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

Coming in, Firm Productivity leaned positive (278 papers); this week, a counter-signal appears. - Strengthened: platform attention is reallocated, not just scaled, with an 8.56M-user Netflix randomized controlled trial (RCT) diffusing consumption from superstars to the middle tail while a preregistered search RCT finds AI overviews pull clicks on-platform and away from publishers. - Better measured: the publisher and user-experience costs of AI-answer search are now quantified causally (-18.8 percentage points in external click-through rate (CTR), small declines in trust and sessions). - Challenged: name-inference-driven equity fixes are not uniform, with new evidence suggesting that benefits concentrate on Western-legible women's names and often miss culturally ambiguous ones.

What Moved & What Held

Coming in, the standing view was that AI reallocates attention and effort: recommender and answer-synthesizing AI shift who gets traffic and rents; AI tools mostly augment knowledge work by moving time from routine processing to higher-value tasks; and governance and fairness interventions carry uneven benefits with auditing pitfalls.

This week adds clearer causal weight on the attention reallocation story in opposite directions across intermediaries: a massive Netflix holdback shows recommendation upgrades grow engagement while reducing superstar concentration, and a preregistered search experiment shows AI overviews sharply reduce outbound referrals and nudge down trust and session depth. It also qualifies equity tactics by documenting a legibility gap in name-based interventions and flags a common measurement trap in event-time designs around user-triggered AI features. Still holds this week: autonomy without human scaffolding remains limited, agent reliability requires repeated audits, and task-augmentation gains are heterogeneous across roles and settings.

Top Papers

• Confirms · established Recommendation quality and the concentration of consumption: Experimental evidence from Netflix (Guy Aridor, Winston Chou, Nathan Kallus, Antoine Scheid, Allen Tren, Kevin Zielincki; RCT, high evidence) - In a 60-day randomized holdback on a global platform with 8,559,252 subscribers, improved recommendations increase engagement and shift recommendation share away from superstars toward the middle tail, lowering title-level concentration by about 5.7 percent. This independently corroborates the standing view that recommender upgrades can broaden consumption rather than amplify hits, at scale. - So what: If this holds, attention and revenue risk sits with superstar-heavy catalogs and contracts, not just with the long tail. - Full numbers

• Confirms · established AI in search reduces publisher referrals without improving user experience: Experimental evidence (Stephanie T. Wang, Jeffrey Gleason, Yakov Bart, Christo Wilson, Danae Metaxa; RCT, high evidence) - A preregistered randomized controlled trial (N=1,100) finds AI-answers mode cuts click-through to external sites by 18.8 percentage points, reduces news and Reddit clicks, and slightly lowers trust and sessions versus current search; removing AI overviews modestly raises referrals without perceived quality gains. This tightens causal estimates around publisher harm and neutral-to-worse UX under AI-answers, aligning with prior concerns about on-platform answer capture. - So what: If this generalizes, publisher revenue exposure to search format shifts is larger and less offset by UX gains than many models assume. - Full numbers

• Extends · suggestive Beyond automation: AI and the human value of sell-side analysts (Devin Shanthikumar, Il Sun Yoo; quasi-experiment, high evidence) - Using difference-in-differences and event studies around US 10-K inline XBRL (iXBRL) adoption and bank AI investments, analysts at AI-invested banks are associated with timelier, bolder, and more accurate forecasts and expanded coverage, consistent with a shift from public-data processing to private information gathering. This extends augmentation evidence into high-skill finance, indicating reallocation toward higher-value tasks rather than displacement. - So what: If this holds, earnings-season information production could become more uneven across institutions, raising model and market-structure risk for firms relying on slower or thinner coverage. - Full numbers

Also Notable

• Extends · established Whose doctor does the AI recommend? An algorithm audit of reputation and demographic signals in large language model-assisted physician choice (Syeda Anshrah Gillani, Mirza Samad Ahmed Baig)
  Large preregistered conjoint audits find large language model (LLM) recommenders heavily weight ratings and fees and exhibit smaller, systematic gender/ethnicity and first-position biases, sharpening equity and steering concerns in healthcare referrals.

