# Bank AI Disclosure Tracker: data downloads

What US banks say about artificial intelligence in their SEC filings: annual reports (10-K), quarterly reports
(10-Q) and earnings releases and investor presentations (8-K exhibits), from 2023 onward.

Licence: CC BY 4.0. Free to use, share and adapt with credit to the Bank AI Disclosure Tracker.
Not investment advice. Labels are generated by an AI model and checked against a hand-labelled sample; they can
be wrong. Quotes are verbatim from the filings, and every passage links to its source.

## Files

Each file comes as CSV (opens in Excel or Google Sheets) and Parquet (for analysis tools). In CSV files, list
values are separated by "; ".

### banks.csv: one row per bank per report year

| Column | Meaning |
| :- | :- |
| bank, cik, slug, state | Bank name, SEC CIK number, tracker page id, head office state |
| fiscal_year | Report year: the fiscal year the annual report covers (not the year it was filed) |
| size_group | Bank size: total assets of the bank's FDIC-insured bank subsidiaries at the fiscal year end (Under $1B, $1B to $50B, $50B and above) |
| assets_bn | Those total assets, in billions of US dollars |
| passages | Passages in the annual report that mention AI (outside legal notices) |
| substantive | Passages that say something about AI beyond a passing mention |
| max_detail | Highest detail level: 1 General, 2 Names an area, 3 Concrete example |
| using_ai_now | The report says the bank is using AI now |
| testing_or_planning | The report says the bank is testing or planning AI |
| explains_controls | The report explains how AI is controlled (committee, policy, model risk management, human review, training) |
| sees_ai_as_risk | The report lists AI as a risk |
| risk_without_controls | Lists AI as a risk but no text was found on how AI is controlled. This is about the report, not the bank's practice |
| mentions_genai, mentions_agentic | Mentions generative AI; mentions agentic AI or AI agents |
| names_ai_company | Names an AI product or company |
| concrete_detail | At least one passage gives a concrete example (detail level 3) |
| use_case_areas, vendors_named | Areas where the bank says it uses or plans AI; AI products or companies named |
| new_passages, removed_passages | AI passages new since, or no longer present from, the previous year's report |

### paragraphs.csv: one row per passage that mentions AI

| Column | Meaning |
| :- | :- |
| passage_id | Stable id: accession number plus a text hash |
| bank, cik, slug | The bank |
| filing_kind, form, accession, filing_date | The filing (annual, quarterly or earnings material) |
| fiscal_year, fiscal_quarter | Report period |
| section | Part of the filing, for example Risk factors, Management's discussion, Earnings release or Investor presentation |
| page, speaker | Page or slide number; the executive quoted, for earnings releases |
| is_legal_notice | Part of a forward-looking statements notice; never counted |
| under_review | The label could not be checked automatically and is held back from all counts |
| what_it_says | Using AI now; Testing or planning AI; Explains how AI is controlled; Sees AI as a risk; General statement about AI; Standard wording or passing mention |
| detail | General, Names an area, or Concrete example |
| use_case_areas, technologies, vendors_named, governance_signals, risk_framing | Labels from the passage |
| standard_wording_banks | When most of the passage is standard wording: how many other banks use the same sentences that year |
| new_this_period | 1 when the passage is new compared with the bank's previous filing of the same kind |
| evidence_quote | The part of the passage that supports the label, copied exactly |
| text | The passage, verbatim |
| link | Opens the filing on sec.gov at this passage |

The Parquet version of this file also has the label codes behind these columns: mention_type, detail_level and section_code.

### filings.csv: one row per filing

Accession, bank, filing kind, form, filing date, report period, passage counts, highest detail level, the same
yes or no columns as banks.csv, share of standard wording, new passages, and passages held back for review.

## How the data is made

See the methods page on the site. In short: every 10-K, 10-Q and earnings 8-K filed since 1 January 2023 by US
banks filing with the SEC (SIC codes 6021 and 6022, plus seven large bank holding companies filed under other
codes) is downloaded from EDGAR; passages are found with a published term list; an AI model (Claude Sonnet 4.6 on
Amazon Bedrock) labels each passage; standard wording is detected by matching sentences across banks; bank size
comes from FDIC call report data.