Abstract
<p>‘Safety climate’ refers to the shared attitudes about safety within an organisation. While traditionally measured via self-report surveys, it can now be measured by analysing the language in written incident reports. Using large-language model-based machine text analysis, we examined whether text-derived measures predict future safety outcomes in four high-risk industries (i.e., rail, nuclear, mining and aviation) and found that safety climate operates through two distinct pathways. First, stronger safety climate predicts less harm and operational failures, suggesting a preventative effect. Second, a stronger safety climate predicts more formal reporting and documentation, reflecting a culture in which workers feel comfortable raising concerns. To ensure these findings were not analytical artefacts, we tested 810 reasonable variations of each analysis using held-out data for validation. Some findings proved robust across all approaches; others depended on specific methodological decisions. This framework demonstrates how organisations can use large-language model-based text analysis to track safety culture dynamically, while validating these measures against real-world outcomes.</p>