Capital Market Consequences of Generative AI, Early Evidence from the Ban of ChatGPT in Italy

What happens to valuations and liquidity when a regulator suddenly restricts access to generative AI?

Jeremy Bertomeu, Yupeng Lin, Yibin Liu, and Zhenghui Ni ask how markets price generative AI in their paper « Capital Market Consequences of Generative AI: Early Evidence from the Ban of ChatGPT in Italy ».

They exploit Italy’s March 31, 2023 ChatGPT ban as a natural experiment, using matched European firms and difference-in-differences designs on stock returns, analyst forecasts, and bid-ask spreads.

Their key takeaways include:

• During the ban, Italian firms with higher GenAI exposure underperform by about 8.7% relative to matched European stocks, and about 9.2% when adding industry fixed effects.
• A high-exposure minus low-exposure strategy yields about -1.5% over a three-day window and about -5.48% over the 20 trading days spanning the ban period.
• Firms with higher investor attention to ChatGPT show about -6.8% differential return during the ban, and firms with higher AI-topic mentions show about -8.7%.
• The ban’s return impact is more negative for younger firms (about -6.8%) and for smaller firms by market capitalisation (about -7.1%), suggesting access constraints hit challengers harder.
• Analysts in Italy cut forecast production relative to foreign peers covering the same firm: forecast likelihood fell by 10.9%, and forecast frequency by 29.0% during the ban.
• Bid-ask spreads increase during the ban, and the widening is smaller for firms with more analyst coverage, higher foreign ownership, and higher institutional ownership.

This research suggests regulatory constraints on data use and tool access can transmit into equity returns and market liquidity, with distributional effects across firm size and age.

Generative AI access can be a priced productivity input, not just a technology narrative through data privacy and governance mechanisms.

As a limitation, the ban can affect both firms’ fundamentals and information processing costs, and analysts may circumvent the ban (for example via VPN), so analyst effects may be a lower bound.