The latest developments in AI have the potential to significantly support qualitative analysis of interview transcripts. This study explores the utility of the OpenAI o1-model to assist in efficiently obtaining reliable summarizations of safety culture interviews. Analysis shows that the current approach has the clear potential to significantly improve efficiency of interviewers by providing a concise report that summarizes multiple interviews according to a pre-defined format. However, some hallucinations are present in the generated report. Additional work will aim at reducing their presence, but such hallucinations also emphasizes that LLMs should primarily assist, rather than replace, interviewers in creating a definitive report.

In this article, the authors apply the intervention mapping (IM) protocol to develop safety leadership training for a rail infrastructure maintenance company. The IM protocol helps to create an evidence-based intervention in a structured way, based on concrete evidence. The application of IM within the occupational safety domain is limited, a research gap that this article bridges with the development and testing of a safety leadership intervention to promote safety behavior among managers. The company was positively and actively engaged in the training program thanks to the IM protocol. The local support group took full advantage of the opportunities to provide input during the development of the training’s various components. Despite this, interpersonal problems within the leadership team itself, such as a lack of psychological safety, were not identified during the needs assessment. These issues had an impact on the overall effectiveness of the training, as they manifested during the training when managers met physically for the first time in several years (due to the coronavirus). Our IM protocol will be adjusted accordingly for future applications, and we hope that sharing our experiences will enable fellow researchers to avoid this problem.

In the Netherlands, approximately 2.300 workers have a serious reportable accident at work every year, of which around 60 are fatal (Inspectie SZW, 2020; Bellamy et al., 2014). Safety practitioners employ many methods to improve occupational safety for workers within their companies. Interventions might, for example, be aimed at improving companies’ overall ‘safety culture’, at the introduction of a safety management system (e.g. Robson et al., 2007), or at improving the compliance of workers to specific safety rules (e.g. Peuscher and Groeneweg 2012; Bryden et al., 2016). However, the effectiveness of many of those interventions remains largely unclear (Dyreborg et al, 2015). The Dutch National Institute for Public Health and the Environment (RIVM) has started a project with the ultimate goal of developing a database filled with effective safety interventions. Developers can submit their interventions using a fixed protocol. To support this project, we developed a survey, which was sent to all members of the Dutch Society for Safety Science (NVVK). In the survey, we used a list of 48 predefined descriptions of common interventions. Respondents could indicate whether they made use of these common interventions and the extent to which they considered these effective. The survey thus provided an extensive overview of the use and perceived effectiveness of 48 specific safety interventions. In the future, these insights can support the development and testing of more effective safety interventions.

Although major hazard industries have no legal obligations to demonstrate a good safety culture, it is an increasingly important factor for risk-based inspections. This paper describes the development and evaluation of a relatively simple method for external indicative assessment of the process safety culture in major hazard industries. As a first step, a tool was developed. Second, the tool was successfully tested in a workshop with major hazard inspectors. After this ‘proof of concept’ and a few adaptations to improve the tool, quantitative evaluation and validation were carried out with a representative sample of major hazard inspectors in the Netherlands. In this stage, the tool comprised 30 questions, of which 19 sufficiently frequent valid responses were given for the validity test. The evaluations demonstrate that without specific preparations, inspection teams can use the tool to discriminate between plants with poor or good safety cultures. That is sufficient for ranking as additional input for risk-based inspection strategies. The resulting indicative ‘scores’ for safety culture correlate meaningfully with the quality of process safety management systems and the number of (legal) non-compliances, which were identified by the inspection team at the same occasion. The use of the tool can be further improved by training inspectors, making minor adaptations of the inspection process, and broadening the scope of a few questions. It is expected that external assessments of safety culture will become increasingly important not only for inspectorates, but also for a range of other external stakeholders.

Quantitative Risk Analysis (QRA) is a standard tool in some high-risk industries (such as the on- and offshore exploration and production and chemical industry). Presently, existing knowledge concerning human error likelihood and human reliability assessment is insufficiently represented in QRAs. In this paper we attempt to implement the quantification of the human factors in a QRA, which we call QRA+. We analysed a specific incident scenario: the risk of overfilling chemical storage tanks that operate at atmospheric pressure. This scenario was chosen because it is a relevant example of a high-risk scenario in the chemical industry. We identified relevant technological and human parameters within this scenario through on-site visits and interviews with site-experts. The quantitative knowledge concerning the technological parameters was obtained from officially documented SIL statistics, whereas the Standardized Plant Analysis Risk-Human Reliability analysis (SPAR-H) was used to quantify the human factors. Beta distributions were used to model failure probability distributions to account for the uncertainty inherent in dealing with human reliability. For seamless integration of existing qualitative and quantitative knowledge, we made use of a Bayesian Belief Network. The resulting model provides an integrated and more accurate estimation of the failure probabilities for both technological and human factors and the uncertainty surrounding such probability estimates. Furthermore, it gives insight in where these failure probabilities originate and how they interact. This will allow companies to identify those parameters they need to influence to get optimal results concerning their management of risk.

Last minute risk assessment (LMRA) is a well-known work method to support employees’ risk perception. However, little is known about the effectiveness of LMRA in providing this support. Here, we describe an eye-tracking experiment with which we attempted to gain more insight into the relationship between LMRA and risk perception and to assess the difference between generic and specific supporting questions. Employees from an international energy production and desalination company participated in this experiment by assessing photographs portraying a (staged) work situation and deciding whether it was safe enough to continue activities and which risk factors were present or absent. The results show a consistent interaction effect over several parameters between work experience and the type of supporting questions, indicating that generic and specific supporting questions should be considered complimentary to each other. Furthermore, the results revealed several other challenges concerning real-world application of the LMRA.