Companies have discovered how to game the system in the age of AI-powered investment analysis—and the techniques they're using reveal fundamental vulnerabilities in how artificial intelligence interprets financial disclosures.
Bradford Levy, a University of Chicago professor researching AI applications in financial markets, has documented how corporations are strategically restructuring SEC filings to exploit weaknesses in language models. The tactics don't involve changing information—that would be securities fraud—but rather manipulating how AI tools process and prioritize that information.
The core exploit is elegantly simple: AI models are less likely to scrutinize content in the middle of documents. So companies shuffle sections, moving unfavorable disclosures—say, details about declining margins in a key business segment—from prominent positions to the middle of lengthy 10-K filings where AI parsers are statistically less likely to weight them heavily in analysis.
In Levy's classroom, students demonstrate the technique by "shuffling paragraphs around so they can move section seven to the front of the 10-K" without altering actual content. The reshuffling changes how AI-driven investment tools like AlphaSense, Bloomberg Intelligence, or ChatGPT-powered analysts interpret the document's overall sentiment and risk profile.
This represents a fascinating arms race. On one side, investors have gained unprecedented analytical power from AI tools that can process thousands of filings instantly, extract key metrics, and identify patterns humans would miss. Hedge funds and institutional investors increasingly rely on these tools for initial screening and analysis. On the other side, corporate IR departments and outside counsel have figured out how to downplay information they'd prefer received less attention—not by hiding it, which regulators prohibit, but by exploiting how AI attention mechanisms work.
The SEC hasn't issued formal guidance on this practice yet, but the cat-and-mouse game raises profound questions about disclosure effectiveness in the AI era. The spirit of securities law requires material information be presented clearly and prominently. But if companies can technically comply while knowing AI tools will systematically underweight certain disclosures, does that satisfy the disclosure obligation?
Levy's students now include a counter-exercise where some attempt to "unwind the positive bias that other students attempt to put in the disclosure so that companies cannot game the situation." It's arms race dynamics playing out in real-time: as AI detection methods improve, gaming techniques will evolve in response.
The practical implications for investors are significant. Those relying heavily on AI-powered screening and analysis tools should recognize they're potentially missing material information that companies have deliberately de-emphasized. Human oversight remains essential—not just because of "the dangers of hallucinations" that Levy notes, but because adversarial actors are actively trying to manipulate the AI's understanding.
For AI vendors providing financial analysis tools, this exposes a critical vulnerability in their products. AlphaSense, Bloomberg, and others will need to develop more sophisticated document analysis that doesn't fall prey to simple structural manipulation. That likely means weighting sections based on regulatory requirements rather than document position, applying adversarial training to detect gaming attempts, and incorporating human expert review of AI-flagged anomalies.
The broader lesson extends beyond finance. As AI systems become gatekeepers of information—whether analyzing corporate filings, medical records, legal documents, or academic research—adversarial actors will develop techniques to manipulate how those systems interpret data. The examples will be technical and domain-specific, but the pattern is universal: any automated system with known biases can be gamed.
Some might argue this is just companies exercising discretion about how to present information, no different than emphasizing positive metrics in earnings press releases while burying concerning data in footnotes. But there's a distinction: press releases are explicitly promotional documents, while SEC filings are regulatory disclosures meant to provide material information to investors. Gaming AI analysis of regulatory filings sits in murkier ethical and potentially legal territory.
For retail investors, the takeaway is straightforward: don't rely exclusively on AI summaries of financial filings, no matter how sophisticated the tool. The companies you're analyzing may be actively trying to manipulate those summaries. Read the actual filings, particularly the "Risk Factors" sections and MD&A, regardless of where they appear in the document structure.
The SEC will eventually need to address this. Possible solutions include standardizing filing structures to prevent gaming, requiring digital tagging of material information regardless of location, or explicitly prohibiting structural manipulation intended to confuse AI analysis. Until then, investors face an uncomfortable reality: the AI tools they're relying on have known vulnerabilities, and the companies they're analyzing know exactly how to exploit them.
The numbers don't lie, but executives sometimes do—and now they've learned to do it in ways that fool the machines.
