A subtle, yet potentially dangerous shift is underway in one of the most influential narrative sections of financial reports: the Management’s Discussion and Analysis (MD&A). Companies are increasingly crafting these disclosures not just for human shareholders or regulators, but for machines.

This shift isn’t theoretical. In a recent article, I argued that artificial intelligence is no longer just analyzing MD&A sections after publication; it’s shaping how they’re written in the first place. I identified three key pressures: exposure (avoiding AI-driven sentiment red flags), competition (benchmarking language against peers), and reputation (anticipating how an algorithm might interpret a phrase). While that initial work theorized the concept of AI-induced disclosure pressure and examined the correlation between narrative tone and financial performance, it lacked a quantifiable measure for how disclosures were specifically optimized for machine readers or for detecting subtle forms of algorithmic manipulation.

Since then, many have asked: Can we actually measure this shift? And more importantly: Can regulators detect when narrative tone crosses the line from optimism to manipulation?

The answer is yes. And now we have the tools to do it.

A Simple, Testable Framework – Prompted and Scored by AI

In a new working paper, I present two empirical tools designed to measure how firms tailor their MD&A disclosures to algorithmic readers. The analysis draws on 80 MD&A sections from 20 large S&P 100 firms between 2021 and 2024. Both tools were implemented using two leading generative AI models: ChatGPT and Gemini, based on carefully calibrated prompts developed for each platform. Each model received the same strict scoring rubric and instructions, applied to the full text of the MD&A without manual editing or cleaning.

The first tool, the AI Orientation Score, ranges from 0 to 5 and assesses how machine-optimized a disclosure is based on measurable language, keyword usage, structured formatting, impersonal tone, and tonal consistency. A higher score suggests that the MD&A was likely crafted with machine readers in mind.

The second tool, the AI Manipulation Exposure Index (AI-MEX), also scored from 0 to 5, captures rhetorical red flags that may signal tone manipulation. These include upbeat language despite poor financials, absence of key performance indicators, vague aspirational claims, and excessive repetition of positive terms.

Importantly, these tools are transparent, replicable, and simple enough to apply using public AI models and publicly available filings, making them accessible to regulators, analysts, and journalists alike.

What the Results Show – and Why It Matters

The results are clear. Between 2021 and 2024:

• AI Orientation Scores rose steadily, indicating growing use of algorithm-friendly narrative techniques in MD&A disclosures.
• AI-MEX scores were significantly higher for firms with weak fundamentals, such as negative net income or deteriorating cash flow.
• By 2024, over 60 percent of companies scored high on both indices, signaling a dual strategy of structural optimization and tone management.

While some variation exists between the two AI models, the average scores were highly correlated across the sample. In most cases, ChatGPT and Gemini arrived at nearly identical conclusions. Where discrepancies did arise, typically in cases of high rhetorical ambiguity, they served as useful flags. These disagreements are not noise; they are diagnostic cues for human reviewers to take a closer look. Rather than dismiss these differences, we treat them as valuable signals, much like volatility in analyst forecasts. High consistency confirms robustness; disagreement points directly to areas where rhetoric may be outpacing substance.

A Crucial Complement, Not a Casual Commentary

The MD&A has long played a critical role in corporate reporting. It serves as the narrative complement to the audited financial statements, a place where management explains results, articulates risks, and outlines forward-looking plans. In many cases, it’s the only section where strategy, performance, and uncertainty are discussed together in plain language.

But that value depends on trust. If the MD&A becomes a stage for algorithmic simulation, structured for machine approval but detached from financial reality, its usefulness is at risk. Investors, analysts, and regulators may begin to discount it, weakening a key layer of market transparency.

That’s why these tools matter.

AI-MEX offer a way to protect the integrity of the MD&A by identifying when the narrative form no longer matches economic substance.

What Regulators Can Do  Now

This isn’t a call for new legislation; it’s a call for smarter oversight using tools already available.

• Incorporate AI Orientation and AI-MEX as diagnostic tools when reviewing MD&A disclosures, especially from firms under performance pressure.
• Treat significant disagreements between AI models as a signal, not a flaw. They point to exactly the kind of interpretive ambiguity that deserves closer attention.
• Elevate the MD&A’s regulatory status – not by restricting narrative content, but by holding firms accountable for how tone, metrics, and structure align.
• Encourage the use of concrete, comparable key performance indicators in forward-looking statements, rather than allowing vague optimism to dominate.

Restore Confidence in the Narrative

The MD&A was designed to illuminate, not obscure, the firm’s performance and outlook. In today’s algorithm-driven environment, the challenge is no longer just disclosure, it’s truthful, balanced communication for both human and machine readers.

Back in 2008, Professor Feng Li demonstrated how managers at underperforming firms often resorted to complex, hard-to-read MD&A language to obscure weak fundamentals. Today, complexity has given way to clarity, but it’s clarity crafted to satisfy algorithms, not necessarily to inform stakeholders.

We now have practical tools to detect when that balance is lost. In most cases, AI models converge; in ambiguous cases, their disagreement is itself a red flag. Either way, it’s no longer a guessing game.

Regulators don’t need new legislation, just the resolve to use what’s already available. If they fail to act, the MD&A risks becoming a polished narrative shell: credible in form, but hollow in substance.

This post comes to us from Professor Keren Bar-Hava at Hebrew University of Jerusalem. It is based on her recent paper, “The AI-MEX Framework: Two Empirical Tools for Detecting Algorithm-Oriented Narrative Design and Manipulation in MD&A Disclosures,” available here.