AI risk has moved rapidly into the language of British corporate reporting, but detailed explanations of how companies manage it remain unusual. An independent research paper analyzing 9,821 annual reports from 1,362 UK-incorporated listed companies found that the share of reports mentioning AI as a risk rose from 2.8% in 2020 to 41.2% in 2025. Only 4.3% of all 2025 reports, however, contained AI-risk disclosures that the study classified as substantive.

The project, called the AI Risk Observatory, used a two-stage language-model pipeline to examine reports published from 2020 through 2025, plus partial 2026 data that the author treated separately. It first filtered passages using a fixed list of AI-related terms, then classified 24,189 extracted passages by whether they described adoption, risk, a vendor, a realized harm or only a general mention. Further classifiers grouped risk categories, named providers and the specificity of the disclosures.

The method was checked against 474 human-annotated passages drawn from 30 reports across 15 companies. For the high-level mention labels used in the headline trends, recall ranged from 0.91 to 0.97, while precision varied by label. Risk detection reached 0.82 precision and 0.97 recall. Agreement weakened for finer-grained categories: mean similarity between the model and reference labels was 0.33 for the ten-part risk taxonomy. The paper therefore treats detailed category findings more cautiously than broad trends.

Thousands of annual-report pages pass through a two-stage AI classification system.
The study used an LLM pipeline to classify 24,189 AI-related passages from 9,821 annual reports.

The headline trend accelerated after generative AI entered mainstream business discussion. AI-risk mentions rose from 4.8% of reports in 2022 to 9.8% in 2023, then tripled to 30.4% in 2024 before reaching 41.2% in 2025. Adoption disclosures increased from 13.8% in 2020 to 45.2% in 2025. Even so, 58.8% of the 1,561 reports processed for 2025 did not mention AI risk at all.

Companies also began naming a wider set of concerns. Strategic and competitive risk remained the most common category in 2025, appearing in 27.4% of all reports, followed by cybersecurity at 25.7% and operational or technical risk at 24.2%. Information-integrity concerns such as misinformation and deepfakes rose from one report in 2020 to 162 reports in 2025. Workforce-impact mentions increased from three reports to 132 over the same period.

Disclosure differed sharply by listing segment. Among reports tagged as Main Market filings, 57.8% mentioned AI risk in 2025, compared with 6.4% of AIM reports. FTSE 100 reports reached 68.6%. The author says the gap may reflect differences in governance rules, company size and reporting resources, but the text alone cannot determine why one group discloses more than another.

Sector results also exposed uneven reporting. Data Infrastructure had an AI-risk mention rate of 10% across 20 reports in 2025, even though its AI-adoption disclosure rate was 55%. Energy reported AI risk in 20% of 140 reports and adoption in 21%. Because some sector samples were small, the paper describes rates based on fewer than 30 reports as indicative rather than definitive.

Many AI warning flags stand above mostly closed boxes, while a few reveal concrete safeguards.
In 2025, 41.2% of reports mentioned AI risk, but only 4.3% included disclosures rated substantive.

Named vendor disclosures clustered around major providers. Microsoft appeared in 98 reports in 2025, followed by Google in 47, Amazon in 42 and OpenAI in 38. Yet only 19.2% of all 2025 reports named any AI vendor, so the observed concentration may not represent companies’ full supplier relationships. Vendor passages also scored as more substantive partly because naming a provider automatically satisfied one element of the study’s specificity rubric.

The largest gap was between mentioning risk and explaining it. Of the 643 reports that mentioned AI risk in 2025, only 67 were rated substantive, meaning they described a specific mechanism along with concrete controls, named systems or measurable targets. Most risk disclosures were classified as moderate: they identified an area such as compliance or cybersecurity but did not describe mechanisms or mitigations. The study says this could reflect the conventions of annual reports, immature governance, compliance signaling or a strategic preference for vagueness; it cannot distinguish among those explanations.

Annual reports also proved poor at revealing realized harm. Automated screening identified possible harm passages in only seven of the 9,821 reports, and manual review found five passages across four reports that described actual AI-caused incidents. The author warns that this says little about the real frequency of harm because incidents may surface instead through litigation, enforcement records or specialist databases.

The findings measure disclosure, not corporate exposure or governance quality. A company with detailed language is not necessarily safer, and silence could mean low relevance, weak recognition or a deliberate choice not to disclose. The validation set relied on one annotator, did not independently cover AIM filings or early years, and the specificity score involves subjective boundaries. The paper nevertheless shows how LLMs can turn thousands of annual reports into a repeatable monitoring signal—provided that signal is not mistaken for the underlying reality.