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

• Strengthened: Evidence that ethics-governance and institutional accountability, not model capability alone, shape adoption and perceived legitimacy gained weight this week.
• Newly observed: A randomized experiment finds that disclosing AI as the decision-maker in corporate social responsibility (CSR) lowers perceived competence and authenticity, and a separate survey links over-explaining generative AI (GenAI) to reduced delegation via cognitive overload.
• Better measured: Simulations indicate AI adoption can lower true market beta but typical samples lack power to detect it, helping explain why many asset-pricing tests may return nulls.

What Moved & What Held

Coming in, the standing view was that AI's economic effects hinge on governance, trust, and organizational design, with transparency a necessary but insufficient condition for legitimacy and with full automation in expert work rarely succeeding because tacit judgment matters.

This week adds causal and synthesis weight: a systematic distrust review synthesizes reasons disclosure alone often does not repair trust; a randomized controlled trial (RCT) finds AI-attribution can backfire in CSR; a survey reports a transparency-overload delegation paradox; qualitative small and medium-sized enterprise (SME) evidence describes hidden verification labor; and Monte Carlo work illustrates why market-level AI effects can be empirically hard to see. Still holds this week: organizational redesign and human integration remain key constraints on AI productivity and acceptance.

Top Papers

• New · suggestive The antecedents, consequences and repair mechanisms of public distrust: A systematic literature review — Agni Shanti Mayangsari, Badri Munir Sukoco, Reinhard Bachmann, Juansih, Elisabeth Supriharyanti (systematic review, suggestive)

  • Synthesizing 54 studies across sectors and geographies, the review suggests transparency alone is often insufficient to repair distrust and that relational and structural reforms are frequently required alongside disclosure.
  • So what: If this generalizes: over-reliance on transparency tools risks stalled adoption because the trust gap sits in governance and accountability rather than documentation.
  • Full numbers

• Extends · established AI-Led or Human-Led? Disclosure of the CSR Decision-Maker, Motive Attribution, and Perceived CSR Authenticity — Keonyoung Park, Dongqing Xu, Jiamin Xie (RCT, high evidence)

  • A randomized controlled experiment finds that stating an AI system made the CSR decision reduces perceived competence, fairness, and authenticity relative to human-led disclosure.
  • So what: If this holds, reputation risk from AI-attribution can offset intended legitimacy benefits, especially where stakeholders read moral intent into corporate actions.
  • Full numbers

• New · descriptive Ethnographies of Human‐AI Collaborations: What Arrangements of Expertise Emerge? — Netta Avnon (ethnographic synthesis, descriptive)

  • Across professions and platforms, the synthesis identifies four recurring human-AI arrangements, noting that attempts at full automation frequently fall short, while hybrid co-production and algorithmic control are linked to divergent productivity and power outcomes.
  • So what: In this sample, realized value depends on role and accountability design, so productivity gains may be uneven even with the same technical system.
  • Full numbers

Also Notable

• New · descriptive Branching regulatory biographies of GPT-4 and DeepSeek-R1: A BOAP-informed analysis of regulatory reachability in the European Union, the United States and China — Yiping Cao Documentary analysis argues open-weight forks can increase competition but may fracture regulatory reachability, while hosted branches may remain more directly governable.
• New · suggestive Organizational AI adoption and financial risk: a Monte Carlo asset pricing framework and portfolio analysis — Evanthia K. Zervoudi, Apostolos Christopoulos Simulations indicate firm-level AI adoption can lower true market beta but the effect is often hidden by estimation noise in realistic samples.
• New · suggestive Maintenance Dynamics in Digital Health Public Goods: A Quasi-Experimental Study of Issue Resolution During the 2025 USAID Funding Disruption — Jinyou Sheng, Amy Finnegan A donor shock is associated with differential changes in bug resolution, with revenue-driven open-source projects slowing while donor-driven ones stayed stable in this natural experiment.
• New · descriptive Invisible gains: how resistance contests the promised social benefits of generative AI in SMEs — Alejandro Ramirez Dutch SME cases report hidden verification work and accountability gaps accompanying GenAI, complicating net efficiency claims.
• Extends · suggestive The primacy of ethical governance: Unraveling the AI-HRM adoption paradox in an emerging economy — Aunchistha Poo-Udom A Thailand survey links ethical governance more strongly than technical ease to human resources (HR) practitioners’ intent to use AI.
• Tension · suggestive How Does Green and AI ‐Related Innovation Activity Influence Ecological Decoupling in the U.S.? Evidence From Aggregate and Industry‐Adjusted Measures — Md. Rashed, Md. Kamal Uddin, Md. Naeemur Rahman, Mohammad Fakhrul Islam, A. K. M. Mohsin, Md. Faisal‐E‐Alam Time-series associations suggest long-run national improvements alongside short-run industry-intensity rebounds when AI and green innovation co-move.
• New · descriptive Should Businesses Trust AI Advice? A Methodology to Audit the Ethical Integrity of Chatbots — Manuel Chaves-Maza The Adaptive Ethical Evaluation Protocol (AEEP) audit reports consistent cross-model differences and high correlation with expert ratings, presenting an operational ethics screen for large language model (LLM) advisors.
• Extends · suggestive Using large language models for legal decision-making in Austrian value-added tax law: a comparative study — Marina Luketina, Andrea Benkel, Christoph G. Schuetz Fine-tuned and retrieval-augmented generation (RAG) LLMs assist routine value-added tax (VAT) analysis yet still hallucinate, indicating limits to full automation in advisory workflows.
• Extends · suggestive Dual-pilot policy, supervisory technology and overseas technology investment: Evidence from high-tech firms in China — Hanrui Wu, Jingyi Li, Yao Chen, Dewen Liu A quasi-experimental panel links complementary place-based pilots to higher overseas tech investment, moderated in an inverted-U by supervisory procurement.
• Tension · descriptive Decision delegation to GenAI agents in travel planning: Responsible AI signals, delegation levels, and the transparency paradox — Sanjit K. Roy, Gaganpreet Singh, S. Mostafa Rasoolimanesh, Ronnie Das, Ali N. Tehrani Survey evidence associates reliability/accountability signals with higher delegation but finds excessive explainability is associated with lower delegation via cognitive overload.
• New · suggestive When diversity cuts both ways: top management team heterogeneity, AI strategic orientation, and firm value creation efficiency — Zhidi Yin, Jiamei Che In this panel, associations link top management team (TMT) heterogeneity to lower value-creation efficiency but higher disclosed AI strategic orientation, indicating disclosure may not track implementation.
• New · descriptive Artificial intelligence adoption in accounting and auditing: The technology–organisation–environment (TOE) framework perspective — Betül Alkan Interviews indicate managerial support is associated with adoption while data-integration and regulatory uncertainty are reported as constraints, especially outside Big Four contexts.
• New · framework Minds and machines: Rethinking absorptive capacity in the age of artificial intelligence — Mattia Pedota Proposes AI-specific absorptive capacity dimensions around data/model acquisition and human–AI orchestration.

