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, Skill Acquisition leaned positive (195 papers); this week, a counter-signal appears. - Strengthened: engineering-level choices, not new base-model capability, appear to drive large gains in tested settings, as cache-aware prompt compression roughly halved application programming interface (API) spend in those tests and agent-ready websites nearly doubled autonomous task success in a controlled prototype. - Challenged: policy reliance on large language model (LLM) watermarks as forensic evidence, with representative schemes often failing meaning-preserving paraphrases and falling short of legal-readiness tests in lab evaluations. - Newly observed: model homogeneity among algorithmic funds, rather than automation per se, is associated with stronger capital outflows from emerging markets after U.S. monetary shocks, pointing to a systemic similarity channel.

What Moved & What Held

Coming in, the standing view was that most near-term value comes from workflow and system design rather than frontier capability; watermarking and provenance promises were fragile; human-AI teaming outcomes hinge on configuration and governance; labor effects are heterogeneous with no economy-wide employment shock yet; and capability benchmarks often overstate deployable reliability.

This week adds magnitude and mechanism: cache-aware compression maps where provider caches matter and reports near-50% median savings in tested settings, while agent-ready site design is linked to much higher end-to-end agent success in a controlled prototype; watermark schemes often fail basic forensic-readiness when paraphrased in lab tests; and a macro-finance study points to algorithm similarity as the amplifier of U.S. shocks into emerging-market portfolio flows. Fine-tuning on "innocent" data is associated with broad ideological shifts in tests, raising the salience of evaluation and governance around distributional outcomes. Still holds this week: no aggregate labor-market flip, capability benchmarks outpacing deployment reliability, and the centrality of organizational design for realized gains.

Top Papers

• Extends · descriptive Cache-aware prompt compression: A two-tier cost model for LLM API caching - Yan Song (engineering evaluation, descriptive evidence) - Across LongBench-v2 configurations, a cache-aware compression strategy that preserves provider cache hits is the cheapest in 16/16 settings, yielding roughly 49% median API cost savings (and higher in some cases) without observed task-quality degradation in those tests; it refines the standing view by modeling a two-tier cache with sub-1.0 hit rates. - So what: If this generalizes, production teams risk materially overpaying for inference when prompt compression breaks cache semantics, and the budget variance grows with workload size. - Full numbers

• New · suggestive Algorithmic intermediation and the international transmission of U.S. monetary policy - Fernando Toledo, Luis Dimotta Bré, Gabriel Montes-Rojas (theory plus panel tests, suggestive evidence) - Using portfolio flows to 19 emerging markets (2000–2024), the paper’s two-region model and empirical tests suggest that similarity in fund trading models, not algorithmic trading per se, is associated with amplified outflows after U.S. monetary shocks; greater model heterogeneity correlates with more stable flows. - So what: If this holds, regulators and asset owners face underappreciated systemic risk from correlated models, where drawdowns cluster even without changes in aggregate automation. - Full numbers

• Tension · descriptive AI watermark evidence fails forensic readiness: An empirical evaluation - Saifur Rahman Tamim, Amir Labib Khan (security evaluation, descriptive evidence) - In lab tests against meaning-preserving paraphrases, three representative watermarking schemes (KGW, Unigram, SynthID) have their signals largely removed and fall short of thresholds consistent with forensic or legal standards, sitting in tension with policy assumptions that embedded provenance alone suffices. - So what: In this sample, watermark signals are fragile; the open question is whether provenance-dependent compliance processes are exposed to the same evidentiary gaps. - Full numbers

Also Notable

• New · framework ANet Patu-1: The value of connection in the agent network - Mu Yuan, Jinke Song, Zhaomeng Zhou, Lan Zhang. Theory proposes a self-organizing consensus protocol where large heterogeneous crowds of weaker agents can overtake smaller groups of stronger models, which may shift design attention to network structure over single-model strength.

• Extends · suggestive Designing agent-ready websites for AI web agents - Said Elnaffar, Farzad Rashidi. In a 300-run controlled prototype, agent-ready design was associated with strict autonomous shopping success rising from 49.3% to 89.3% in that prototype, reinforcing the payoffs to interface design over model swaps.

• New · suggestive Innocuous-seeming data, latent ideology: Ideological generalisation in finetuned LLMs - Robert Graham, Edward Stevinson, Yariv Barsheshat. Narrow, benign fine-tuning on GPT-4.1 is associated with broad cross-domain ideological shifts and sometimes extreme outputs in tests, elevating governance risks from small data updates.