• New · established Event-time confounding under bursty human dynamics (Michael Iannelli, Alan Ai)
  Argues that aligning analyses on self-timed AI events can mechanically inflate post-event engagement because events occur within ongoing tasks, a measurement warning for digital behavior studies.

• Extends · established Targeting support using job seekers' biases: A randomized experiment (Bruno Crépon, Aurélien Frot, Christophe Gaillac)
  Personalized motivational machine learning (ML) nudges raise search effort and short-run reemployment for pessimistic job seekers, while generic recommendations lift applications without reemployment gains, highlighting heterogeneity-aware targeting.

• New · descriptive UpgradeBench: A decision-centric benchmark for upgrading fine-tuned LLM specialists (Ye Chen, Weining Zhang)
  Specialist adapters' advantages sometimes persist across base-model upgrades but often erode, suggesting upgrade durability is task- and lineage-dependent.

• Extends · established TRACE: Agentic catalog enrichment with multi-source evidence grounding (Rohan Kumar, Steven Xu, Kyle MacDonald, Matthew Long, Bernice Chow, Mac VanRenterghem, Sudeep Das)
  A verify-before-write judge in an e-commerce pipeline achieves near-expert attribute extraction and lifts checkout conversion by 0.48 percent in an RCT, a practical gain for catalog quality.

• Extends · established The legibility gap: How gender equity interventions redistribute recognition across cultures (Binglu Wang, Jose Cervantez, Jiahui Xue, Katherine L. Milkman, Dashun Wang)
  Experiments and observational analyses find name-based gender inference and citation-diversity prompts largely benefit Western-legible women's names, missing many others.

• New · framework The order of binary experiments under endogenous stopping (Zihao Li)
  Characterizes sequential information dominance via directed Kullback-Leibler (KL) divergences, informing design of costly observation policies.

• Extends · suggestive Auditing self-evolution in financial agents: Capability gains, security drift, and execution-interface mismatch (Jialong Li, Jialing Zhu)
  In AgentDojo-style audits (a standardized agent evaluation suite), some self-evolution methods improve accuracy in tests but also coincide with higher exposure to prompt injection and unauthorized state changes, a safety trade-off.

• Extends · suggestive Science under sanctions: The impact of the entity list on Chinese academic research (Xiaodie Pu, Xintong Wang, Di Tong, Alain Yee Loong Chong)
  Stacked event studies suggest entity-list shocks are associated with lower publication counts but higher average bibliometric quality and a reorientation of collaborations toward domestic intermediaries.

• Tension · suggestive Stranded credentials: how a skill-signaling market absorbed generative AI (Song Yao)
  Kaggle medals predict performance mainly within one year and lose value faster in the AI era, indicating accelerated credential obsolescence.

• New · descriptive Pricing the risk of runtime compression: Anytime-valid admission and a served-output law for compressed serving state (Fanzhe Wei, Li Liu)
  In one production deployment, a formal-bounds admission ledger was associated with about half as many fallbacks to exact serving while making quality-capacity trade-offs explicit.

• Extends · suggestive Governing generative AI in organizations: a design theory and quasi-experimental field study of sociotechnical guardrails (Maikel Leon)
  A stepped-wedge field study estimates that guardrails are associated with fewer hallucinations and better auditability but slower tasks and some circumvention.

• New · descriptive No task fails every time: Why one-shot audits are structurally blind to agent damage (Shiven Khurdi)
  Repeated, ground-truth state-diff audits reveal stochastic, sometimes irreversible agent damage that single-run audits miss.

• New · descriptive Frontier AI forecasting has a measurement problem: An audit of progress evidence (Fabricio F. Costa)
  Public progress records are sparse, versioned, and concentrated in few sources, undermining simple compute-to-capability extrapolations.

