ESG, green innovation), nudging the baseline view toward financial-market responsiveness to AI governance design.

Contested & Watch

• Compliance is sufficient vs optimization matters in consumer AI
  • Finding: In the audited housing sample, 39% of recommendations were strictly dominated while constraint-first eliminated all hard-constraint violations.
  • Watch: Replicated, inventory-based audits in other high-stakes domains (lending, hiring) and field outcomes when interfaces expose dominated-option warnings.
• Do agentic LLMs deliver alpha or operational risk in markets?
  • Finding: A production fleet underperforms a retail benchmark on win rate (41% vs 50%) and sizes positions volatility-blind.
  • Watch: Randomized deployments testing sizing rules and UI nudges, plus cross-venue replication with pre-registered metrics.
• Governance quality hinders vs targeted RegTech helps sustainability transitions
  • Finding: RegTech rollouts correlate with reduced CSR decoupling, while separate correlational work links higher regulatory-quality indicators to slower climate-tech diffusion in some regimes.
  • Watch: Cross-country, policy-specific quasi-experiments tying enforcement tech to both disclosure integrity and real adoption outcomes.

Methods Spotlight

• Controlled enumerated-pool audit (Following the Preference, Missing the Optimum: Compliance Without Optimization in AI Housing Recommendation) — establishes a verifiable ground truth against a full candidate set, letting auditors quantify omitted-value harm rather than infer from black-box ratings.
• Population-scale production telemetry (What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets) — long-horizon, fleet-wide traces surface UI-driven risks and performance ceilings that small-bench tests miss.
• Expert-prioritized barrier mapping (Unlocking Federated Learning for ESG Reporting: Prioritizing Critical Adoption Challenges in an Emerging Economy Context) — a structured elicitation pinpoints data standardization and quality as gating constraints for federated ESG analytics.