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: Large preregistered randomized controlled trials (RCTs) plus a real fundraising deployment suggest frontier conversational AIs out‑persuade trained experts and convert more donations in these settings, estimating a sizable gap in this context under realistic incentives.
• Strengthened: Production A/B (split) tests from DiDi, paired with industrial annotation and exporter panels, add weight to firm‑level key performance indicator (KPI) gains from large language model (LLM) integration while indicating dependence on data, workflow, and financing complements.
• Challenged: A dynamic cross‑country panel reports no independent green‑growth association with national AI ecosystem indices once persistence and governance are accounted for in this sample, pushing back on simple scale-up claims from micro successes.

What Moved & What Held

Coming in, the standing view was: AI delivers task- and firm-level productivity gains in targeted deployments; distributional impacts are uneven with low-skill exposure; and macro/productivity payoffs hinge on complements like skills, governance, and data. Persuasion capabilities were rising but mostly shown in constrained lab settings; recommendation and visibility concerns were plausible but under-measured.

This week, persuasion moves from plausible to more credibly estimated: AIs outperformed incentivized expert humans in preregistered trials and increased donations in one deployment. Micro-to-macro divergence sharpened: more firm-level gains (dispatch gross merchandise value, annotation time cuts, digital export correlations) alongside a cross-country result that green growth does not automatically follow AI ecosystem build-out without institutional capacity in this analysis. Labor effects are better parsed by mechanism, with instrumental variables (IV) evidence separating augmentation (benefits to high-skill) from automation (wage pressure on low-skill), and audits suggest AI-mediated recommendations can entrench brand incumbency or overweight eco labels. Still holds this week: AI’s gains are conditional on complements; distributional asymmetries persist; and macro reweighting awaits wider, governance-aware adoption and longer panels.

Top Papers

• Extends · established Conversational AIs out‑persuade trained human experts and raise far more donations: Kobi Hackenburg, Caroline Wagner, Luke Hewitt, Ben M. Tappin, Ed Saunders, Hannah Rose Kirk, Helen Margetts, Christopher Summerfield (preregistered RCTs + field, high evidence) - Across ≈19k conversations and a live fundraising test, frontier AIs out‑persuade skilled, incentivized humans in these trials and deployment. This extends prior small-lab persuasion results to expert comparators and real money, lending more confidence that AI-driven persuasion could matter in markets.
• Extends · established LLM-derived user profiles raise dispatch prediction AUC and lift GMV in live DiDi tests: Tengfei Lyu, Zirui Yuan, Xu Liu, Kai Wan, Zihao Lu, Li Ma, Hao Liu (production A/B test, high evidence) - Integrating LLM-based behavioral profiles into a ride-hailing dispatcher improved prediction area under the curve (AUC) and nudged gross merchandise value (GMV) and completion rates upward in a 14‑day online test, strengthening the view that targeted LLM integration can move core KPIs in production in this setting.
• Extends · suggestive Augmentation AI links to more high‑skill work and wages, automation exposure to lower wages for low‑skilled workers: David Marguerit (IV-panel dissertation, medium evidence) - Using novel exposure measures and an IV strategy, the work separates augmentation from automation exposures: augmentation is associated with expanded high-skill roles and wages, automation correlates with wage compression and adverse outcomes for low-skill workers in this setting, refining heterogeneity relevant to labor-policy design.

