GPT Applications for Construction Safety: A Use Case Analysis

Abstract

This study explores the use of Large Language Models (LLMs), specifically GPT, for different safety management applications in the construction industry. Many studies have explored the integration of GPT in construction safety for various applications; their primary focus has been on the feasibility of such integration, often using GPT models for specific applications rather than a thorough evaluation of GPT’s limitations and capabilities. In contrast, this study aims to provide a comprehensive assessment of GPT’s performance based on established key criteria. Using structured use cases, this study explores GPT’s strength and weaknesses in four construction safety areas: (1) delivering personalized safety training and educational content tailored to individual learner needs; (2) automatically analyzing post-accident reports to identify root causes and suggest preventive measures; (3) generating customized safety guidelines and checklists to support site compliance; and (4) providing real-time assistance for managing daily safety tasks and decision-making on construction sites. LLMs and NLP have already been employed in each of these four areas for improvement, making them suitable areas for further investigation. GPT demonstrated acceptable performance in delivering evidence-based, regulation-aligned responses, making it valuable for scaling personalized training, automating accident analyses, and developing safety protocols. Additionally, it provided real-time safety support through interactive dialogues. However, the model showed limitations in deeper critical analysis, extrapolating information, and adapting to dynamic environments. The study concludes that while GPT holds significant promise for enhancing construction safety, further refinement is necessary. This includes fine-tuning for more relevant safety-specific outcomes, integrating real-time data for contextual awareness, and developing a nuanced understanding of safety risks. These improvements, coupled with human oversight, could make GPT a robust tool for safety management.

1. Introduction and Literature Review

The construction industry is considered to be one of the most dangerous industries due to the numerous hazards that workers may encounter on jobsites [1]. In 2022, the construction industry accounted for 19.2% of all worker fatalities in the United States, which means that one out of every five worker deaths occurred in this industry [2]. Given the high incidence of workplace fatalities in the construction industry, several efforts have been made to improve safety performance [3]. Considering the low safety managers to worker ratio, automating the safety management process has emerged as a need to elevate safety in the construction industry [4]. Several methods have been proposed for this purpose, such as using Building Information Modeling (BIM) to identify potential safety issues [5,6], virtual reality (VR) and augmented reality (AR) to provide automated safety training for workers [7,8,9], using robots and artificial intelligence (AI) to detect and manage safety hazards [10,11], and language models to analyze safety documents for hazard causes [12]. Despite these advancements, the injury rates in construction remain notably high. Therefore, it is clear that further attention is needed to improve safety performance in the construction industry [13].

In recent times, advancements in computational capabilities and the vast amount of available data have contributed to the surge in popularity and effectiveness of Natural Language Processing (NLP) across various sectors. NLP, a subfield of artificial intelligence, trains computers to interpret human language both in written and spoken forms, akin to human comprehension [14]. NLP combines rule-based modeling of human language with statistical, machine learning, and deep learning techniques to enable computers to comprehend the meaning and context of human language, including the speakers’ or writers’ intentions and emotions [15]. This technology holds particular promise for the construction industry, which is often laden with vast documentation, reports, and correspondence. Given the industry’s pressing need to enhance safety and operational efficiency, the application of NLP presents an opportunity for transformational change. In the past few years, the construction industry has increasingly adopted natural language processing (NLP) to increase automation [16,17]. For instance, it has been applied in areas like document management [18], risk management [19], quality assurance [20,21], scheduling [22], and building information modeling (BIM) [23]. Furthermore, the use of NLP and AI on accident datasets has also been recommended by European Statistics on Accidents at Work (ESAW) [24,25]. NLP has especially proven valuable in enhancing safety management in construction projects. Specifically, NLP is used to analyze safety reports [26], extract safety information and knowledge from safety codes and documents [27], and identify the root causes of hazards [28]. Moreover, NLP-based tools have also been developed to generate content from existing information for safety training and education [29], and to determine and analyze the relationship between different safety risk factors [30]. By serving these diverse purposes, NLP has proven to be a promising tool for achieving safer and more efficient construction environments [26].