• Tension · descriptive AI with authority, from application to silicon (Jason Hickey)
  A specialist-led agent fleet documents a case in which a verified software-to-silicon stack and tapeout occurs in five weeks, a standout case against the general limits seen in autonomous science benchmarks.

• Tension · descriptive BC-Bench: Evaluating agentic engineering in a domain-specific language for ERP (Haoran Sun, Klaus Marius Hansen)
  In this evaluation, gains on public code benchmarks did not reliably transfer to enterprise resource planning (ERP) domain-specific language (DSL) tasks, a warning on external validity.

• New · framework Foundation models for partial causal identification (Alexis Bellot, Anish Dhir)
  Proposes how prior-fitted causal foundation models can be used to recover identified sets for partially identified queries.

• Tension · descriptive ContractScrub: A benchmark for final review of legal contracts (Yejin Bang, Kirsty Fielding, Brandan Oliver, Brian Birke, Nabeel Seedat, Andrew M. Bean)
  Frontier LLMs still struggle on long-context legal scrubbing (best macro recall ~0.75, F1 <0.65), highlighting limits in compliance-heavy domains.

• New · suggestive Credit without ground truth: Auditing step-level credit assignment in LLM agents against executed replay (Haiyue Zhang)
  Executed-replay audits suggest popular credit signals track fluency, not causal contribution, complicating reward design.

• Extends · suggestive Platform labour participation and the division of household labour: evidence from the 2023 Chinese Social Survey (Mingzhe Cui, Han Wang, Chang Li, Dingxuan Wang, Yingying Zheng)
  Platform work correlates with more household outsourcing, higher spouse employment, and lower education spending in China, suggesting nontrivial household reallocations.

• New · framework Beyond weaponized interdependence: the corporate chokehold in frontier semiconductor governance (Nik Hynek)
  Argues firms' operational control over service, qualification, and scheduling can cap effective production scale beyond what state export controls dictate.

• Extends · descriptive CentaurBench: Benchmarking LLM capabilities on augmenting vs. automating real-world work tasks (Pattaraphon Kenny Wongchamcharoen, Kris Gulati, Min Min Fong, Abhishek Nagaraj)
  Model rankings diverge for augmentation versus automation, suggesting different selection criteria for assistive use.

• Extends · suggestive A jagged frontier: Evaluating robustness of code agents to semantics-preserving transformations (Hasan Najib Mahmud, Shreya Gupta, Isha Chaudhary, Nathaniel Enis, Ravi Mangal, Gagandeep Singh, Corina Pasareanu)
  In these tests, semantics-preserving code rewrites often reduce repair success and raise costs, with robustness varying by model and scaffold.

• Extends · descriptive StagedWorkspace: A versioned workspace for knowledge-work agents (Yining Hua, Hongbin Na, Yifan Zhou, Akshay Kalose, Cyrus Ayubcha, Levi Lian)
  Binding parsed and native-file views via hashes is associated with better agent performance on knowledge-work tasks.

• Confirms · descriptive ASI-Bench: At the dawn of artificial superintelligence (Junwei Zhou et al.)
  In this benchmark, agent performance drops sharply without human methodological guidance across 60 scientific tasks, reinforcing limits to autonomous discovery.

• Extends · suggestive Institutional empowerment: How national pilot zones for innovative development of artificial intelligence affect corporate ESG performance? (Kunzhe Yuan)
  Staggered difference-in-differences (DiD) on Chinese A-share firms links AI pilot zones to modest environmental, social, and governance (ESG) score gains, consistent with resource reallocation and green innovation.

• New · descriptive China's diffusion-forward AI strategy: The “AI race” in political economic context (Hao Chen, Meg Rithmire)
  Documents a policy model prioritizing embodied AI in manufacturing with investor-state financing and decentralized-hierarchical governance.

What Moved

• Concentration dynamics in recommendations: Against the long-running worry that recommenders amplify superstars, the 8.56M-user Netflix RCT shows diffusion toward the middle tail with higher engagement. Relative to earlier mixed correlational evidence, this is a clean causal push toward lower concentration when quality improves.