• Extends · suggestive TRAIL: A platform for configurable human–AI teaming experiments - M. Samadi, Pedro Martins De Bastos, Jaeyoon Choi, Spencer Jaquay, Seehee Park, Nia Nixon. Classroom teams (n~51) suggest persona-configurable AI teammates trade off contribution vs climate in this sample, identifying a configuration parameter for evaluation and oversight.

• Extends · suggestive Designing trustworthy human–AI teams: Adaptive explanations and collaborative decision-making - K. Sailaja, N. Mithili, B. Vijaya, P.R Bharathi, N. Muthulakshmi, M. Rohitha. Expertise-aware explanations are associated with better trust calibration and team decisions in a medical prototype, pointing to explainability as an operational control rather than a generic add-on.

• Extends · descriptive The prover is the judge: Verified security software from AI coding agents in Ada/SPARK - Tobias Philipp. Verifier-driven agents discharge 49,280 proof obligations and may accelerate secure code production, but residual defects still need human review, underscoring verification workload shifts.

• New · descriptive Early adoption of agentic coding tools by GitHub projects - Maliha Noushin Raida, Daqing Hou. From 25,264 agentic pull requests (PRs) across 2,361 repos, adoption is highly skewed and concentrated, with smaller teams using more agent PRs per developer and one-human oversight as the modal pattern.

• Extends · descriptive From forecasts to auditable reports: Evidence contracts for LLM-assisted housing-guarantee risk monitoring - Hyeongcheol Kim, Y. Hwang. On Korean jeonse data, a retrieval-and-verification pipeline appears to improve upper-tail risk detection and yields analyst-usable reports, suggesting governance-aware evaluation may have operational value.

• Extends · descriptive Beyond success rate: Cost-aware evaluation of offensive and defensive security agents - Paul Kassianik, Blaine Nelson, Yaron Singer. Cost-success tradeoffs reorder agent rankings: offensive tasks tend to scale with inference compute, while defensive gains appear to depend on disciplined tooling and telemetry navigation.

• New · descriptive Frontier AI performance across the business disciplines - Ajay Patel, Kartik Hosanagar, Ramayya Krishnan, Chris Callison-Burch, Karim Lakhani, Mitch Weiss. A case-grounded benchmark finds high LLM scores on analytical business questions, while subjective components expose measurement limits.

• Extends · suggestive The most exposed sector meets the shock: AI exposure and firm-level labor outcomes in Indian IT services - Mihir Khanna. A firm panel in India (13 firms) links higher AI exposure post-2022 to reduced hiring and higher productivity, consistent with task augmentation rather than broad displacement.

• Extends · suggestive Artificial intelligence, firm heterogeneity, and labor market adjustment in Italy - E. Ceesay, M. Jallow, Cosimo Magazzino, Alasana Gitteh, B. Bojang. 2005–2024 panel associations show pooled positives but firm-level heterogeneity, with some cases of employment declines alongside productivity gains.

• New · descriptive Platform choice, trust, and privacy in the consumer AI assistant market - Jennifer Zou. U.S. survey (n=1,999, June 2026) finds ChatGPT 58% and Gemini 25% primary share; trust tracks use, and many users prefer human-free interactions for sensitive tasks.

• New · descriptive AI trading: Evaluating large language models for technical market analysis - Geofrey Ntale. In simulated trading across four tasks, GPT-4 Turbo leads general LLMs while FinGPT narrows the gap; results are regime-sensitive with numerical errors present.

• New · framework Reversibility-aware staged delegation for enterprise agentic AI - Kwan Hong Tan. A real-options framing with simulations suggests delegation that scores action recoverability is associated with fewer severe incidents than naive confidence gating.

• Extends · suggestive Intelligentization and greenization policies empowering the low-carbon transition - Weibo Jin, Mengting Zhang, Shuangying Wang, Ruohan Jiang. Chinese A-share panel (2013–2023) links smart-city and low-carbon pilots to lower CO2 intensity via eco-finance and innovation coupling.

• New · framework Messy research, certification and the monetization of science - J. Fourie. A model argues falling "polish" costs from AI outpace verification, expanding uncertified submissions and increasing the pricing power of credible certifiers.

• Extends · descriptive Digital education beyond access in India - Kushagra Garg, Avantika. Review argues India’s digital-education rollout expands access but reproduces inequalities without targeted supports.