Recent advancements in NLP have been significantly driven by the emergence of Large Language Models (LLMs), such as OpenAI’s Generative Pretrained Transformer (GPT) [31] and Google AI’s Bard [32]. These models, trained on extensive text datasets, generate text with human-like fluency, answer questions, and execute a host of language-related tasks with good accuracy [33]. Their inherent strength lies in their extensive pretraining on diverse data, enabling versatility across various applications [34,35]. Given the significant progress made in NLP through Large Language Models, there is a compelling rationale to use them for automating construction safety management. Industries across the spectrum are already tapping into the potential of LLMs [36,37], with applications ranging from coding assistance [38] to education [39] and content generation for advertisement purposes [40]. Notably, the construction industry is also beginning to integrate LLMs, with preliminary efforts on safety training [41] and scheduling [42]. These studies largely emphasize preliminary demonstrations or speculative potential without critically examining how LLMs perform in the nuanced, high-risk context of construction safety. They often present isolated applications rather than synthesizing insights across use cases or rigorously evaluating LLM effectiveness and limitations. As a result, it remains unclear whether LLMs offer tangible benefits beyond automation, particularly in tasks that require domain-specific reasoning, compliance with safety standards, and high levels of reliability. This gap underscores the need for a systematic assessment of LLMs such as GPT across key safety management tasks to better inform their potential role in real-world construction environments.

In recent times, the wave of popularity of LLMs has led to a significant increase in research proposing their application across a wide range of construction-related tasks. However, much of this momentum appears driven by the trend itself rather than a rigorous evaluation of the LLMs’ relevance, effectiveness, or limitations. Consequently, there is growing concern that this rapid adoption of LLMs may eclipse more evidence-based, methodical approaches necessary for addressing the complexities inherent in construction research. This trend is particularly concerning in areas like construction safety, where the misuse of technology can lead to ineffective or even dangerous outcomes. While existing studies have explored general applications of AI and NLP in construction, few have conducted a critical synthesis of how LLMs specifically perform when tasked with safety-critical responsibilities that demand contextual awareness, regulatory alignment, and reliable outputs. There remains a notable absence of systematic evaluations that examine both the capabilities and boundaries of models like GPT when applied across distinct safety functions. Without such evaluations, it is difficult to determine whether current LLMs can truly support the nuanced decision-making required in high-risk environments. Therefore, this study aims to investigate the application of LLMs in construction safety management, with a focus on critically assessing their potential benefits, limitations, and practical viability. By conducting a structured use case evaluation, this research not only explores the practical implications of using LLMs in safety management but also helps to establish a more grounded foundation for future integration of LLMs into safety-critical domains. While several LLMs have emerged recently, GPT stands out as particularly relevant for this study.

GPT is a Large Language Model, based on the transformer architecture. It operates by predicting the next word in a sequence, using patterns learned from large-scale text datasets during its pretraining phase. This autoregressive structure allows GPT to generate fluent and contextually relevant responses across a wide range of topics [43]. Its widespread popularity provides several resources and community support, indicating its potential for more extensive future use in safety-related fields. The model’s training on diverse datasets and advanced language processing abilities are especially suited for interpreting and applying complex safety protocols and guidelines in varying contexts, which is crucial for construction safety. Finally, its potential for integration with other technologies positions GPT as a promising choice for developing safety management solutions, anticipating its growing role in safety applications in the future. However, its outputs are shaped entirely by statistical correlations in the data, without any embedded understanding of meaning or intent. Additionally, GPT lacks the ability to verify facts, access real-time information, or incorporate multimodal inputs unless explicitly integrated through external tools or prompt engineering. These architectural features can lead to limitations in high-risk domains like construction safety, including hallucination of inaccurate information, shallow responses in technical scenarios, and limited adaptability to dynamic site conditions. Furthermore, without fine-tuning on domain-relevant safety documents or feedback loops, GPT’s outputs may not fully reflect regulatory nuances or site-specific risks. Recognizing these architectural constraints is critical when evaluating GPT’s performance in construction safety applications [44,45,46]. For this study, GPT-4o was selected as the language model due to its accessibility and relevance to general users. As a publicly available version of GPT, it reflects the model capabilities that are most likely to be used in real-world construction environments.

