The First Annual Reports of the LLM Era
Language Change in 150 SEC 10-K Filings, 2019–2024
Published March 2026
Abstract
This study examines 150 SEC Form 10-K filings across 50 of the largest US public companies to measure language change across three observation points: FY2019 (pre-pandemic baseline), FY2022 (pre-enterprise LLM baseline), and FY2024 (the first filing cycle following widespread enterprise LLM accessibility). Using a mechanically verified 6-dimension rubric scored on a 0-to-2 scale, the analysis tracks six editorial forms of prose drift: distinctive claim softening, hedge proliferation, framing separation, named specifics loss, sentence inflation, and generic new section insertion. Excluding structurally induced cybersecurity disclosures, the annual rate of editorial prose drift was 24.5 per cent higher in the 2022–2024 interval than in 2019–2022, rising from an average of 0.71 to 0.88 points per company annually. The composition of drift also evolved: while early drift was heavily concentrated in sentence inflation (29.2 per cent share), later drift became highly diffuse, with hedge proliferation (25.0 per cent), sentence inflation (25.0 per cent), and framing drift (22.7 per cent) sharing dominance. Sector-level acceleration varied sharply, led by industrials (+175 per cent) and technology (+86 per cent), while financial services drift slowed (-31 per cent). The study establishes that in the first reporting cycle following enterprise LLM tool deployment, corporate disclosure prose changes faster and in less concentrated, harder-to-detect patterns.
Suggested citation
Foster-Fletcher, R. (2026). The First Annual Reports of the LLM Era: Language Change in 150 SEC 10-K Filings, 2019–2024. MKAI Working Paper.
PDF size: 139 KB
Record metadata
- Type
- Published working paper
- Theme
- Public reporting and disclosure
- Status
- Published
Methods
- Longitudinal corpus comparison of 150 SEC Form 10-K filings across 50 large US public companies (FY2019, FY2022, FY2024)
- Deterministic passage extraction across Items 1, 1A, and 7
- Structured 6-dimension scoring rubric (0 to 2 scale, max 12) with mechanical extraction verification
- Annualized interval rate comparison and sector-level composition profiling