Corporate Communication with AI-Equipped Investors

How to cite

Ni, Zhenghui; Lin, Yupeng; Su, Yichen; Xu, Nianhang. 2025. "Corporate Communication with AI-Equipped Investors." AEA RCT Registry. AEARCTR-0016695. https://doi.org/10.1257/rct.16695-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Zhenghui Ni* , Yupeng Lin , Yichen Su , Nianhang Xu

* Primary investigator

Registry ID
AEARCTR-0016695
DOI
https://doi.org/10.1257/rct.16695-1.0
First registered
Last updated
Status
On going
Trial dates
2025-04-01 – 2025-10-31
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/16695

Abstract

The ability to acquire and process information is central to capital market efficiency (Grossman and Stiglitz 1980; Sims 2003). Advances in artificial intelligence (AI) have the potential to reduce investor-side frictions in information processing, but their impact on firms’ disclosure strategies remains unclear. This paper implements a large-scale randomized controlled trial (RCT) with 4,049 publicly listed firms in China to study how managers respond to perceived AI-equipped investors. Our experiment leverages the institutional setting of online earnings conferences (OECs), regulator-endorsed text-based platforms where investors pose real-time questions to managers. We exogenously vary only the perceived identity of questioners—framed either as conventional human investors or as AI agents—while holding informational content constant. This design addresses two central challenges in the literature: the dual unobservability of investor AI adoption and managerial beliefs about AI readership, and the difficulty of isolating exogenous variation in information demand. The setting also allows us to separate market reactions to questions themselves from reactions to managerial responses, opening a microstructure perspective rarely feasible outside an RCT. By examining both voluntary disclosures during OECs and subsequent mandatory reports, the experiment provides causal evidence on whether, how, and under what conditions generative AI monitoring disciplines managerial disclosure behavior and alters the informativeness of capital markets.

Experimental design

This study examines how firms respond to perceived differences in investor identity during online earnings conferences (OECs) in China. OECs are regulator-endorsed, text-based events where investors submit questions to management in real time. Unlike traditional earnings calls, OECs are public, standardized, and widely accessible, providing a unique opportunity to experimentally vary investor inquiry styles while observing both managerial responses and contemporaneous market reactions.

The experimental sample consists of OECs held between April 14 and June 6, 2025, a period covering mandatory earnings conferences following annual report disclosures. We begin from all conferences announced on the official disclosure portal (cninf) and apply a series of pre-registered exclusions (e.g., offline calls, joint conferences, non-mainstream platforms, and Q1-only events). After exclusions, the final sample includes 2,734 OECs across nine mainstream platforms.

Each participating firm receives four standardized investor questions, covering:

Current financial performance,
Future growth outlook,
Current industry performance and peer comparison,
Future industry trends.

The control group (T0) receives these questions in conventional human phrasing, typical of actual investor inquiries on OEC platforms.

The treatment group (T1/T2) receives the same questions, but presented in an AI-like style: phrased with robotic linguistic features and prefixed with tags (e.g., “[inquiry-001]”). Where platforms allow, we further reinforce this identity with chatbot-style profile images. Thus, treatment varies only in the perceived identity of the questioner, while holding informational content constant.

Design details

Randomization method
Randomization is conducted using a coin-toss procedure in office by a computer. Each conference is independently assigned with 50% probability to the treatment or control group. Because conference schedules are announced gradually and unevenly across dates, randomization occurs on a rolling basis: for each daily batch of newly announced conferences, assignment is determined the day before the scheduled OEC. This ensures independence across conferences and avoids ex ante sample imbalance.
Randomization unit
The unit of randomization is the online earnings conference (OEC), which corresponds to a single firm’s disclosure event. Each OEC is independently randomized into treatment or control using a simple coin toss on the day prior to the scheduled conference.
Clusters
Approximately 2,700 online earnings conferences (OECs) (one OEC per firm) during Apr 14–Jun 6, 2025. (Unit: OEC/firm)
Observations
Approximately 2,700 firm–OEC observations (primary analysis unit). Secondary analyses at the question level: ≈ 10,800 question–OEC observations (4 standardized questions per OEC).
Arms
Two-arm design at the OEC (firm) level: Control (human style): ~1,350 OECs Treatment (AI-styled): ~1,350 OECs
Primary outcomes
the primary endpoints are: (i) probability of manager response, (ii) tone/neutrality of voluntary disclosure, (iii) stock price and trading volume reactions at high frequency, (iv) shifts in mandatory disclosure tone.

Cited by

No citation data available for this registration yet.

Keywords