What Moved

• Governance and trust: A systematic review synthesizes evidence that transparency without relational and structural repair rarely rebuilds trust, and the CSR RCT provides causal evidence that AI-led disclosure can depress perceived authenticity; paired with a survey suggesting transparency overload, the balance tilts further toward governance quality over disclosure volume.
• Roles and hidden labor: Ethnographic synthesis and SME cases reinforce that hybrid arrangements create new verification and coordination work that often goes unrecognized, while a workflow pilot suggests load can fall when roles and tooling are redesigned, highlighting heterogeneity that is likely across firms.
• Market measurement: Monte Carlo asset-pricing work clarifies that even if AI adoption shifts true betas, finite-sample noise can mask the signal, implying that recent nulls in empirical tests may be power-limited rather than effect-absent.

Contested & Watch

• Transparency helps or hurts acceptance

  • Finding: An RCT finds AI-led CSR disclosure lowers perceived competence, fairness, and authenticity (online participants, randomized treatment).
  • Standing evidence: A systematic review finds transparency is necessary but insufficient and works better with relational and structural reforms (54 studies, cross-sector).
  • Watch: Multi-site field tests that pair disclosure with accountable governance to see if backfire effects attenuate.

• Full automation versus expert augmentation

  • Finding: In VAT advisory, fine-tuned and retrieval-augmented LLMs assist routine analysis but still hallucinate and miss client context, requiring human oversight in these tests (comparative experiments).
  • Standing evidence: Ethnographies report that attempts to remove humans from expert work largely fail due to tacit knowledge and contextual judgment.
  • Watch: Logged field deployments with measured error costs and escalation rates in legal and tax settings.

• Hidden labor versus cognitive-load relief

  • Finding: In Dutch SMEs, GenAI introduces unrecognized verification work and accountability gaps around outputs.
  • Standing evidence: An interventionist platform study suggests AI can reduce cognitive load for accountants and shift effort toward strategic advice when workflows are redesigned.
  • Watch: Time-use panels that split visible and hidden tasks before and after role redesign with recognition changes.

• Regulatory reachability in open-weight ecosystems

  • Finding: A regulatory-biographies analysis argues that redistribution via open weights fractures regulator reach while hosted branches stay more reachable.
  • Standing evidence: Case work on climate-risk analytics finds private rating firms centralize and standardize metrics, concentrating governance power in intermediaries rather than at the model level.
  • Watch: Provenance and audit pilots that test branch-sensitive obligations across hosted and redistributed deployments.

Methods Spotlight

• Adaptive Ethical Evaluation Protocol (AEEP), Should Businesses Trust AI Advice? A Methodology to Audit the Ethical Integrity of Chatbots: A branched audit that reports strong correlation with expert judgments, enabling repeatable ethics screening of large language model (LLM) advisors.
• Monte Carlo CAPM for AI adoption effects, Organizational AI adoption and financial risk: a Monte Carlo asset pricing framework and portfolio analysis: Quantifies mechanical beta shifts from adoption and the finite-sample power limits that obscure them in practice under the capital asset pricing model (CAPM).
• Dynamic fsQCA with DEMATEL-AISM, How multiple pressures shape enterprises’ Innovation Paths? A case analysis of China’s new energy vehicle (NEV) sector: Maps equifinal policy-market-resource configurations that yield high innovation and identifies core-periphery influence among pressures using fuzzy set Qualitative Comparative Analysis (fsQCA) with Decision-Making Trial and Evaluation Laboratory (DEMATEL) and a variant of Interpretive Structural Modeling (AISM).