Despite the growing use of LLMs in various domains, there has been no systematic evaluation of their role in construction safety management. Prior work tends to focus on feasibility or isolated applications without assessing the breadth and depth of LLM capabilities across multiple safety functions. This gap limits our understanding of how tools like GPT can be reliably and ethically applied in safety-critical environments. This study addresses this need by evaluating GPT’s performance in structured use cases aligned with key construction safety tasks. These use cases are personalized safety training and education, automated post-accident analysis, automated generation of safety guidelines and checklists, and real-time safety management assistance. Construction safety practices are informed by several foundational theories that guide training, analysis, and decision-making processes. Personalized safety training is grounded in constructivist learning [47] theory and adaptive instructional design, which emphasize tailoring educational experiences to individual needs, enabling learners to build knowledge through active engagement and contextual feedback. Root cause analysis theory and systems thinking [48] underpin post-accident evaluations, facilitating a deeper understanding of how organizational and procedural factors contribute to safety incidents. The development of safety guidelines and checklists reflects principles of procedural knowledge and compliance theory [49], focusing on translating regulatory standards into structured, task-specific protocols. Real-time safety management draws from situational awareness theory and decision support frameworks [50], which emphasize timely access to information, hazard recognition, and responsive actions in dynamic site conditions. These theories collectively provide a robust foundation for improving safety outcomes in complex and high-risk construction environments.

The findings of this study are expected to have implications for both the construction industry and the field of safety management research. The study outcomes offer practical insights for effectively integrating GPT into current construction safety management processes, identifying areas where such models can be applied with efficiency and safety. Furthermore, this study identifies the specific strengths and weaknesses of GPT, aiding researchers and decision-makers in making informed choices regarding GPT adoption for safety while also understanding its limitations and leveraging its potential benefits.

Point of Departure

The potential of GPT in improving construction safety is significant, but a comprehensive understanding of their specific value, strengths, and limitations in safety applications is lacking. To fill this gap, this study aims to critically examine and demonstrate the effectiveness of GPT in various construction safety management applications. The study demonstrates the practical application of GPT through various use cases. This helps in comprehending where GPT can be most beneficial and where it might have limitations. It also highlights the opportunities and challenges in its usage for different safety management applications. This analysis also involves identifying the various capabilities that GPT must demonstrate for effective safety management purposes in the future.

While some studies [41,51,52] have explored the integration of GPT in construction safety for various applications, their primary focus has been on the feasibility of such integration, often using GPT models for specific applications rather than a thorough evaluation of GPT’s limitations and capabilities. In contrast, this study aims to provide a comprehensive assessment of GPT’s performance based on established key criteria, which include interactivity, critical analysis, and comprehension, across different safety management contexts (explained in more detail in Section 2.1). This approach allows for the identification of overarching patterns, potential generalizations, and critical limitations. Additionally, it facilitates cross-application learning, where insights from one case can inform others.

2. Research Method

This study illustrates the use of GPT in different safety management applications through use cases and evaluates its capabilities and limitations under each application. The use case approach helps predict a technology’s behavior under various scenarios as it interfaces with users or other technologies [53,54]. Each use case includes scenarios that illustrate how the technology interacts with the users or other systems to achieve a specific objective. These descriptions typically include the steps that the user or technology will take, expected outcomes, and possible variations in interaction [53]. This approach provides a structured and detailed analysis of technology interactions in real-world scenarios, facilitating a deeper understanding of both its practical applications and potential limitations. Due to its inherent strengths, this approach has been used by researchers in other fields to evaluate LLMs. For example, Saka et al. [54] demonstrated the capabilities and limitations of GPT for the construction industry through use cases in different fields such as sustainability, project management, and safety. In another study, Raj et al. [55] identified the areas of use cases for GPT integration in business operations and analyzed its impacts on the efficiency and effectiveness of the use case areas. Moreover, Ghesmi et al. [56] investigated GPT’s ability in construction cost analysis and bid pricing through a case study of a bridge rehabilitation project. In the context of our study, a use case would detail the specific ways in which GPT could be applied to safety management tasks, such as identifying hazards, training workers, or suggesting safety guidelines. It would describe the interactions between the GPT system and its users, thereby providing insights into its limitations and capabilities in each application area. This would allow us to understand its shortcomings and capabilities for each application area.

