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

Correction: In last week's author note, I described the linked NFW Reader piece incorrectly. It was about using Copilot in Word for causal-impact work, not Copilot Studio. The link itself was correct. I also included a reader-facing link for the OES Dashboard that did not resolve publicly; the correct link is the OES Dashboard. The archive has been corrected.

Most discussion of AI agents starts with worker productivity. "The Agentic Economy" asks a second economic question: what changes when assistant agents and service agents can communicate directly? The visual Reader piece maps the lower communication costs that could follow, the choice between platform-run walled gardens and an open web of agents, and six markets that could shift.

The Delta

Coming in, Research Productivity leaned positive (373 papers); this week, a counter-signal appears. - Strengthened: a large randomized trial finds AI-assisted reviewers match humans while AI-led teams miss many more errors, supporting the view that oversight, not autonomy, yields higher quality on complex verification tasks in this task. - Better measured: quasi-experimental US clickstream data indicate ChatGPT Search generates outbound clicks in only 5.2% of sessions and is associated with about a 9.4% decline in traditional search use during rollout, providing a magnitude for referral compression in this sample. - Strengthened: population-wide Norwegian registers using multiple difference-in-differences designs find no clear early displacement of young workers after the ChatGPT release in highly exposed occupations.

What Moved & What Held

Coming in, the standing view was that AI works best as an augmentor alongside skilled humans, not a drop-in replacement; early labor-market impacts look limited and heterogeneous; firm-level productivity gains appear in pockets while macro productivity remains uncertain; and AI intermediaries may redirect value in digital markets.

This week adds hard causal evidence that AI-led workflows can fail on critical error detection even when AI assistance helps, national administrative data that add weight against near-term displacement fears for early-career workers, and estimated magnitudes for how AI search compresses outbound web referrals. Orchestration and memory studies report sizable cost and latency improvements without observed quality loss in tested tasks, and in South Korea, hours adjustments appear in high-exposure industries rather than headcount losses. Still holds this week: augmentation over automation, limited immediate displacement, uneven firm gains, and unresolved micro-to-macro scaling.

Top Papers

• Confirms · established AI-assisted teams outperform AI-led teams but not human-only teams in assessing research reproducibility in quantitative social science (Abel Brodeur, David Valenta, Alexandru Marcoci, Juan P. Aparicio, Derek Mikola, Bruno Barbarioli, Rohan Alexander, Lachlan Deer, Tom Stafford, Lars Vilhuber, Gunther Bensch, Fabio Motoki, Mohamed Abdelhady, Yousra Abdelmoula, Ghina Abdul Baki, Tomás Aguirre, Sriraj Aiyer, Shumi Akhtar, Farida Akhtar, Melle R. Albada, Micah Altman, David Angenendt, Zahra Arjmandi Lari, Jorge Armando De Leon Tejada, David Rodriguez Arana, Igor Asanov, Anastasiya-Mariya Noha, Rebecca Ashong, Tobias Auer, Francisco J. Bahamonde-Birke, Bradley J. Baker, Söhnke M. Bartram, Dongqi Bao, Lucija Batinovic, Tommaso Batistoni, Monica Beeder, Louis‐Philippe Beland, Carsten Gero Bienz, Christ Billy Aryanto, Cylcia Bolibaugh, Carl Bonander, Ramiro Bravo, Egor Bronnikov, Stephan Bruns, Nino Buliskeria, Sara Caicedo-Silva, Andrea Calef, Juan Sebastian Cano Arias, Gustavo A. Castillo Alvarez, Solomon Caulker, Simonas Cepenas, Arthur Chatton, Zirou Chen, Ngozi Chioma Ewurum, Anda-Bianca Ciocîrlan, Felix J. Clouth, Jason Collins, Nikolai Cook, Cesar Cornejo, Joao Craveiro, Jonathan Crechet, Jing Cui, Niveditha Chalil Vayalabron, Christian Czymara, Carlos Daniel Bermúdez Jaramillo, Hannes Datta, Lien Denoo, Arshia Dhaliwal, Nency Dhameja, Elodie Djemai, Erwan Dujeancourt, Uğurcan Dündar, Thibaut Duprey, Yasmine Eissa, Youssef El Fassi, Ismail El Fassi, Keaton Ellis, Ali Elminejad, Mahmoud Elsherif, Aysil Emirmahmutoglu, Giulian Etingin-Frati, Emeka Eze, Jan Fabian Dollbaum, Jan Feld, Andres Felipe Rengifo Jaramillo, Guidon Fenig, Victoria Fernandes, Lenka Fiala, Lukas Fink, Mojtaba Firouzjaeiangalougah, Sara Fish, Jack Fitzgerald, Rachel Forshaw, Alexandre Fortier-Chouinard, Louis Fréget, Joris Frese, Jacopo Gabani, Sebastián Gallegos, Max C. Gamill, Attila Gáspár; randomized controlled trial) - In a randomized controlled trial that assigned 288 researchers into 103 teams across human-only, AI-assisted, and AI-led arms, AI assistance matched human reviewers on overall reproducibility while AI-led teams reproduced only about 37% and missed more major coding errors. This supports the standing view that augmentation helps but autonomy degrades quality on complex verification tasks in this task. - So what: If this holds, delegating high-stakes verification to AI carries higher undetected error risk and reputational exposure than many current quality-assurance models assume. - Full numbers

