Can Regulation Curb AI Adoption Narrative–Practice Decoupling?An Agent-Based Model of the A-Share Market
Using an agent-based model of the A-share market, this paper demonstrates that while market forces alone weakly penalize the gap between firms' AI narratives and actual implementation, a regulatory combination of repeat-offender escalation and targeted enforcement is most effective at curbing such decoupling, whereas mechanisms like short selling and simple disclosure labels prove ineffective.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of finance, companies constantly tell stories to attract investors. They speak of future plans, new technologies, and strategic shifts, hoping these narratives will convince people to buy their stock. For decades, economists have understood that there is often a gap between what a company says and what it actually does. When a company claims to be transforming its business but lacks the budget or the staff to back it up, it creates a disconnect. This is not necessarily a lie, but it is a mismatch between the story told to the public and the reality of the business operations. The question that has long puzzled regulators and market watchers is whether the market itself can spot this gap and punish the companies that are merely talking, or if investors are easily swept up by the excitement of the story regardless of the facts.
A new study tackles this problem by looking at the Chinese stock market, specifically the A-share market, where artificial intelligence has become a dominant topic of conversation. The researchers wanted to know if the current rules of the game are enough to stop companies from riding the wave of hype without doing the hard work of implementation. To find out, they did not just look at past stock prices; they built a digital laboratory. They created a computer simulation that mimics the behavior of thousands of real companies, millions of individual investors, and a regulatory body. By feeding this virtual market with real data from over 3,000 listed firms, they could watch how different rules would play out when a wave of excitement about artificial intelligence swept through the system.
The researchers started by measuring two distinct things for every company in their simulation. First, they measured the "talking," which is the intensity of the artificial intelligence narrative found in a company's annual reports. This is the language used to describe AI ambitions. Second, they measured the "doing," which represents the actual, observable work being done. This was tracked by counting how many companies were hiring people with AI skills and how many were filing patents related to the technology. When a company talked a lot about AI but showed little evidence of hiring or patenting, the simulation flagged this as a decoupling, a separation between the story and the substance.
The core of the experiment involved injecting a sudden, massive wave of excitement about artificial intelligence into this virtual market, similar to the real-world surge that happened when new AI tools became widely known. The researchers then ran the simulation over and over again, changing the rules of the game each time to see which regulatory approach worked best. They tested nine different methods that regulators use around the world. Some methods were like a blanket rule, punishing any company that showed a gap between talking and doing, regardless of the size of the gap. Others were more like a targeted spotlight, focusing only on companies that were clearly exaggerating. Some relied on whistleblowers to report bad actors, while others simply asked companies to explain their actions if they didn't comply with a standard.
The results revealed that the market, left to its own devices, is surprisingly weak at punishing these exaggerations. Even without any new rules, the simulation showed that investors were not effectively distinguishing between the companies that were truly building AI and those that were just talking about it. The gap between the story and the reality did not lead to a significant drop in stock prices on its own. This suggests that relying on the natural discipline of the market is not enough to stop the hype.
When the researchers introduced different regulatory tools, the outcomes varied dramatically. The most effective approach was a system of escalating penalties for repeat offenders. In this scenario, if a company was caught exaggerating its AI efforts, the penalty for the next offense would be significantly higher than the first. This created a powerful incentive for companies to stop talking and start doing, or at least to stop lying about it. The simulation showed that this method was far more successful than simply making the initial penalty larger. It was the structure of the rule, not just the size of the fine, that drove the change.
Another highly effective method was targeted enforcement. Instead of punishing every company that had a gap between talking and doing, regulators in this scenario focused only on those with a positive gap—companies that were talking more than they were doing. This approach avoided punishing companies that were actually doing more than they were talking about, which kept the incentives for genuine innovation intact. When these two methods were combined, they produced the strongest results, effectively curbing the narrative decoupling and aligning stock prices more closely with the actual work being done.
In contrast, several popular regulatory ideas proved to be ineffective in this specific environment. Simply asking companies to disclose their AI plans or to explain why they weren't complying with a standard did little to change behavior. The simulation showed that these "comply or explain" rules, which are common in many markets, failed to stop the hype because the cost of talking was too low and the risk of being caught was too small. Similarly, allowing investors to bet against companies by short-selling their stocks had almost no effect. This was largely because the rules of the Chinese market make it very difficult for investors to take these short positions, meaning the mechanism simply could not function as a check on bad behavior.
The study also found that the design of the regulation matters more than the sheer size of the punishment. A system that penalizes companies based on a fixed threshold, regardless of whether they are genuinely innovating or just bragging, can actually hurt the market. It can discourage real companies from investing because they fear being punished for having a gap that is actually due to the slow pace of real-world implementation. The most successful rules were those that were smart enough to distinguish between a company that is genuinely trying and one that is simply lying.
Ultimately, the research suggests that to stop the disconnect between AI narratives and reality, regulators need to move away from broad, one-size-fits-all rules. Instead, they should adopt a system that gets stricter the more a company is caught exaggerating, and that focuses its scrutiny specifically on those who are talking more than they are doing. The study indicates that a combination of these two approaches offers the best path forward. It is a finding that relies on a computer model of the market, not a direct observation of the future, but it provides a clear map of how different rules might shape the behavior of companies in the age of artificial intelligence. The simulation shows that without the right kind of pressure, the market will continue to reward the most articulate storytellers rather than the most substantive builders.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.