The Commonplace · Nov 02
|
The Commonplace
Weekly Research Digest · November 02, 2026
|
The Delta
What Moved & What Held
Coming in, the standing view was that AI often lifts output quality and organizational efficiency when paired with governance and workflow design, but capital deepening and replacement cycles weigh on firm-level productivity accounting; short-run learning gains are uneven and junior on-the-job formation can erode; and labor-market effects remain mixed with polarization pressures. These baselines rest on prior evidence about heavy and recurring capital expenditures (capex) and financing (infrastructure carry and balance-sheet pull), governance-contingent efficiency and measurement payoffs, documented quality gains in randomized controlled trials (RCTs), and risks to apprenticeship.
This week adds three pieces: verification capacity emerges as a likely bottleneck that helps explain why headline AI potential may translate into much smaller realizable value, clarifying why some firm-productivity claims may not show up in practice without better checking; metacognitive scaffolds again raise AI-assisted output quality with transfer, and a field-calibrated credential study finds no post-AI collapse in the predictive power of admissions metrics for licensure; finally, new theory and quasi-field evidence point to synchronization and firm-boundary reorganization as first-order constraints and channels, with limited within-model agent heterogeneity to diversify behavior. Still holds this week: AI assistance improves written-output quality; governance and workflow shape net organizational gains; and investment and adoption pressures remain elevated.
Top Papers
Key: each paper is tagged Relation (New, Confirms, Extends, Tension, Challenges) and evidence status (established, suggestive, framework, descriptive). Study design (RCT, quasi-experiment) is shown separately in parentheses. full key
Newframework
The Price of JudgmentVerification Capacity, AI Fluency, and the Realizable Value of AI
Agelo Thalassinidis
theoretical accounting, framework evidence
The paper prices verification time, verifier wages, redo costs, and escaped-error losses and estimates that about $649bn, roughly 22% of a $3.01tn headline AI value, is realizable today in this accounting; modest gains in verification speed and detection can move material value on these estimates.
So what: if this holds, ROI models that omit verification capacity understate value at risk and overstate near-term gains.
Newestablished
Do Academic Credentials Still Predict Professional Competence After Generative AI?
Doaa M. Bamasoud, Mohammed Ayid Alqahtani, Rasheed Mohammad
quasi-experiment, high evidence
Using US cohort-linked law-school credentials and first-time bar outcomes spanning pre- and post-2022, the credential-to-licensure relationship does not appear to weaken and is, if anything, stronger after generative AI's release.
So what: if this generalizes, overcorrecting away from formal credentials risks mis-screening where licensure-quality still aligns with admissions signals.
Newestablished
The Moral Character of Humans Behind AI Shapes Behavioral Trust in Large Language Models
Cassandra Grützner, Paweł Niszczota
RCT, high evidence
In a preregistered two-stage experiment, higher-scoring upstream humans prescribe more other-regarding AI behavior and downstream users trust an AI more when told its guide had high moral character, with larger trust losses from negative disclosures.
So what: If this holds, reputational negatives about human shapers can disproportionately depress user trust in AI-assisted processes you rely on.
Also Notable
Newdescriptive How Heterogeneous Is Homo Silicus? Measuring Risk-Preference Dispersion in Populations of Large Language Model Agents Gabriel Song: Within-model personas produce little durable risk-preference heterogeneity, with near-zero between-agent variance without personas and much smaller dispersion than human benchmarks.
Newframework Decorrelating the Advice Layer: A Calibrated Fragility Metric and Tested Mitigations for Synchronized Liquidation in Agentic Wealth Management Saurabh Kohli: In calibrated simulations, agentic automation amplifies index tails roughly with the square root of adoption, and decorrelation and staged execution are associated with lower tail liquidation with little welfare loss in these tests.
Newdescriptive Artificial Intelligence Investment and Economic Outcomes in the United States Harrison Margolin: AI-related capex rises to historic shares of GDP while short asset lives highlight return timing risk and employment effects remain weakly linked to exposure.
Newsuggestive AI and the Boundary of the Firm: Evidence from Radiology Marco Francesconi, Xiaoyu Xia: AI-enabled verification is associated with easier outsourcing and a recentering of reads to larger platforms, without clear increases in per-radiologist output.