• Confirms · suggestive Labor Market Consequences of Generative AI: Early Evidence from Norway (Dennis Facius, R. Iacono; quasi-experiment using population registers) - Using Norwegian population registers from 2015 to March 2025 and multiple quasi-experimental strategies (within-firm composition difference-in-differences, occupation-level synthetic difference-in-differences, and firm-level shift-share) that exploit the November 2022 ChatGPT release, the authors find no robust displacement of early-career workers in highly exposed occupations and no clear effects across other cohorts. This aligns with the baseline that near-term employment effects are muted and heterogeneous. - So what: If this holds, overestimating immediate displacement risk could misdirect training budgets and social insurance planning toward the wrong time horizon. - Full numbers

• Confirms · suggestive Answering Without Referring: How AI Search Rewrites the Web's Economic Bargain (Qiaoni Shi, Kai Zhu, Kai Gu; quasi-experimental clickstream study) - US Comscore desktop clickstream data indicate ChatGPT produces outbound clicks in only 5.2% of conversation sessions and that broader access to ChatGPT Search is associated with about 9.4% lower traditional search use during rollout, with clicks skewing to specialized sites rather than ad-heavy portals. This provides quantitative backing for the claim that AI intermediaries compress referral flows that fund the open web. - So what: If this generalizes, referral-dependent publishers and ad markets face sharper traffic and revenue compression than current operating plans contemplate. - Full numbers