Also Notable

• New · established Whose hotel does the AI recommend? An algorithm audit of reputation signals in LLM-assisted hotel selection: Mirza Samad Ahmed Baig, Syeda Anshrah Gillani, Asher Ali (randomized conjoint audit, high evidence) - Guest ratings and price dominate, while list position and eco labels causally sway recommendations with price-equivalent shifts, quantifying manipulability in assistant-mediated choice in this design.
• New · descriptive AgentFairBench: Do LLM Agents Discriminate When They Act?: Triveni Morla, Rohith Reddy Bellibaltu, Manpreet Singh, Manmeet Singh Kapoor (benchmark study, medium evidence) - Introduces a low-cost benchmark for action-level disparities across hiring, lending, and triage; initial runs find no bias for one model in this setting, and the tool enables broader regulatory testing.
• Confirms · descriptive AI Adoption in Local Government: Productivity, Systemic Risk, and Institutional Resilience: Abayomi Ogunrinde, Carmen De‑Pablos‑Heredero (systematic review, medium evidence) - Synthesizes 68 studies: municipal productivity gains are feasible, contingent on digital maturity, data quality, and governance (consistent with complementarity-heavy baselines).
• Confirms · descriptive From testbeds to high-stakes work: a review of Human-AI teaming domains and teaming factors: Shaida Kargarnovin, C. Hernandez, D. Reiners, C. Cruz-Neira, G. Bochenek, Waldemar Karwowski (systematic review, medium evidence) - Highlights adjustable autonomy, calibrated transparency, and coordination as recurring success factors, while flagging a dearth of long-run field evidence.
• Extends · suggestive Robots, Employment and Wages: Evidence from Turkish Labor Markets: Uğur Aytun, Yılmaz Kılıçaslan, Oytun Meçik, Ü. Yapici (shift-share quasi-experiment, medium evidence) - Robot exposure is associated with higher district employment via firm expansion but coincides with smaller intensive margins for incumbents, aligning with expansion-plus-displacement patterns.
• Extends · suggestive The Impact of Artificial Intelligence Development on Firms’ Educational Composition of Labor: Yanxing Shen (panel correlational, medium evidence) - Chinese firm panels link AI development to a higher share of highly educated workers and a lower share of low-educated workers, with innovation capability as a mediator.
• New · framework Strategic Feature Selection: Jivat Neet Kaur, Pratik Patil, Divya Shanmugam, Emma Pierson, Michael I. Jordan, Nika Haghtalab, Meena Jagadeesan, Ahmed Alaa, Serena Wang (theory, medium evidence) - Jointly choosing features and ridge penalties can outperform blunt exclusion when inputs can be gamed, informing regulator and platform design.
• Extends · descriptive All Smoke, No Alarm: Oracle Signals in Agent-Authored Test Code: Dipayan Banik, Kowshik Chowdhury, Shazibul Islam Shamim (codebase analysis, medium evidence) - 80% of agent-written tests lack strong verification oracles; stronger oracles correlate with higher merge odds, pointing to tooling gaps in agentic coding.
• New · framework A Risk Decomposition Framework for Pre-Hoc Fine-Tuning Prediction: Yuxiang Luo, Chen Wang, Nan Tang (theory, medium evidence) - Derives lower bounds on uncertainty decay and budget-optimal probing for pre-hoc performance prediction, guiding fine-tuning investment.
• New · suggestive Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems: Xi Chu, Yupeng Hou (experiments, medium evidence) - Commercial LLMs favor familiar brands in these tests; tiny rating edges or authority-style cues can flip choices in this setup, implying high manipulability and first-mover advantages.
• New · framework Model Validation of Agentic AI Systems: A POMDP-Based Framework for Belief-State, Forecast, and Policy Validation: Matthew Francis Dixon (theory + case study, medium evidence) - Decomposes agent risk into beliefs, forecasts, and policy for targeted validation, illustrated in portfolio management, using a partially observable Markov decision process (POMDP) framing.
• New · suggestive When Agent Automation Becomes Profitable: Quantifying and Insuring Autonomous AI Risk through Trace-Economic Underwriting: Binyan Xu, Xilin Dai, Fan Yang, Kehuan Zhang (testbed, medium evidence) - Mapping traces to dollar exposure reduced pricing error for agent-insurance in a testbed, sketching a potential market mechanism for automation risk.
• Extends · suggestive How Does Artificial Intelligence Policy Boost Green Innovation in Manufacturing?: Fengyi Li, Tingting Zheng, Hongmei Li (difference-in-differences, medium evidence) - China’s AI Pilot Zones are linked to higher firm-level green patents, with fintech-enabled financing relief as a suggested channel.