Using this approach, the study was conducted by following three steps. Figure 1 demonstrates an overview of the steps of the study:

Step 1: Safety Use Case Identification: The objective was to identify safety use cases for integration of GPT. This included the identification of different application areas where GPT has been used or has the potential to be used in future.

Step 2: Safety Use Case Integration in GPT: This approach entailed assigning specific roles to GPT, corresponding to the responsibilities inherent in each safety management area. The foundational concept used in the integration of GPT was to make GPT emulate the role of human-driven processes in each use case.

Step 3: Safety Use Case Outcome Analysis: Qualitative assessments were carried out by evaluating the responses provided by GPT using four evaluation criteria: (1) Interactivity, (2) Evidence-based Reasoning, (3) Critical Analysis and Extrapolation, and (4) Knowledge and Comprehension.

The following section discusses these three steps of the study in further detail.

2.1. Step 1: Safety Use Case Identification

A preliminary literature review was undertaken to identify potential safety use cases for GPT. The review included the studies that use LLMs for construction safety management tasks. Given that LLMs are advanced implementations within NLP and text mining, it is reasonable that areas that previously employed these technologies are likely to explore the use of LLMs if they have not already done so. Therefore, these areas were also included. The review uncovered four areas where GPT could play a role in safety management processes, which include the following:

Use Case #1: Personalized Safety Training and Education: using GPT to develop and deliver personalized safety learning content.

Use Case #2: Automated Post-Accident Analysis: using GPT to automatically analyze accident reports to determine the cause of accidents and recommend preventive measures.

Use Case #3: Automated Generation of Safety Guidelines and Checklists: using GPT to automate the creation of safety guidelines and checklists.

Use Case #4: Real-time Safety Management Assistance: using GPT to provide real-time assistance in day-to-day safety management tasks.

The following sections discuss each of these identified use cases in further detail.

2.1.1. Use Case #1

Personalized Safety Training and Education: Personalized training and education involve tailoring instructional approaches to meet the unique needs, preferences, and goals of individual learners [57]. This approach differs from traditional models by emphasizing personalized content, flexible pacing, adaptive technologies, and continuous assessment. Personalized training has been shown to enhance engagement, improve outcomes, increase satisfaction, and address diverse learning needs, ultimately leading to a more effective and inclusive learning experience [58]. Personalization is especially crucial for safety training as it requires equipping individuals with the ability to make context-specific decisions in response to dynamic safety challenges. Personalized safety training, characterized by individualized feedback mechanisms, is essential for the cultivation of critical cognitive skills by providing learners with direct, tailored feedback on their decision-making processes [59].

GPT can deliver accessible and personalized training by dynamically generating tailored scenarios and interactive dialogues based on individual learner profiles and responses. These models can analyze inputs from learners in real-time, offering instant, customized feedback and guidance to enhance decision-making skills [60]. LLMs and their predecessors in NLP have been effective in personalized training and education. For instance, Zhou et al. [45] demonstrated the personalization of safety training through the use of chatbots, which were developed using natural language processing techniques to deliver tailored content via mobile phones. This method significantly enhanced workers’ comprehension of safety regulations. More recently, Hussain et al. [39] focused on personalizing safety training by integrating GPT models into VR environments, which provided customized instructions based on individual attributes of migrant workers, helping to overcome language barriers and improve practical safety skills. Similarly, Jamil Uddin et al. [37] investigated the use of GPT in construction safety training, aiming to enhance hazard recognition among construction students. The use of LLMs and NLP in general for personalized education and training extends beyond the construction sector. For example, Sharma et al. [61] introduced an adaptive generative AI system for customized education, which is personalized through real-time sentiment analysis that adapts the educational content based on the learner’s emotional responses, detected via webcam. The growing need for automating personalized safety training in the construction industry, coupled with the increasing utilization of LLMs for personalizing learning content, highlights significant opportunities to integrate LLMs into safety education and training within construction. However, it is crucial to evaluate the limitations and capabilities of GPT for this specific application to ensure its effective and appropriate use in enhancing safety training outcomes.