Also Notable

• Confirms · established Experimental Evidence on the Learning Impact of Generative AI (Zara Contractor, Germán Reyes), Undergraduate randomized controlled trial finds AI access lifts immediate test scores by 0.27 SD with one-week persistence concentrated among students who use AI for explanation, consistent with augmentation over automation in learning.
• Extends · suggestive AI, Output, and Employment (Andrew Johnston, Christos A. Makridis), US administrative employer data associate higher occupational AI exposure with about 7% higher output and 4% higher employment where collaboration is essential, alongside a lower labor share, pointing to distributional shifts even when headcounts grow.
• Extends · suggestive Artificial Intelligence Exposure and Working Hours: Evidence From South Korea (Taiwon Ha), Industry-level exposure measures link higher AI exposure to larger post-2022 declines in weekly hours with no pre-trend differences, suggesting hours may adjust before employment in this context.
• New · descriptive AI Adoption in S&P 500 Firms (Yang Yu, Martin Fleming, Lucy Hampton, Christophe Combemale, Neil Thompson), A 10-K based deep-integration measure rises to 11% in 2025, quadrupling since 2022, with a J-shaped profit association and no clear capex or productivity signals in filings, documenting who is moving first.
• Confirms · established Artificial Intelligence and the Digital Economy: Impact on Employment, Productivity, and Market Structures (Dr. Snehal Mistry, Siddharth Thakkar), Systematic review of 78 studies reports firm-level productivity gains and routine task displacement often offset by complementary roles, with uneven distributional outcomes.
• Extends · suggestive The carbon reduction effect of China’s national AI innovation pilot zone policy (劉南勳, Shuqing Wang, Yuanhong Peng), Staggered city-level quasi-experiment links AI pilot-zone designation to about 6% lower urban CO2, rising to 15.6% with spatial spillovers, suggesting policy-associated environmental effects.
• Extends · suggestive Artificial Intelligence and Urban Green Productivity in China: The Role of Green Computing Capacity and Transmission Channels (Xiaoxiao Tian, Wei Guo, Jingyu Liao), City panels associate AI patenting with higher green productivity where green computing capacity and industry bases are strong, highlighting the role of complements.
• New · descriptive Seeing is Free, Speaking is Not: Uncovering the True Energy Bottleneck in Edge VLM Inference (Junfei Zhan, Haoxun Shen, Mingang Guo, Zixuan Huang, Tengjiao He), On-device profiling finds output token decoding dominates energy and latency for vision-language models, shifting where efficiency gains are likely to come from.
• New · suggestive The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI (Muayad Sayed Ali, Aliaksandra Novik, Anji Boddupally, Artem Yavorskyi, Chris Nickerson, Daniel Rica, Emily DuGranrut, Felix Leung, Garrett Prince, Grace Barnett, Heath Robinson, Hosain Al Ahmad, Jesse Resnick, Juan Carlos Farah, Jyothi Swaroop Meruga, Leonid Kuznetsov, Luke Gorham, Marie Schmoll, Michael Paciullo, Saumya Das, Sharath Sheripally, Tommy Griscom, Mykyta Osadchyi, Neha Mantri, Nick Westrum, Olivia Benowitz, Parikshith Kulkarni, Radik Chernyshov, Rakshith Vasudev, Rohith Nadimpally, Vikas Gangadevi, Waseem AlShikh), Controlled engineering experiments report 30–45% lower token costs and latency with orchestration layers while holding measured task quality constant, pointing to a potential lever for enterprise deployments.
• New · suggestive Shared Selective Persistent Memory for Agentic LLM Systems (Sanjana Pedada, Aditya Dhavala, Neelraj Patil), A memory design achieves 96% task completion while cutting tokens and time versus naive persistence in tested tasks, suggesting architecture choices can mitigate cost-quality tradeoffs.
• Tension · descriptive 3100 Opinions on Code Review in an AI World: Building Causal Theory from Practitioner Discourse (Shyam Agarwal, Courtney Miller, Christian Kästner, Bogdan Vasilescu), Repository mining suggests agent-authored pull requests are reviewed less and merged faster, but patterns flip under alternative analytic choices, underscoring governance ambiguity around AI-generated code.
• New · framework Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops (Mingguang Chen, Licheng Wang, Bo Qu), Survey distinguishes verifiable, bounded self-refinement from speculative open-ended self-improvement and argues current systems remain compute and grounding constrained.
• Extends · suggestive Two-player Alternate Uses Test: A Controlled Testbed for Interactive Human-AI and Human-Human Co-Creation (Babak Hemmatian, Anita Keshmirian, Yijun Lin, Shravan Ramamoorthy, Maryam Jahadakbar, Eli Khuri-Reid, Jingtong Wang, Sarah Hadjarab, Sindre Veum, Pranav Gupta, Deepak Somaya, Lav R. Varshney), In a small controlled setting, GPT-4 partners generate originality on par with human partners, conditional on motivation and prior exposure.
• New · framework Analysis of labor productivity in the context of technological transformations (T. Martyn, V. Nitsenko, V. Kyrylenko, O. Tkachenko, Y. Stavska, O. Kulhanik), Proposes an "AI phase" where organizational frictions and AI debt slow macro productivity even as micro gains accrue.
• Extends · suggestive Asymmetric effects of renewable energy and artificial intelligence on green growth: evidence from G20 countries (Olfa ZARRAD, Maha Bouattour, Sourour Guidara, Kamel Helali), Panel autoregressive distributed lag (ARDL) model suggests AI contributes short-run green gains and amplifies renewables' long-run effects, with CO2 levels damping impacts.
• New · framework What AI Cannot Learn: A Cognitive Science Perspective on Human-Centered Strategic HRM (Daniel Altieri, Zohra Damani, Cynthia Nebel), Argues current generative models lack experience-based judgment, offering a structure to preserve human accountability.
• Confirms · suggestive Artificial Intelligence-Aided Strategic Information System Tools for Competitive Position in the Market: A Systematic Review (Uchenna Nzenwata, Rabiu Ayantayo, Fiyinfolu Okadare, Samuel Owolabi), Synthesis indicates business intelligence (BI), decision-support, customer relationship management (CRM), and generative tools speed decisions and improve customer metrics when organizational readiness is in place.
• Extends · suggestive Technology adoption and bias in officiating: automated Ball-Strike System implementation in Korean Baseball (Jimin Song, Ji Hyuk Kang, Richard J. Paulsen), Natural experiment links automation to reduced advantages for high-status batters, illustrating redistribution effects from precision technologies.
• Extends · suggestive The Synergistic Effect of Digital Industry Agglomeration and Digital–Intelligent Technology on Green Productivity (S Liu, Hongyu Hè), Simultaneous equations suggest digital agglomeration and intelligent tech jointly reduce innovation misallocation and raise green productivity with diminishing returns.
• New · descriptive Leading in the Digital Age: Digital Leadership Capabilities, Organisational Innovation Climate, and AI Adoption Intention Among SMEs in Nigeria (A. Idowu, Y. Babalola), Cross-sectional evidence is associated with leadership capabilities and stronger AI adoption intention, partly via innovation climate.
• New · suggestive Conceptualization of causes and implications of AI adoption behavior in hospitality and tourism (Mohammad Alimohammadirokni, Ali Iskender, Nasrin Rasouli), Mixed methods point to context, strategy, and training as adoption drivers, with perceived efficiency and sustainability gains.
• New · descriptive AI AND THE TRANSFORMATION OF THE LABOR MARKET: THE SOCIAL CONSEQUENCES OF AUTOMATION AND THE NEW EMPLOYMENT UNCERTAINTY (N. Baigabylov, Alimzhan Yessenovabylov), Secondary synthesis suggests net job gains but large task churn and skill obsolescence, reiterating distributional exposure.