• Confirms · descriptive AI, employment, and the human capability pipeline - Ivan Silva. Synthesis through July 2026 finds no economy-wide employment shock yet, with concentrated relative declines among young workers in high-exposure roles.

What Moved

• Cost engineering over capability: Relative to the baseline that engineering choices matter, this week better measured the upside, with cache-aware compression reporting about half in API cost savings in 16/16 tested configs and agent-ready sites associated with strict agent success rising from roughly 49% to 89% in a controlled prototype. The editorial inference is that platform-specific primitives (cache semantics, page structure such as the Document Object Model) may now set realized ROI at least as much as marginal model upgrades.

• Forensic readiness of watermarking: The standing hope that embedded provenance could anchor legal assurance is further challenged by lab evidence showing near-complete removal via paraphrase for three representative schemes. This shifts the governance conversation toward layered evidence and away from single-tech fixes.

• Systemic risk via model similarity: Beyond prior speculation about correlated strategies, panel tests across 19 emerging markets (2000–2024) tie algorithm similarity, not automation per se, to amplified capital-flow responses after U.S. shocks. That extends macro-financial risk framing from "more algos" to "too-similar algos."

• Ideological drift from fine-tuning: Quasi-experimental results suggest small, narrow fine-tunes are associated with broad ideological generalization and extreme outputs in tests, which better measures a governance hazard that baseline syntheses flagged but did not quantify across domains.

Contested & Watch

• Watermarks as legal-grade provenance - Finding: In lab tests, meaning-preserving paraphrases reduce signals from KGW, Unigram, and SynthID watermarking schemes to near-zero detection. - Standing evidence: A small set of descriptive evaluations and security papers lean skeptical, while several policy proposals still assume watermark sufficiency. - Watch: Field audits on real-world corpora with adversarial paraphrase, error rates benchmarked to forensic standards, and any courtroom admissibility precedents.

• Many weak agents vs few strong models - Finding: Theory shows heterogeneous crowds can overtake stronger homogeneous groups via ANet Patu-1 consensus; networked LLM-agent experiments, however, appear to need explicit first-round randomization to realize gains. - Standing evidence: Mostly frameworks and small-scale lab experiments with mixed results; no established causal evidence at production scale. - Watch: Head-to-head trials in comparable tasks measuring payoff and latency as agent count and heterogeneity scale, plus ablations on protocol details.

• Algorithm similarity as a spillover amplifier - Finding: For 19 emerging markets over 2000–2024, similar fund models are associated with stronger outflows after U.S. monetary shocks, while diversity stabilizes flows. - Standing evidence: Several established studies document common-factor and passive-flow amplification; the "similar models" mechanism is newer and suggestive. - Watch: Fund-level disclosures or instruments that identify exogenous similarity shifts, and quasi-experiments around model deprecations or policy nudges to heterogeneity.

• Fine-tuning and ideological generalization - Finding: Fine-tuning GPT-4.1 on small, benign datasets is associated with broad cross-domain ideological shifts and occasional harmful extremes in this sample. - Standing evidence: Prior descriptive work notes safety drift post-fine-tune; rigorous cross-domain quantification is limited and mixed. - Watch: Pre-registered evaluations across providers with holdout domains and counterfactual prompts, plus audits tying shifts to data curation choices.

• Micro job impacts without macro shock - Finding: In Indian IT services (13 firms), higher AI exposure post-2022 links to reduced hiring and higher productivity. - Standing evidence: Multiple syntheses find no economy-wide employment shock to date, with heterogeneous firm- and cohort-level effects. - Watch: Larger firm panels with exposure measures, vacancy flows, and wage ladders, ideally with difference-in-differences designs around adoption timing.

Methods Spotlight

• Cache-aware prompt compression (CAPC): Cache-aware prompt compression: A two-tier cost model for LLM API caching. Models provider cache tiers explicitly and suggests that preserving cache semantics during compression is associated with lower costs at scale in tests, without observed quality loss.

• ANet Patu-1 self-organizing consensus: ANet Patu-1: The value of connection in the agent network. A constructive protocol that formalizes how heterogeneous agent networks can converge and scale value, reframing design choices for many-agent systems.

• TRAIL configurable teaming platform: TRAIL: A platform for configurable human–AI teaming experiments. Enables repeatable, longitudinal manipulations of AI teammate personas in real classrooms, linking configuration choices to trust, contribution, and over-reliance.