2.1.2. Use Case #2

Automated Post-Accident Analysis: Post-accident analysis in construction safety management involves retrieving and examining relevant accident case reports to understand the tacit knowledge about when, why, and how accidents occurred, thereby establishing measures to prevent future accidents [19,62]. This analysis is critical for improving safety as it provides actionable insights that can significantly mitigate risks and reduce the frequency of future incidents. Traditionally, this task has been performed by human experts because it requires a nuanced understanding of complex safety standards, legal implications, and situational specifics. Human analysts are adept at interpreting subtle cues and contexts that might be overlooked by automated systems, making their role indispensable in the thorough and accurate assessment of accidents.

With the latest developments in NLP, there is potential to augment this traditional approach. AI can assist by rapidly processing large volumes of accident reports and extracting key patterns and data points that might not be immediately obvious to human analysts. This capability speeds up the analysis process and enhances the depth of insights gained by identifying trends and correlations across multiple incidents [62]. For instance, Kim and Chi [26] automated the extraction of critical safety-related information from unstructured accident reports through NLP, specifically identifying the causal factors of each incident by categorizing hazard objects, positions, and associated work processes, thereby enhancing the effectiveness and precision of post-accident analysis. Similarly, Pan et al. [63] developed a graph-based deep learning framework to automate the identification of accident types, injury types, and body part factors from unstructured construction accident reports. Moreover, Xu et al. [64] developed a text mining and deep learning approach to automate the identification of hot work accident causes from unstructured accident reports, thereby providing detailed insights into causation that aid safety managers in formulating targeted interventions for improved on-site safety management. In another study, Zhang [65] developed a hybrid structured deep neural network with Word2Vec for the automatic classification of construction accident causes, which offers detailed insights into causation through post-accident analysis, facilitating the implementation of precise preventive measures to enhance safety.

Building on these advancements, LLMs offer further opportunities to enhance post-accident analysis in construction safety management. LLMs excel in understanding and generating human-like text, which allows for a more nuanced extraction of data from accident reports, including implicit information and subtleties that may not be explicitly stated. By applying advanced sentiment analysis, context recognition, and predictive analytics, LLMs can provide deeper and more accurate insights into the causes of accidents and potential preventative strategies [66].

2.1.3. Use Case #3

Automated Generation of Safety Guidelines and Checklists: The generation of safety guidelines and checklists is a structured process that involves identifying potential hazards associated with specific job tasks, determining appropriate control measures to mitigate these risks, and documenting these findings in detailed, step-by-step instructions to ensure workplace safety and compliance [67]. This process is critical for safety as it provides clear, actionable instructions to workers, helping to prevent accidents and ensuring that safety protocols are consistently followed. Traditionally, this task has been performed by human experts because it requires a deep understanding of the specific work environment, job tasks, and relevant safety regulations. Human analysts excel at interpreting complex safety data and tailoring guidelines to the unique conditions of each worksite, ensuring that all potential hazards are addressed comprehensively.