What Moved

• Human-AI collaboration limits: The multi-team randomized trial narrows uncertainty on team design by finding that AI-led verification performs substantially worse even as AI assistance helps, which tightens the boundary between augmentation and autonomy relative to past smaller studies and reviews. Engineering studies on orchestration and memory complement this by reporting where cost and latency gains come without observed quality loss in tested tasks.

• Early labor-market effects: Norway’s national registers add weight to the limited-displacement baseline by finding no near-term job losses for young workers in high-exposure roles after ChatGPT, while Korean industry data point to hours reductions rather than headcount cuts. Together, they shift the near-term concern from unemployment to work intensity, pay, and task composition.

• Market structure and referrals: US clickstream evidence estimates magnitudes for AI search’s low outbound-click rate and cannibalization of traditional search, moving from hypothesis to measured impact on the web’s referral economy and implying redistribution across publishers and AI intermediaries.

Contested & Watch

• Can AI-led teams safely replace human oversight in high-stakes verification? - Finding: In a 288-person randomized controlled trial, AI-led teams reproduce about 37% of results and miss more major errors than humans or AI-assisted teams. - Standing evidence: Multiple lab and field experiments suggest augmentation improves quality while full automation underperforms, but some agent architectures report high task completion in constrained domains. - Watch: Replications in other high-stakes domains with preregistered error audits and cost-quality frontiers.

• Are near-term labor impacts showing up more in hours than in jobs? - Finding: In South Korea, high-AI-exposure industries see larger post-2022 declines in weekly hours with no pre-trend differences. - Standing evidence: Several quasi-experimental and administrative studies find muted early displacement and heterogeneous effects, with some output and employment gains where collaboration is required. - Watch: Matched employer-employee panels with hours, earnings, and vacancy data to track substitution versus reallocation within occupations.

• Will AI search durably compress web referrals and ad revenue? - Finding: US desktop clickstream indicates 5.2% outbound clicks from ChatGPT sessions and about 9.4% lower traditional search usage during rollout. - Standing evidence: Prior descriptive work and platform reports hinted at substitution but lacked quasi-experimental measures and destination mix detail. - Watch: Mobile and international panels linking referral changes to publisher revenue and content production responses.

• Do firm-level AI gains translate to macro productivity on current trajectories? - Finding: S&P 500 disclosures show rapid deep adoption with a J-shaped profitability association but no clear capex or productivity signals in filings. - Standing evidence: Many firm studies report productivity improvements, while macro productivity effects remain inconsistent and may lag due to organizational frictions. - Watch: Sectoral multi-factor productivity in high-adopting industries and input-output spillovers that would close the micro-to-macro gap.

• Does AI exposure raise employment where collaboration is essential but lower labor share? - Finding: US employer data link a one standard deviation rise in AI exposure to about 7% higher output and 4% higher employment in collaborative tasks, alongside a lower labor share. - Standing evidence: Experimental evidence shows augmentation boosts task productivity, and several observational studies hint at shifting factor shares. - Watch: Wage-bill decomposition by task exposure and longitudinal worker earnings in high-collaboration occupations.

Methods Spotlight

• Multi-arm randomized trial of human-only versus AI-assisted versus AI-led teams (Brodeur et al.): rare causal evidence on collaboration design in a real scientific task clarifies where autonomy failed in this task.
• Population-wide administrative designs exploiting the ChatGPT release as an availability shock (Facius and Iacono): combining within-firm composition difference-in-differences, synthetic difference-in-differences, and shift-share on national registers tightens inference on early labor effects.
• Hardware-level energy profiling for vision-language inference (Zhan et al.): isolates output token decoding as the dominant energy and latency driver on edge devices, reframing optimization targets for deployment economics.