Reckoning with the Political Economy of AI: Avoiding Decoys in Pursuit of Accountability
Source: https://arxiv.org/abs/2604.16106v1 ↗
Full text: arXiv preprint ↗
Vertesi, boyd, Taylor, and Shestakofsky argue that AI accountability debates are not failing by accident — they are being shaped by the same networks of power they nominally critique.
The concept of 'decoys' is analytically sharp: it names a mechanism by which critics, journalists, and policymakers are recruited into legitimising the very structures they think they are challenging.
This is a political economy argument in the tradition of Galbraith or Veblen applied to contemporary AI governance — the question is not whether AI is good or bad, but who controls the terms of evaluation.
For product directors who work inside or alongside large AI platforms, this paper offers a structural map of why so many accountability efforts feel productive while changing little.
The author lineup — Janet Vertesi, danah boyd — brings serious institutional credibility to what might otherwise read as polemic.
Central argument
Vertesi, boyd, Taylor, and Shestakofsky argue that the 'Project of AI' is fundamentally a world-building endeavor by networked elites who consolidate wealth and power by enrolling critics, researchers, policymakers, and the public into reinforcing that project through what the authors call 'decoys' — framings that appear to challenge AI (debates about bias, fairness, capability, governance) but actually stabilize and co-constitute AI's political economy. Drawing on Castells' theory of network-making power and STS concepts of material political economy, they contend that accountability efforts consistently fail because they target technical symptoms rather than the underlying infrastructure of capital, relationships, and institutional arrangements that make AI possible. The paper identifies five such decoys and argues that genuine accountability requires shifting analytic attention to those material political-economic networks — the financiers, alliances, and physical infrastructures — rather than the technical particulars AI companies prefer critics to debate.
Critique
The paper's central concept of 'decoys' carries an inherent epistemic risk: by framing virtually any existing accountability effort (bias audits, fairness research, regulatory proposals) as potentially complicit in the Project of AI, the argument becomes difficult to falsify and risks dismissing productive incremental work without a clear threshold for what engagement would not constitute co-optation. The authors gesture toward 'grappling directly with the material political economy' as the alternative, but the paper — at least in these opening sections — does not specify what forms of critique or intervention would be sufficiently outside the decoy logic, leaving practitioners and policymakers with a diagnosis but no actionable traction. This creates a tension between the paper's call for accountability and its structural tendency to delegitimize the most accessible levers people currently have.
Why it matters for product
For a CPO, the paper's core move is directly applicable to product strategy: just as scholars get captured by debating AI's technical properties while missing the power structures underneath, product teams risk spending roadmap energy on surface-level AI ethics gestures — responsible-use policies, bias disclaimers, explainability features — that satisfy internal stakeholders without interrogating how their product's dependency on a handful of foundation model providers reproduces the very lock-in and extraction the paper describes. The concept of 'network-making power' also reframes vendor and partnership decisions: when a product integrates deeply with an AI platform, the organization is not just making a technical choice but enrolling itself into a specific configuration of dependencies and power relations that will constrain future strategic optionality. This argues for treating AI infrastructure choices as political economy decisions — scrutinizing who controls the pipelines, data, and pricing — rather than purely as build-vs-buy cost calculations.