Advancements in AI (especially NLP) have provided opportunities to automate this process by quickly analyzing large volumes of safety data and generating initial drafts of safety guidelines and checklists. These systems can identify common hazards and control measures from past reports, project documents, and standards, accelerating the generation process and ensuring critical information is included [68]. For example, Wang and El-Gohary [69] utilized NLP and deep learning to automate the extraction of fall protection requirements from construction safety regulations. The extracted data was organized into a knowledge graph, developing automated tools for compliance checking by providing a framework to assess whether construction site operations adhere to the stipulated safety regulations. Extracting guidelines from regulations or project documents using NLP has also been explored. For example, Borjigin et al. [70] developed a framework that utilizes NLP to extract information from compliance-checking construction regulatory documents, combined with template matching and multidomain semantic modeling.

With recent developments in LLMs, there are new opportunities for refining and enhancing the creation of safety guidelines and checklists. LLMs can generate safety guidelines and checklists automatically by processing vast amounts of textual data, such as past accident reports, safety regulations, and industry best practices. They can extract relevant information, identify patterns and common hazards, and then compile this data into coherent and comprehensive safety documents.

2.1.4. Use Case #4

Real-time Safety Management Assistance: Real-time assistance refers to providing immediate support and information as needed. In the context of safety, real-time safety assistance involves delivering instant guidance and information to managers or workers during operations. For example, if an unfamiliar piece of equipment is introduced, safety managers might want to know the safety requirements of its operation. Similarly, if a hazardous situation arises, such as exposure to hazardous materials, a construction manager might want to immediately know the protocols for addressing the issue and ensuring worker safety. Traditionally, real-time safety assistance has been a human-driven process, where safety managers or experts are responsible for finding the necessary information. This often involves searching through relevant standards and guidelines or seeking advice from other experts, which can be time-consuming and may not always provide immediate solutions.

In the context of Industry 4.0, it is viewed as a virtual assistant or query-answering system designed to provide information support through natural language-based human–machine dialogue [71]. A safety assistant, in this context, is defined as a real-time virtual assistant that provides safety information to support safety managers or workers. The concept is not new in construction and NLP has been used to develop real-time virtual assistants before. For instance, Wang et al. [72] used NLP to assist in information extraction from building information models, facilitating the development of a query-answering system that supports construction project team members by analyzing natural human language queries, identifying relevant building information, and generating comprehensible responses. Similarly, Lin et al. [73] proposed an innovative NLP-based methodology for efficient data retrieval and visualization from cloud BIM. More recently, LLMs have also been utilized to develop real-time safety assistants; for example, Wang and Issa [52] employed LLMs and transfer learning to develop a construction safety chatbot that enhances access to OSHA fall protection regulations. By using models like T5 and GPT-3, their system generates questions from regulatory texts, improving safety guideline comprehension and potentially reducing workplace accidents. In another study, Tian et al. [74] suggested a deep learning, NLP-based technique that can provide automated responses to safety hazard inquiries.

These research efforts collectively illustrate the potential of GPT to be utilized as a safety management assistant capable of responding to safety inquiries from workers. Further evidence of this lies in the extensive research on chatbots [75] that have been successfully employed for various purposes in fields beyond construction, indicating a promising avenue for the deployment of LLMs in similar capacities.

2.2. Step 2: Safety Use Case Integration in GPT

Four detailed scenarios were created to better illustrate how GPT could be used in each use case. Drawing upon existing safety protocols and typical tasks in the construction industry, these scenarios were designed to authentically replicate situations where GPT could be effectively deployed for each of the four application areas. Using inputs from safety standards [76], training resources [77,78], and historical incident data from OSHA and NIOSH [79,80], the scenarios were tailored to present realistic yet demanding challenges, providing an opportunity for a thorough evaluation of GPT capabilities in the next step of the study.

Table 1. Safety context and GPT’s role under each use case.