• Extends · suggestive Computing power infrastructure and corporate financialization: evidence from China’s supercomputing centers: Jianxiang Zhang, Maoguang Wang, Xinzi Xia (staggered DiD, medium evidence) - Supercomputing rollout is associated with lower financial-asset holdings and more capital expenditures (capex) and R&D, consistent with real-investment reallocation.
• Extends · suggestive Digital pathways to high-quality and sustainable agricultural exports: evidence from the EU market: Xiaofei Chu, Yanlei Hou (staggered DiD, medium evidence) - Digital adoption correlates with 24% higher export value, market entry, and certification uptake, especially for small and medium-sized enterprises (SMEs).
• Extends · established Contaminated Collaboration: Measuring Gender Bias Transfer in LLM-Assisted Student Writing: Ariyan Hossain, Kazi Kamruzzaman Rabbi, Farig Sadeque, S M Taiabul Haque (RCT, high evidence) - Biased AI assistance increased stereotyped language and occupation choices among students in this RCT, confirming bias transfer from tool to human output in this task.
• Extends · suggestive Intelligent Manufacturing Dynamic Capabilities and Corporate Green Innovation: Can Ding, Jianxin Xu, Jing Li (difference-in-differences, medium evidence) - Intelligent-manufacturing pilots coincide with greener innovation via capability-building, with stronger intellectual property protection amplifying effects.
• New · suggestive Exploring the Relationship Between Human-Centric AI and firm Idiosyncratic Risks: Zhen-yuan Ralph Liu, Yu‑Ting Wang, Jia-jia Yan, Shivam Gupta, Mihalis Giannakis (panel correlational, medium evidence) - Human-centric AI strategies associate with lower idiosyncratic volatility, moderated by digitalization and executive ownership.
• New · descriptive Artificial Intelligence as Game Changer in Cybersecurity: What We Learned in 2025-2026, and how this is relevant for Africa: Mikael Alemu Gorsky (policy analysis, high quality) - Argues that frontier LLMs may be increasingly used in cyber operations, while African institutions face restricted access that could widen capability gaps.
• New · descriptive Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines: Pratyush Kumar (measurement study, high quality) - Household brands account for about 73% of mentions in AI search responses while niche brands are around 11% in this sample, indicating concentration in AI-mediated discovery.
• Extends · descriptive Speeding up the annotation process in semantic segmentation industrial applications: Marta Fernandez-Moreno, Margarita Guerrero, Rosalia Rementeria, Pablo Mesejo, Raul Moreno (quasi-experimental, medium evidence) - Unsupervised pre-annotation reduced labeling time by ~78% in this context and releases a useful industrial dataset.
• Confirms · descriptive Labor Market The Impact of Artificial Intelligence on Employment Skills: D. Bano, Saba Shaukat, Maria Kanwal (systematic review, medium evidence) - Reaffirms declining routine content and rising demand for data/machine learning and digital communication skills, with training gaps unfilled.
• New · descriptive FFinRED: An Expert-Guided Benchmark Generation and Evaluation Framework for Financial LLM Red-Teaming: Chaeyun Kim, Daeyoung Park, Junghwan Kim, Jinyoung Jeong, Eunji Song, Yongtaek Lim, Minwoo Kim (benchmark + sandbox, medium evidence) - Finance-specific red teaming reduced critical false negatives by about half in their evaluations and is already used in a regulatory sandbox.
• New · framework Forecasting AI-Era Productivity: The Intellectually Converged Human Framework and a Missing Cognitive Mediator in Production Function Theory: Kwan Soo Shin, In Seok Kang (theory, medium evidence) - Proposes “convergence capacity” as a mediator for total factor productivity from AI, offering a structural account of cross-country divergence.
• New · framework What Capital After Labor? Forecasting the Talent ROI Transition in the Human-AI Era: Kwan Soo Shin, In Seok Kang (quasi-experimental + panel, medium evidence) - Uses Korea’s 52‑hour law as an early signal to project SG&A (selling, general, and administrative) pressure from time compression and a shift to output accounting by 2032.
• Tension · suggestive Path Dependence, Governance, and the Limits of AI-Led Green Growth: Chantal Chelala, Rosette Ghossoub Sayegh, Nisrine Hamdan Saadé (System‑GMM dynamic panel, medium evidence) - After accounting for persistence and endogeneity, national AI ecosystem indices show no independent green-growth effect in this model and sample, with government effectiveness emphasized instead. System‑GMM is a dynamic panel estimator using generalized method of moments.
• Extends · suggestive LLM-Mediated Human-AI Interaction in Search and Rescue: Impact of Expertise on Attentional Allocation: Elahe Oveisi, Hemanth Manjunatha (RCT with eye-tracking, medium evidence) - LLM guidance improved per-step efficiency but not total rescues in a simulated setting; novices relied more passively while experts verified, highlighting interface trade-offs.

