Data as Liability

Fatih Kansoy; Yuhao Huo

Working Paper · Mar 2026 New

Data as Liability

with Yuhao Huo (University of Cambridge) · University of Oxford Department of Economics Discussion Papers

PDF

Summary

When firms move from inventing AI to actually deploying it, where does the hidden governance cost show up? In roughly 84,000 firm-years of US 10-K filings, this paper shows that the answer is in breach-risk disclosure: AI adoption, not AI invention, is the margin most strongly associated with higher stated attention to data-breach risk.

The effect is economically meaningful at about 5 per cent of the sample mean, survives a non-AI digitisation placebo, and is strongest when AI deployment is customer-facing. The central wedge is visible in the figure below: once adoption and invention enter together, adoption stays positive while invention becomes economically negligible.

Figure comparing the adoption coefficient ladder with the adoption-versus-invention wedge. — Headline figure: the adoption-versus-invention wedge. The paper's core result is not that AI discussion matters in general, but that adoption-specific language is the margin that tracks breach-risk attention.

At a Glance

Research question: Does AI create a disclosure-visible data liability, and is that effect driven by adoption rather than invention?
Core finding: AI adoption is linked to higher breach-risk attention; AI invention adds little once both margins are estimated together.
Data: About 84,000 US listed-firm annual filings, 1994-2023, with layered text and LLM classification.
JEL: O33, G38, D83, L86, K22

Citation

Kansoy, Fatih, and Yuhao Huo. 2026. “Data as Liability.” University of Oxford Department of Economics Discussion Papers.

@article{kansoy2026data,
  title   = {Data as Liability},
  author  = {Kansoy, Fatih and Huo, Yuhao},
  journal = {University of Oxford Department of Economics Discussion Papers},
  year    = {2026}
}