Use Cases	Scenario Context
Use Case #1: Personalized Safety Training and Education	Safety Context: a session that includes safety training with hazard recognition, hazard management, and safety compliance as learning objectives. GPT Role: creating and delivering the above safety training content.
Use Case #2: Automated Post-Accident Analysis	Safety Context: a detailed real accident report obtained from NIOSH to be analyzed. GPT Role: analyzing the accident report and extracting relevant safety information to identify the causal factors and suggest preventive measures.
Use Case #3: Automated Generation of Safety Guidelines and Checklists	Safety Context: safety guidelines and checklists for welding while working at height. GPT Role: creating safety guidelines and checklists, ensuring alignment with the OSHA safety standards and site-specific requirements.
Use Case #4: Real-time Safety Management Assistance	Safety Context: queries from three key areas of safety management: (1) hazard recognition, (2) hazard evaluation, and (3) hazard control. GPT Role: providing practical answers to the questions by assuming the role of a safety assistant that provides support to the safety manager.

demonstrates different safety contexts that GPT will be tested in and also its roles.

After outlining each scenario, the details of each use case context were provided to GPT using the instructional prompts shown in

Table 2. Instructional GPT prompts for each use case.

Use Cases	Instructional GPT Prompts
Use Case #1: Personalized Safety Training and Education	Contextualizing Safety: The instructional prompt provided the topic of “working on height” with hazard recognition, hazard management, and safety compliance defined as the learning objectives of the training. GPT was instructed to incorporate specific training material from OSHA and NIOSH material to create the learning content. The training progression was also provided to GPT as follows: instruction on first learning objective → pause for user’s questions → assessment scenario for the learning objective → feedback on user performance → subsequent scenarios based on user performance → feedback → repeat for the next learning objective. Clarifying GPT’s Responsibility: GPT was tasked to introduce itself, create interactive and accurate safety training content, deliver training, answer user-asked questions, create safety scenarios and questions to evaluate user’s understanding, and create follow-up scenarios based on user’s performance until a satisfactory result is conceived.
Use Case #2: Automated Post-Accident Analysis	Contextualizing Safety: GPT was tasked to analyze an accident report to achieve three specific objectives that include extracting key information from accident reports, identifying the root cause and causal factors of accident, and proposing preventive measures. Clarifying GPT’s Responsibility: GPT’s responsibility was using its knowledge base, reasoning capacity, and relevant standards and guidelines to answer 9 key questions framed around the objectives listed above.
Use Case #3: Automated Generation of Safety Guidelines and Checklists	Contextualizing Safety: GPT was provided with a construction scenario that included specific task (arc welding), specific component of the building (roof framing of a steel structure building), and specific location (third floor roof). Clarifying GPT’s Responsibility: GPT was instructed to generate safety guidelines and checklists for the construction scenario using OSHA regulations and standards. Moreover, GPT was tasked to provide relevant OSHA standards to be considered for the work
Use Case #4: Real-time Safety Management Assistance	Contextualizing Safety: GPT was tasked to provide answers to diverse safety-related questions that are typically expected from safety assistants from three areas: (1) hazard recognition, (2) hazard evaluation, and (3) hazard control. Clarifying GPT’s Responsibility: GPT was tasked to provide detailed answers with references from OSHA standards and reliable safety resources and come up with follow-up questions to enhance user’s understanding of the topic.

. The instructional prompts were employed to clarify GPT’s responsibility in interacting with the data, specifying the type of processing it should undertake and outlining the expected outcomes for each use case. (See Appendix A for detailed descriptions of these instructional GPT prompts to create the four use cases).

As detailed in Appendix A and Table 2, prompts were designed to reflect realistic construction safety scenarios and the types of responsibilities GPT would be expected to assume in practice. Each use case assigned GPT a distinct role, such as safety trainer, safety manager, or assistant, to align with the specific task. Prompts were structured with neutral, context-relevant language and avoided suggestive phrasing to minimize response bias. Importantly, the tasks were designed to mirror the actual responsibilities, expected interactions, and intended outcomes associated with each professional role in a real-world construction safety setting. For Use Case #1 (Personalized Safety Training and Education), GPT was instructed to act as a certified safety trainer and conduct an interactive height safety training session, including small talk, OSHA-cited content delivery, question-and-answer interaction, and scenario-based knowledge assessment with adaptive difficulty. In Use Case #2 (Automated Post-Accident Analysis), GPT was prompted to analyze a detailed accident report and respond to structured questions regarding hazard identification, OSHA standards, injury details, and causal analysis. For Use Case #3 (Automated Generation of Guidelines and Checklists), GPT was asked to generate safety documentation based on a specific construction scenario while citing relevant OSHA regulations. In Use Case #4 (Real-time Safety Management Assistance), GPT played the role of a virtual safety assistant, beginning with conversational onboarding and user profiling, then offering support by answering user questions based on OSHA, NIOSH, and other credible safety sources, while proposing relevant follow-up queries to deepen understanding.