What Moved

• Persuasion at scale: Multiple preregistered trials plus a live fundraising deployment provide a more precise estimate of AI’s persuasive edge over trained humans in these contexts, shifting from lab-suggestive to field-grounded and more policy-relevant. Relative to earlier small or student-sample studies, this expands populations, incentives, and outcomes to real donations.
• Micro wins, macro caution: Additional firm-level uplifts (dispatch GMV, annotation speed, digital exports) strengthen the case that targeted LLM integration can pay in production, while a governance-aware cross-country panel challenges expectations that national AI build-outs independently drive green growth. The juxtaposition is editorial: it indicates complements and institutions are likely the binding constraints to aggregate gains.
• Distributional parsing by mechanism: New IV-based separation of augmentation vs automation helps separate AI types and associated labor outcomes (benefits to high-skill vs wage pressure on low-skill), and a robotization study in Turkey shows regional employment expansion alongside intensive-margin cuts, together refining how we read net gains versus incidence.
• Market structure via AI interfaces: Fresh audits and measurements observe that LLM assistants and AI search concentrate attention on incumbents and can be swayed by small rating or authority cues in these experiments and samples; relative to prior theoretical concerns, this week adds experimental and scaled measurement evidence.

Contested & Watch

• This finding: National AI ecosystem indices show no independent green-growth effect once persistence and governance are modeled (36 countries, 2017–2023, System‑GMM). Standing evidence: several firm/pilot difference-in-differences studies (medium evidence) find AI and intelligent-manufacturing policies associate with higher green innovation locally. One study does not reweight this. Watch: harmonized country panels linking micro adoption to macro emissions-intensity with governance interactions and instrumented adoption.
• This finding: Robot exposure is associated with increased district employment in Turkey via firm expansion (2014–2021, shift-share IV). Standing evidence: mixed robotization results across OECD settings (medium-to-high evidence) often show small employment effects with wage pressure. One study does not reweight this. Watch: matched employer-employee data separating entry, exit, hours, and wages across emerging and advanced economies.
• This finding: AIs out‑persuade expert humans and raise donations in the field (≈19k conversations plus one deployment). Standing evidence: earlier small RCTs and online crowdworker studies (suggestive) found modest AI persuasion gains. One study does not reweight this. Watch: platform-scale A/B tests with disclosure treatments, demographic heterogeneity, and long-run attitude durability.
• This finding: LLM assistants overweight eco labels and list position in hotel recommendations, with price-equivalent shifts (~$12/night), and AI search visibility concentrates on household brands (100k+ responses). Standing evidence: general recommender and SEO biases (descriptive) but sparse assistant-specific quantification. One study does not reweight this. Watch: ecosystem-level logs from assistants and AI search with randomized ranking/label nudges.
• This finding: LLM guidance in simulated search-and-rescue improves per-step efficiency but not total victims saved; novices become more passive (RCT with eye-tracking). Standing evidence: firm RCTs in support roles (high-to-medium evidence) often show output and sometimes quality gains. One study does not reweight this. Watch: end-to-end outcome trials with verification prompts, attention-aware UI, and expert-novice splits.

Methods Spotlight

• Large preregistered persuasion RCTs plus a real fundraising field test: AI systems out-persuade expert humans in these trials. Combines preregistration, expert comparators, and monetary outcomes to credibly size persuasion effects.
• LLM-derived semantic profiling with live production A/B: ProfiLLM at DiDi. Illustrates how to embed LLM features in core dispatch and validate with business KPIs, a template for productized experimentation.
• Counterfactual matched profiles for action-level fairness: AgentFairBench. Provides cheap, reproducible tests that vary only protected-attribute proxies to evaluate agent actions, not just text outputs.

The Week Ahead

• Stand up domain-specific A/B tests and instrument for complements before scale-up; treat workflow, data quality, and financing channels as first-class levers.
• Prepare disclosure and throttling policies for AI-mediated persuasion in campaigns, fundraising, and platforms; test effects on conversion and trust.
• Build governance capacity alongside AI spending, skills, data stewardship, and inter-agency coordination if you want macro or green-growth payoffs.
• Add verification checkpoints and attention-aware UI in high-stakes human-AI teaming; audit for oracle strength in agent-authored tests.
• Track AI search and assistant visibility for brand concentration; run controlled rank/label experiments to preempt manipulation and anticompetitive dynamics.

