The language of bank risk

What moved

• AI entered the vocabulary from literally zero: 0.0% of banks' 10-Ks mentioned artificial intelligence or machine learning in 2016 (none of that year's filers did, JPMorgan included), against 87.8% in 2026.
• "Bank failure" went from background vocabulary (15.3% of 2023 filings, most filed just before Silicon Valley Bank failed in March 2023) to 74.9% in the first full post-SVB filing season.
• "Uninsured deposits" began climbing BEFORE the 2023 failures: 16.3% of 2021 filings, 56.4% of 2022 filings, 89.6% by 2024. The page reports the counted timing and does not assign a cause.
• Cyber reached 100.0% of filings in 2024: the first year in the panel in which every bank's 10-K mentions it.
• Pandemic language peaked at 99.5% in 2022 and was still at 87.3% in 2026: risk sections accrete language faster than they shed it.

Quarterly resolution around the 2023 failures

Adding the 10-Qs gives roughly four observations per bank-year, which resolves the turn the annual series blurs. Silicon Valley Bank failed on March 10, 2023: filings lodged in 2023Q1 (mostly 10-Ks, filed in February) against 2023Q2 (10-Qs) straddle it. "Uninsured deposits" went from 62.8% of banks filing in 2023Q1 to 76.1% in 2023Q2; "bank failure" from 14.0% to 75.1%. Composition runs AGAINST these rises: Q2 filings are 10-Qs, far shorter documents than 10-Ks, so the jump happens despite fewer pages in which to say it.

Share of banks’ 10-Ks mentioning each term

By filing year (a 10-K filed in early 2026 covers fiscal 2025). The panel is unbalanced: banks merge, delist and fail out of it, so each year's denominator is that year's filers (189 in 2026).

First year against latest, ordered by movement

Method, verbatim

• Corpus. Primary 10-K documents fetched from SEC EDGAR (public domain), forms exactly 10-K (amendments excluded), filed 2016-01-01 onward, for the fetched slice of SIC 6021/6022 commercial banks: 314 banks with qualifying filings. This is the corpus's documented coverage boundary, not all US banks; the largest banks are in it because the fetch was banks-first.
• Text. The document’s HTML with script and style blocks, all markup and entities removed, whitespace collapsed. One 10-K per bank per filing year (the larger document is kept when a late filer files twice in one year).
• Counting. A mention is one case-insensitive match of the term’s pattern. The share is the fraction of that year’s banks with at least one mention. Nothing is weighted, scored or model-interpreted.
• Quarterly series. The quarterly cards pool 10-K and 10-Q primary documents, one document per bank per filing quarter (the larger kept), with that quarter’s filers as the denominator. Quarters alternate in composition: Q1 is 10-K-heavy, the rest are 10-Q-only, and 10-Qs are much shorter documents; the page reads rises into 10-Q quarters as happening despite that, and the annual 10-K series above is unaffected by the quarterly file.
• The 2026 year is partial. The corpus was fetched in July 2026; most 10-Ks file January-April, so the year is close to complete, but late filers are missing.

Source: SEC EDGAR 10-K filings (US government work, public domain). Structured counterpart: as-filed XBRL · filings methodology.