Corporate Disclosure in the Age of AI
December 15, 2025
Summary
Corporate disclosure is becoming far harder to manage in the age of AI. While companies once focused on safeguarding traditional releases—financial reports, earnings calls, and sensitive R&D details—AI now enables anyone to infer strategic insights from seemingly innocuous information. Large language models can already match expert analysts in identifying peer firms, detect “rule-bending” cultures from accountant job postings, and extract signals about innovation strategy from skill-specific hiring ads. As AI improves, every public communication becomes a potential source of competitive intelligence. Yet relying on AI to craft disclosures can backfire, dulling meaningful signals for sophisticated investors and enabling opportunistic actors to appear credible.
Traditionally concerns around corporate disclosure have been restricted to a select set of releases—financial reports filed with regulators such as the SEC, earnings calls, etc.—that were considered material for financial performance.
Managers carefully curate this information and keep sensitive information under wraps, taking decisive steps to preserve secrecy: curbing loose lips, shoring up cracks in corporate firewalls, and carefully limiting access to competitively sensitive information about, for example, R&D and proprietary technologies. This has always presented a challenge, because a firm’s information environment can be complex and the flow of information has been known to elude managerial control.
In 2023, for example, online researchers happened upon a 38-terabyte cache of sensitive data unintentionally linked to one of Microsoft’s GitHub accounts—including over 30,000 internal MS Teams messages. The culprit was an access token that had been misconfigured to allow permissions for the contents of an entire Azure Storage account rather than, as intended, open-source AI models held in that account. Accidental data leaks like these are especially significant when they vouchsafe behind-the-scenes information that, even if incomplete, could be highly revealing in the hands of professional market analysts. Apple’s electric car project, for example, was leaked by its job posting. In another similar slip, the company also accidentally leaked its own top-secret hardware in software code.
The challenge is increasingly being magnified by AI, which is rapidly democratizing market expertise. To take one example. Investment managers select stocks through analyzing companies’ performance and prospects relative to a carefully researched peer group of competitors. Our recent paper in Review of Accounting Studies, which looked at large publicly traded companies about which data abounds, found that Bard (now Gemini), Google’s chatbot, was able to generate lists of peer firms—a key component of valuation—that were comparable in similarity and accuracy to those prepared by expert human investment analysts. This means that almost anyone with an AI can almost instantly do the detective work on a large company’s competitiveness that once required hours of expert analysis.
The situation was less depressing when it comes to midsize firms, where we found that informed humans outperformed Bard, largely because AI still needs a tremendous amount of relevant and publicly available data in its pre-training process in order to do work on par with flesh-and-blood financial whizzes. Unfortunately, given the pace of improvement in AI’s technology, any remaining human advantage is unlikely to last long.
In this article we argue that that the problem is rooted in the fact that AI can draw inferences from information that has not traditionally been seen as competitively sensitive. To counter this challenge, companies will need to take a more nuanced and holistic approach to managing the flow of information disclosed.
The Data Inside the Data
To be sure, company reports are starting to respond to the AI challenge in disclosures: recent research shows that the structure and tone of company annual reports have been changing as issuers begin to eliminate words known to raise red flags for AI algorithms—for example, “correction” and “restructuring.” But reactive and cosmetic moves such as these won’t be enough to keep firms from being easily outmaneuvered by the rapidly advancing technology. What’s more, they remain within the realm of corporate disclosure as it has been traditionally defined, overlooking the diversity of data that the algorithm can learn from.
The trouble is that the ability to find meaningful patterns within vast data-sets that could be overlooked by traditional methods, means that large language models (LLMs) are increasingly able to surface hidden insights at a scale and speed most companies are not yet prepared for.
One of us (Yi) co-authored a working paper using Google’s BERT LLM to identify “rule-bending” and “rule-following” language in job postings for corporate accountants. For example, a “rule-bending” job ad would invite applications from candidates willing to “explore alternative solutions” or “think outside the box,” while “rule-following” ads would stress “conformance,” “compliance,” “accurate reporting,” etc. These linguistic choices telegraphed actual organizational intent: “Rule-bending” companies were more likely to engage in earnings manipulation in subsequent months and years, based on their 10-K filings. As a result, these companies would be riskier investments.
That wasn’t the only informational nugget to be mined from job postings: Another paper Yi co-authored (published in Contemporary Accounting Research) found that innovative firms facing heavy competition included much more skill-specific information in their job ads, meaning a careful observer could glean valuable hints about their future strategy simply by perusing the ads they were using to recruit specialized talent.
What Should Firms Do?
Given a world in which every innocuous release could be algorithmically mined for strategic insights, leaders might understandably be reluctant to divulge to AI anything it doesn’t already know. Perhaps paradoxically, that could mean in the future that companies actually start to rely more on LLMs, as opposed to humans, in preparing information packages—such as annual reports—for the public. We have seen a similar dynamic at play in education and the labor market, where AI evaluates information prepared with heavy reliance on AI, notably job-seeker résumés.
But this isn’t necessarily a good solution. In the capital markets context, having AI at both ends of the information pipeline may hinder firms from putting their best foot forward. In a 2023 co-authored working paper, Yi analysed more than 100,000 earnings-call transcripts, comparing CEOs’ answers during the Q&A to answers given by ChatGPT to the same analyst questions. As it turned out, the variance between the two answers predicted changes in market outcomes (abnormal trading, for example, or abnormal returns) around the call. When the CEO’s answer varied from the ChatGPT answer, prices moved. When there wasn’t much difference and the CEOs talked like the chatbot, very little market-moving information seemed to get conveyed. If you want to impress the expert investors, therefore, don’t rely on AI.
Non-expert investors should also be wary of companies that rely on AI to attract funding. In another working paper, Yi and his co-authors used AI detection tools to identify crowdfunding proposals on Kickstarter written with the help of LLMs. These were not only more apt to be funded than human-authored proposals, but attracted larger sums as well (interestingly, most of these benefits accrued to start-ups from non-English-speaking countries). But the study also found that content creators using AI as a writing assistant were less likely to deliver their promised projects on time, suggesting that AI-assisted disclosure can enable shady or incompetent actors to claim unearned legitimacy.
In the AI panopticon, no detail is seen in isolation. Technological developments in recent years—culminating in, but not limited to, AI—have given external stakeholders unprecedented power to see through the fog of organizational obfuscation. Job postings shed light on corporate strategy; a tossed-off comment in an earnings call is read like runes by edge-seeking observers. It’s no longer possible to tell who might ultimately subject your disclosures to algorithmic unpacking, which increases the risks both for companies and their potential investors.
In response, firms will need to twice-groom all their disclosures and communications: first, for their immediate audience and the data analytics tools they’ll be using and then for the other relevant stakeholders relying on their AI partners. Maintaining a coherent communication strategy in this context will require processes for quick, collaborative vetting among multiple departments and fluency in a wide range of communication tools.
About the Authors
Yi Cao is an Assistant Professor of Accounting at George Mason University. His research explores how firms communicate with investors and how labor dynamics shape financial reporting quality, drawing on large language models and other machine-learning tools to analyze complex disclosure environments.
Long Chen is Accounting Area Chair and Associate Professor of Accounting at the Costello College of Business, George Mason University. She conducts empirical archival research on topics related to financial reporting and disclosure in both US and international settings, corporate social responsibility, executive profiles, and artificial intelligence.
[Harvard Business Review]
