When AI writes risk: What gets lost in corporate reporting

GenAI is reshaping how companies write about risk. It shortens drafts, simplifies complex language, and makes long disclosures easier to process. On the surface, that sounds like progress.

But in a recent study coauthored with Niccolò Marcarini (Cattolica University of Milan), we found a more uncomfortable result. Analyzing more than 5 million sentences from U.S. companies' risk-factor disclosures between 2019 and 2024 — and using the launch of ChatGPT in November 2022 as a natural dividing line — we documented a 38% rise in AI-generated sentences after that date. And the more AI was involved in drafting these sections, the less investors seemed to learn from them.

The readability trap

Risk disclosures written with more AI involvement are, on average, shorter and easier to read. That sounds like an improvement — and on the surface, it is. For years, these sections have been criticized as bloated, repetitive and nearly impossible to parse. GenAI fixes that. The prose gets cleaner. The structure gets tighter.

But readability is not the same as informativeness.

The same disclosures that got easier to read also became more standardized, more repetitive over time, and less specific to the individual company. The AI, in smoothing the language, also flattened the details that actually matter: the firm-specific exposures, the emerging risks, the signals that tell a sophisticated reader something genuinely new about this company at this moment.

The market noticed. Filings with higher AI usage triggered lower trading volumes, weaker price reactions and slower price discovery in the minutes after release. Investors still reacted — but they appeared to learn less, and to learn it more slowly.

Two ways to read that market signal

The first explanation is friction. When language becomes more templated, it's harder to spot what's new. A disclosure that reads like every other disclosure offers less to extract, especially on busy filing days when dozens of 10-Ks drop at once and analyst attention is already spread thin.

The second is something economists call rational inattention. If investors start to expect that AI-assisted risk sections contain less specific information, they may simply stop devoting resources to reading them carefully. The market doesn't break down — it just learns to ignore a signal it has stopped trusting.

Either way, the implication is the same: more readable doesn't mean more useful.

This isn't just a problem for public filings

Our research focuses on SEC disclosures, but the dynamic shows up anywhere GenAI is used to draft risk documents — audit narratives, board papers, vendor assessments, internal control summaries. Risk work generates enormous volumes of text under tight deadlines, and AI genuinely helps with that. But risk assessment isn't only a writing problem. It's a judgment problem.

A strong risk narrative has to do something harder than sound polished. It has to surface what's materially different, newly emerging, or specific to this business in this moment. If the tool that helps you write faster also nudges you toward language that's fluent and comparable but less sensitive to what's actually changed, you end up with better-looking documents and weaker signals.

The real question

None of this argues for removing AI from the disclosure process. The efficiency gains are real, and in many parts of corporate reporting, standardization is exactly what you want. The more useful takeaway is narrower: firms need to be deliberate about where AI drafts and where humans review.

That means checking whether new risks were genuinely added, whether changed business conditions are reflected in the wording, and whether the final text still contains enough company-specific detail to be decision-useful. One question cuts through the noise: did the company's actual risk profile change more than the narrative did?

For investors and analysts, it suggests a parallel discipline. Don't mistake smoother language for better disclosure. A shorter document may save time while still saying less.

A final, somewhat uncomfortable finding

Higher AI usage in risk sections is associated with more retail trading — but not with a clear deterioration in overall price efficiency. Which suggests the real problem isn't a market breakdown. It's a widening gap: sophisticated investors can supplement these disclosures with other signals. Less sophisticated ones are more likely to take the polished document at face value.

For years, the criticism was that risk disclosures were too long, too boilerplate, and too hard to process. GenAI can fix that. But in fixing it, it may sharpen a different problem — the loss of the company-specific texture that made these sections worth reading in the first place.

The most dangerous sentence in a risk filing might be the one that reads perfectly and says almost nothing new.