Reading List

• AI systems out-persuade expert humans: https://arxiv.org/abs/2606.16475
• ProfiLLM: Utility-Aligned Agentic User Profiling for Industrial Ride-Hailing Dispatch: https://arxiv.org/abs/2606.18803
• Artificial Intelligence, Skills, and Labor Mobility: Understanding the Transformation of Work: https://openalex.org/W7165427545
• Whose hotel does the AI recommend? An algorithm audit of reputation signals in LLM-assisted hotel selection: https://arxiv.org/abs/2606.16344
• AgentFairBench: Do LLM Agents Discriminate When They Act?: https://arxiv.org/abs/2606.16723
• AI Adoption in Local Government: Productivity, Systemic Risk, and Institutional Resilience: Evidence from a PRISMA 2020 Review: https://doi.org/10.3390/systems14060671
• From testbeds to high-stakes work: a review of Human-AI teaming domains and teaming factors: https://doi.org/10.3389/frobt.2026.1733942
• Robots, Employment and Wages: Evidence from Turkish Labor Markets: https://doi.org/10.22440/wjae.12.1.3
• The Impact of Artificial Intelligence Development on Firms’ Educational Composition of Labor: https://doi.org/10.54097/b49e5w74
• Strategic Feature Selection: https://arxiv.org/abs/2606.18867
• All Smoke, No Alarm: Oracle Signals in Agent-Authored Test Code: https://arxiv.org/abs/2606.18168
• A Risk Decomposition Framework for Pre-Hoc Fine-Tuning Prediction: https://arxiv.org/abs/2606.17649
• Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems: https://arxiv.org/abs/2606.17443
• Model Validation of Agentic AI Systems: A POMDP-Based Framework for Belief-State, Forecast, and Policy Validation: https://arxiv.org/abs/2606.17383
• When Agent Automation Becomes Profitable: Quantifying and Insuring Autonomous AI Risk through Trace-Economic Underwriting: https://arxiv.org/abs/2606.16465
• How Does Artificial Intelligence Policy Boost Green Innovation in Manufacturing? A Quasi-Natural Experiment Based on the AI Pilot Zones Policy: https://doi.org/10.3390/su18126139
• Computing power infrastructure and corporate financialization: evidence from China’s supercomputing centers: https://doi.org/10.1186/s40854-026-00946-5
• Digital pathways to high-quality and sustainable agricultural exports: evidence from the EU market: https://doi.org/10.3389/fsufs.2026.1750235
• Contaminated Collaboration: Measuring Gender Bias Transfer in LLM-Assisted Student Writing: https://arxiv.org/abs/2606.15914
• Intelligent Manufacturing Dynamic Capabilities and Corporate Green Innovation: Empirical Evidence from China: https://doi.org/10.3390/su18126053
• Exploring the Relationship Between Human-Centric AI and firm Idiosyncratic Risks: https://doi.org/10.1007/s10796-026-10759-7
• Artificial Intelligence as Game Changer in Cybersecurity: What We Learned in 2025-2026, and how this is relevant for Africa: https://arxiv.org/abs/2606.20102
• Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines: https://arxiv.org/abs/2606.20065
• Speeding up the annotation process in semantic segmentation industrial applications: https://arxiv.org/abs/2606.19934
• Labor Market The Impact of Artificial Intelligence on Employment Skills: https://doi.org/10.63056/academia.5.3(s7).2026.2080
• FFinRED: An Expert-Guided Benchmark Generation and Evaluation Framework for Financial LLM Red-Teaming: https://arxiv.org/abs/2606.19887
• Forecasting AI-Era Productivity: The Intellectually Converged Human Framework and a Missing Cognitive Mediator in Production Function Theory: https://arxiv.org/abs/2606.19794
• What Capital After Labor? Forecasting the Talent ROI Transition in the Human-AI Era: https://arxiv.org/abs/2606.19846
• Path Dependence, Governance, and the Limits of AI-Led Green Growth: A Dynamic Panel Analysis of 36 Economies: https://doi.org/10.3390/su18126274
• LLM-Mediated Human-AI Interaction in Search and Rescue: Impact of Expertise on Attentional Allocation: https://arxiv.org/abs/2606.19514