The Innovation Tax: Generative AI Adoption, Productivity Paradox, and Systemic Risk in the U.S. Banking Sector

This paper evaluates the causal impact of Generative Artificial Intelligence (GenAI) adoption on productivity and systemic risk in the U.S. banking sector. Using a novel dataset linking SEC 10-Q filings to Federal Reserve regulatory data for 809 financial institutions over 2018–2025, we employ two complementary identification strategies: Dynamic Spatial Durbin Models (DSDM) to capture network spillovers and Synthetic Difference-in-Differences (SDID) for causal inference using the November 2022 ChatGPT release as an exogenous shock. Our findings reveal a striking Productivity Paradox'': while DSDM estimates show that AI-adopting banks are high performers ($β> 0$), the causal SDID analysis documents a significant Implementation Tax’’ – adopting banks experience a 428-basis-point decline in ROE as they absorb GenAI integration costs. This tax falls disproportionately on smaller institutions, with bottom-quartile banks suffering a 517-basis-point ROE decline compared to 129 basis points for larger banks, suggesting that economies of scale provide significant advantages in AI implementation. Most critically, our DSDM analysis reveals significant positive spillovers ($θ= 0.161$ for ROA, $p < 0.01$; $θ= 0.679$ for ROE, $p < 0.05$), with spillovers among large banks reaching $θ= 3.13$ for ROE, indicating that the U.S. banking system is becoming ``algorithmically coupled.’’ This synchronization of AI-driven decision-making creates a new channel for systemic contagion: a technical failure in widely-adopted AI models could trigger correlated shocks across the entire financial network.

💡 Research Summary

This paper investigates the causal impact of Generative Artificial Intelligence (GenAI) adoption on productivity and systemic risk within the U.S. banking sector, using a novel panel that links SEC 10‑Q filings with Federal Reserve regulatory data for 809 banks over the period 2018‑2025. The authors employ two complementary econometric strategies. First, a Dynamic Spatial Durbin Model (DSDM) captures network spillovers by constructing spatial weights based on business‑model similarity and geographic proximity, allowing the effect of a bank’s AI adoption to influence its peers over time. Second, a Synthetic Difference‑in‑Differences (SDID) design exploits the exogenous shock of the November 2022 ChatGPT release, creating a synthetic control group from non‑adopting banks to isolate the causal effect of adoption.

The DSDM results show that AI‑adopting banks are high performers: the direct coefficients (β) on AI adoption are positive for both ROA and ROE, indicating that, conditional on other factors, adopters enjoy higher profitability. More importantly, the spillover coefficients (θ) are statistically significant—θ = 0.161 for ROA (p < 0.01) and θ = 0.679 for ROE (p < 0.05). For large banks, the spillover magnitude is even larger (θ ≈ 3.13 for ROE), suggesting that AI adoption creates “algorithmic coupling”: banks’ decision‑making processes become synchronized through shared AI architectures, vendor platforms, and common training data.

The SDID analysis uncovers a striking “Implementation Tax.” After the ChatGPT shock, AI‑adopting banks experience an average decline of 428 basis points in ROE relative to the synthetic control, despite the positive β in the DSDM. This tax is highly heterogeneous: bottom‑quartile banks (by assets) see a 517‑bp drop, whereas top‑quartile banks see only a 129‑bp decline. The authors interpret this as evidence of scale economies in AI implementation—larger institutions can spread fixed licensing, integration, and talent costs across a broader asset base, while smaller banks bear a proportionally larger burden.

The theoretical framework models adoption costs as C_it = c₀ + c₁·S_i^φ + c₂·Complexity_i, where the fixed component c₀ captures software licensing, c₁·S_i^φ reflects scale‑dependent costs, and Complexity_i captures organizational intricacy. The “Implementation Tax” is the short‑run profit drag from these costs. The paper also formalizes algorithmic coupling through an additional correlation term: Corr_AI_ij,t = Corr_base_ij,t + δ·D_AI_it·D_AI_jt·VendorOverlap_ij. Empirically, δ is positive, confirming that shared AI models increase the correlation of banks’ AI‑driven decisions beyond baseline macro‑economic co‑movements.

Policy implications are threefold. (1) Targeted subsidies or technical assistance for small banks could mitigate the disproportionate implementation tax and reduce concentration risk. (2) Regulators should encourage diversification of AI models and vendors to avoid “model monoculture,” thereby limiting the systemic propagation of a single technical failure. (3) Ongoing monitoring of AI adoption patterns and cross‑bank algorithmic dependencies should become part of macro‑prudential supervision.

The paper acknowledges limitations: AI adoption is measured as a binary indicator, obscuring intensity or sophistication differences; vendor and model specifics are inferred rather than directly observed; and the analysis focuses on short‑run profitability, leaving long‑run productivity gains unexamined. Future research avenues include constructing continuous adoption metrics, case studies of AI‑driven operational failures, and macro‑economic modeling of AI’s aggregate impact on the financial sector.

In sum, the study reveals a “productivity paradox” in banking: while AI adopters appear more productive in cross‑sectional snapshots, the causal effect of adoption is a short‑run drag on profitability, especially for smaller institutions. Moreover, the positive spillovers generate a new channel of systemic risk—algorithmic coupling—that warrants close regulatory attention as generative AI becomes entrenched in financial decision‑making.