AI risk disclosures in UK annual reports rise sharply but substantive content remains rare

A new arXiv preprint introduces a reproducible two-stage classification pipeline that analyzed 9,821 annual reports from 1,362 UK-listed companies spanning 2020 to 2025 (with partial 2026 data). The pipeline was validated against 474 human-annotated passages, achieving high recall and moderate label-level agreement.

The study reports a sharp increase in AI-related disclosures over the five-year period: the share of reports mentioning AI risk rose from 2.8% in 2020 to 41.2% in 2025, while AI adoption disclosure increased from 13.8% to 45.2%. Named vendor mentions in reports clustered around a small set of major providers, with Microsoft identified as the most frequently cited.

Disclosure patterns varied significantly by sector and market segment. Companies listed on the AIM market disclosed AI risk at far lower rates than those on the Main Market, and sectors such as Energy and Data Infrastructure lagged behind others in AI risk disclosure.

The authors developed a substantiveness classification to assess the quality of AI risk disclosures. Despite the rise in mentions, most disclosures were not substantive: in 2025, 41.2% of all reports mentioned AI as a risk, but only 4.3% contained AI risk disclosures classified as substantive.

Harm disclosures were nearly absent across the entire corpus, with only seven reports explicitly addressing harms related to AI.