Generative AI in Financial Reporting Elizabeth Blankespoor ...
First large-sample evidence finds generative AI in corporate financial writing, validated with GPTZero GenScore. The detector shows low false positives and flags edits as small as 0.0625 percent. AI appears across releases, call scripts, Risk Factors, MD&A, and S-1s in 2024, concentrated in new text. Rates reach 4.3 percent in Risk Factors, 2.5 in MD&A, and 4.5 in S-1 descriptions. Adoption is higher at smaller firms and big news. AI makes prose clearer and specific.
What it examines
The paper studies how much generative AI is used to write financial disclosures. The authors first validate a commercial AI-text detector (GPTZero) for filings, then apply it to earnings releases, call remarks, risk factors, MD&A, and S-1s (2010--2024), using regressions and “new text” tests to measure usage and linguistic effects.
What it concludes
They find meaningful AI use across disclosures (up to 4.5% of new text in 2024), especially at smaller firms; AI text is clearer, more positive, and more specific. Applications: monitoring, auditing, compliance, IR, research. Limits: detector reliance and lower bounds. Future work: costs, controls, market effects.
Evidence objects
First large-sample study rigorously detects generative AI in financial writing using GPTZeros GenScore: low false positives, detects 0.0625% AI edits, data released, validated in reports, sentence-level insertion, boilerplate stripped, difference-in-differences.
key_findings bullet 1 · key_findings · validation V0
GAI appears across earnings releases, call remarks, Risk Factors, MD&A, and S1s through 2024, concentrated in new text: 4.3% Risk Factors, 2.5% MD&A, 4.5% S1 descriptions; overall, 0.5--1.4%, tone unchanged.
key_findings bullet 2 · key_findings · validation V0
Adoption skews to smaller firms, no IR, big news, low fees; AI reads positive, specific, readable. Limits: undercounting, detector drift, endogeneity. Bottom line: growing; research on costs, controls, markets, regulation.
key_findings bullet 3 · key_findings · validation V0
An early, systematic measurement of generative AI in corporate disclosures through 2024, this paper validates GPTZero for finance texts and documents non-trivial usage (up to 4.5% ). Its empirical focususage metrics, detector validation, and a released datasetoffers novel, actionable evidence relevant to LLM-driven analysis, compliance, and preserving quant text-signal integrity stakes.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
Journal of Accounting and. Economics, 72(2-3), p.101429. de Kok, T., 2025. ChatGPT for Textual Analysis? How to use Generative LLMs in Accounting Research.
Source row: 980 · abstract type: snippet