• Intermediary traffic under AI-answer search: A preregistered RCT quantifies a large drop in outbound clicks and small UX declines with AI answers, strengthening the case that on-platform synthesis reallocates rents away from publishers without clear user-perceived gains in this setting. This contrasts with platform claims of improved satisfaction and pushes the debate toward quantifying publisher compensation mechanisms.

• Task reallocation in finance analytics: Quasi-experimental evidence from US 10-K cycles and bank AI investments extends augmentation findings to sell-side analysts, with faster, bolder, more accurate forecasts and expanded coverage. This shifts weight toward AI complementing expert judgment rather than replacing it in information-intensive roles.

• Equity via name inference: Experimental and observational results show a legibility gap that advantages Western-signaling names, challenging the idea that name-based gender interventions benefit women uniformly. This narrows the scope of evidence for name-inference benefits and spotlights cross-cultural blind spots in citation and recognition tools.

• Methods and measurement: A general caution on event-time designs argues self-timed AI events can mechanically produce post-event lifts, implying some past click- or engagement-based evaluations may overstate treatment effects. Editors' inference: expect re-estimation and design changes in upcoming digital-product studies.

Contested & Watch

• Do platform recommenders amplify or diffuse concentration? - Finding: The Netflix RCT (N=8.56M users, global) shows improved recommendations lower title HHI (Herfindahl-Hirschman Index, a concentration measure) by ~5.7 percent, shifting share from superstars to the middle tail. - Standing evidence: Several observational and smaller-scale studies, mixed identification, split on superstar amplification versus diversification. - Watch: Additional large-scale holdbacks across music, social, and news platforms that report concentration metrics and revenue impacts by tail segment.

• Are AI-answers in search net beneficial for users and publishers? - Finding: A preregistered RCT (N=1,100) shows -18.8 percentage points in external CTR, fewer sessions, and lower trust with AI mode. - Standing evidence: Platform-run reports and lab studies suggest faster task completion and higher satisfaction, mostly correlational or non-preregistered. - Watch: Longer-horizon field experiments with revenue-linked outcomes for publishers and calibrated user-utility measures beyond clicks.

• Does AI augment or displace high-skill analysts? - Finding: Quasi-experimental evidence links bank AI investment and iXBRL to timelier, bolder, more accurate US equity forecasts and expanded coverage. - Standing evidence: Prior RCTs show augmentation in customer support and coding; displacement evidence is mixed and sector-specific. - Watch: Microdata on time allocation, compensation, and exit within analyst teams before/after AI rollout, ideally with instrumented adoption timing.

• Can name-based equity interventions deliver across cultures? - Finding: Experiments show interventions disproportionately lift Western-legible women's names, with muted gains for culturally ambiguous names. - Standing evidence: A handful of audits and field tests, mostly suggestive, note classifier bias but lack outcome-level redistribution estimates. - Watch: Field deployments that report outcome changes by name-legibility strata and alternatives using self-identification rather than inference.

• Do self-evolving agents net raise utility without raising risk? - Finding: In controlled audits, some self-evolution methods increase task accuracy but also coincide with higher prompt-injection exposure and unauthorized state changes. - Standing evidence: Multiple benchmark studies document stochastic failures and audit blind spots, with sparse production evidence on evolution toggles. - Watch: Production A/Bs that jointly track task utility and incident rates before/after enabling self-evolution, with reproducible attack suites.

Methods Spotlight

• Massive production holdback RCT (Netflix paper): An 8.56M-user, 60-day randomized holdback on a live platform provides rare, economy-relevant causal estimates of how recommender upgrades change engagement and concentration.

• Repeated, state-diff grounded audits (One-shot audits paper): By replaying and diffing world states across runs, the method reveals stochastic, irreversible agent damage that single-run audits systematically miss.

• Longitudinal upgrade durability benchmarking (UpgradeBench): Evaluating specialist adapters across actual base-model release sequences, not one-off hops, informs real upgrade-vs-retrain trade-offs practitioners face.