Confirmsestablished Metacognitive Scaffolds Create Productive Friction for AI-Assisted Knowledge Work Viktor Kewenig, Lev Tankelevitch, Sean Rintel, Advait Sarkar, Abigail Sellen: An RCT shows scaffolds for goal-setting and decomposition improve AI-assisted writing and transfer to a subsequent task, even as perceived load rises.
Newdescriptive TaskNet: Building Occupational Task Inventories from Worker Interviews at Scale Jennifer Wang, Andreas Haupt, Isaac Robinson, Phil Lewis, Erik Brynjolfsson, Diyi Yang: Interview-driven task statements accepted by O*NET (Occupational Information Network) expand task coverage and shift AI-exposure estimates for some occupations.
Newsuggestive Practice Makes Perfect: Tacit Knowledge, Generative AI, and Returns to Schooling and Experience Eduardo Levy Yeyati, Darío Judzik, Juan Ignacio Barbiero: Post-ChatGPT, returns to schooling appear to fall in high-exposure jobs and the experience premium appears to rotate, consistent with AI possibly substituting for codified knowledge and complementing early tacit experience.
Newsuggestive Data Centers and the Emerging Contentious Politics of AI Simone Paci, Emily Russell: Since late 2022, data-center announcements are followed by more local mobilization and restrictive bills, suggesting politicized infrastructure siting.
Newsuggestive Valuing Anthropic and OpenAI Under Model Distillation: The Lead over Open-Weight Models Implied by AI Lab Valuations, 2024-2026 Danila Krasnov: A short capability lead can justify large valuations under low elasticity, but price premia can decay on month-scale constants.
Newdescriptive TreasuryQA Bench: Can AI Agents Work an Archive of Scanned Financial Records? Ritiz Tambi, Sahar Shayegan, Joshua Meyer, Sun Kim, Ali Khan, Victor Shih, Mehdi Jamei, Andi Partovi: Top agents solve about two-thirds of archival PDF tasks while recurring OCR (optical character recognition)/blur failures and repeated wrong answers expose brittle spots.
Newdescriptive Auditable large language model scoring of annual-report evidence packs: human calibration, baselines, and reproducibility checks Jie Gao, Junyuan Zhao: An evidence-bounded scoring workflow replicates human disclosure ratings and outperforms lexicon and term frequency-inverse document frequency (TF-IDF) baselines on shared packs.
Newestablished Incentivizing Adoption of Autonomous Delivery: A Randomized Field Experiment Shuai Hao, Zhanzhi Zheng, Yuqian Xu, Anindya Ghose: Multi-stage promotions convert trials to sustained adopters better than one-offs, with reliability and over-deep discounts hurting retention.
Newsuggestive AI Pilot Zones and Labor Market Demand: Evidence from China Jiayi Wei: Pilot-zone designation is associated with higher local postings across skill levels and with wider posted wage gaps between high- and low-skill roles.
Newsuggestive Delegating to the Machine: How People Hand Work to AI, and Whether Equity Markets Priced it Ajai Upadhyaya: Delegation varies by task and country, rises within task clusters, and is not clearly priced in equities with available power.
Newsuggestive Global SHAP Importance and Adjusted Association Diverge for Constructed Survey Indicators: Child Labour and Foundational Learning in Bangladesh Ishtiaq Mahmud Fahim, Samia Sulatana Reti, Sahil Raiyan: Strong negative associations can receive near-zero SHAP (a Shapley-based feature-importance method) importance in boosted models, suggesting feature-importance should not be read as effect size.
Newsuggestive Green Innovation Through Artificial Intelligence: Sectoral and Regional Perspectives of China Tahir Yousaf, Qurat ul Ain: AI adoption correlates with greater firm-level green innovation, especially in eastern regions and non-SOEs (state-owned enterprises).
What Moved
Contested & Watch
Methods Spotlight
TaskNet worker-interview pipeline (P8): scales qualitative interviews into validated task statements accepted by O*NET (Occupational Information Network), shifting exposure estimates and offering a richer base for policy and research on task-level AI effects.
Calibrated fragility index with decorrelation/staging experiments (P5): links flows and returns to a predictive tail-risk metric and tests operational mitigations that substantially reduce synchronized liquidation in the calibrated runs.
TreasuryQA scanned-archive benchmark (P24): end-to-end agent evaluation on realistic government PDFs surfaces OCR/blur failure modes and repeated-error behaviors invisible in cleaner benchmarks.
