The First Annual Reports of The LLM Era

The 2024 filing season was the first full cycle of SEC annual reports produced after enterprise-grade large language model tools became widely available. That made it possible to compare the language of annual filings written in the LLM era against the language that came before it.

I published a study earlier this year doing exactly that, covering 150 10-K filings from 50 large US public companies across three filing years: 2019, 2022, and 2024.

What I found is that the annual rate of language change was 24.5 per cent higher in the later interval than in the earlier one.

The study scored each company on six dimensions of language change. The simplest to explain is the first: can the claim be checked, or can it not? A company that says it manufactures a specific set of products has made a verifiable claim. A company that calls itself "a leader" has not.

The remaining five measure related shifts: whether hedging is accumulating, whether named specifics are disappearing, whether sentences are growing longer without saying more, and whether the framing around concrete examples is loosening. Each one tracks a different way that filing language can become less precise without becoming inaccurate. The full methodology is published on SSRN.

Here is what that looked like in practice.

In FY2022, AMD described itself as "a global semiconductor company primarily offering embedded CPUs, GPUs, APUs, FPGAs, and Adaptive SoC products."

By FY2024, that became: "AMD is the high performance and adaptive computing leader, powering the products and services that help solve the world's most important challenges."

AMD's business did change in that period. It moved further into AI infrastructure, adaptive computing, and data centre competition. The study does not dispute that. It shows that this commercial change was expressed through language that became broader and less testable. Product language can be checked directly, leadership language cannot.

What the later interval looked like

If the study had only asked when each company changed the most, the finding would have been unremarkable. Twenty-five companies showed their sharpest language change between 2019 and 2022, before enterprise LLM tools were widely available. Seventeen showed their sharpest change between 2022 and 2024. Eight were evenly split. On that count alone, the earlier period produced more movement. What stood out about the later period was what the change looked like.

In the earlier period, comparable passages got longer without becoming much more informative. That is a familiar kind of filing drift.

In the later period, the filings started to sound different.

The language became markedly more qualified.

By that I mean wording that leaves more room around a claim: terms such as "may," "could," "seeks," "aims," and "expects," along with similar forms that widen what a sentence can mean without making it false. This kind of language increased measurably across the later interval.

The relationship between the claim and the supporting detail weakened. In earlier filings, a paragraph was more likely to move directly from the point being made to the fact or example supporting it. In later filings, the opening wording was more often broader and more reusable, while the supporting detail was less closely connected to it.

Named specifics receded. A named product became "key offerings." A named jurisdiction became "international markets." A cited statute became "applicable regulations."

Comparable passages grew longer through added qualifiers, framing language, and connective wording. In the earlier interval, this was the dominant pattern. In the later interval, it became one part of a broader shift.

That is what stood out. The later interval did not continue the earlier patterns.

Where it happened and where it did not

But not every company in the dataset moved in the same direction. Seven companies scored 2 or below across all dimensions. NVIDIA, in fact, added product names and country groupings that were absent from earlier filings. Morgan Stanley's identity language stayed close to verbatim across all three of the filings.

The pattern was visible at sector level too. Technology companies showed 86 per cent more language change in the later interval. Industrials showed 175 per cent more. Financial services showed 31 per cent less. These are small subgroups, between 6 and 12 companies per sector, so the figures should be read as indicative.

The first annual reports filed after enterprise LLM tools became widely available look different from what came before. The difference is uneven across companies, sharper in some sectors than others, and broader in form than the earlier drift. The pandemic, legal caution, peer benchmarking, and template convergence are all part of the picture. The study does not claim to separate them. What it documents is a change in the character of filing language that coincided with the arrival of a new set of tools in the environment where filings are produced.

A Traceability Report documenting AI use in the production of this research is published at mkai.org.