To ensure consistency and minimize bias, only the first GPT-generated response was used for evaluation. However, additional output generations were reviewed to confirm response stability and verify that outputs did not deviate significantly from the initial generation. This approach enabled a reliable and reproducible evaluation of GPT’s performance across all use cases while preserving the contextual integrity of the tasks.

2.3. Step 3: Safety Use Case Analysis

After defining the use cases for GPT using instructional prompts, each created use case was assessed using a series of User–GPT interactions.

Table 3. User–GPT interaction under each use case.

Use Cases	User–GPT Interactions
Use Case #1: Personalized Safety Training and Education	GPT Role: GPT was tasked to assume the role of an “expert construction safety trainer” and train a construction worker. User Role: user assumed the role of a new construction worker who has limited understanding of the safety topic that is being taught.
Use Case #2: Automated Post-Accident Analysis	GPT Role: GPT had the role of a construction accident investigator who analyzes the past accidents and develops guidelines and preventive measures for the future. User Role: not applicable.
Use Case #3: Automated Generation of Safety Guidelines and Checklists	GPT Role: GPT assumed the role of a safety expert and provided safety guidelines and checklists for the given construction scenario. User Role: user provides a specific construction settings to GPT and requests guidelines and checklists for the task.
Use Case #4: Real-time Safety Management Assistance	GPT Role: GPT assumed the role of a safety management assistant. User Role: user assumed the role of a construction manager and asked safety-related questions within three key areas of safety management.

provides examples of how such interactions were conducted between users and GPT under each use case. Appendix B also provides all these interactions with complete details.

The User–GPT interaction for each use case was analyzed to evaluate GPT’s performance, capabilities, and limitations. The following criteria were selected for the qualitative assessment of the User–GPT interaction based on the established literature on language model evaluation [81,82,83] and their relevance to the demands of safety-critical domains. The criteria include interactivity, evidence-based reasoning, critical analysis and extrapolation, and knowledge and comprehension. These reflect core competencies required to effectively carry out the tasks represented in each use case, such as delivering training, analyzing incidents, generating safety documentation, and providing real-time assistance. Together, they provide a structured framework for assessing GPT’s ability to support construction safety management in a manner that aligns with both domain expectations and model performance benchmarks.

2.3.1. Interactivity

One of the critical features of GPT is its inherent interactive capability to ensure a user-centric interaction. This includes effective dialogue, contextual understanding, and user-friendly responses. GPT’s proficiency in extrapolating from the content of the user’s input, or from instructions given as a system prompts, makes it an invaluable tool in safety management where prompt and accurate information exchange is crucial.

• GPT Assessment Procedure: to evaluate the interactive ability of GPT qualitatively, two researchers observed and assessed the responses of GPT based on following criteria.

  ○

  Comprehensive Language Processing: GPT’s ability to understand and interpret the nuances and variations in human language was evaluated, particularly its handling of complex queries and generation of coherent follow-up responses.

  ○

  Contextual Understanding: GPT’s capacity to maintain conversational continuity by remembering and integrating previous information and context into its subsequent responses was assessed.

  ○

  Human-like Responses: the degree to which GPT’s responses replicate human conversational patterns was examined, focusing on its ability to demonstrate empathy, personalization, and a natural flow in interactions.

• Assessed Use Cases: This assessment was conducted for use cases #1 (Personalized Safety Training and Education) and #4 (Real-time Safety Management Assistance), where user interaction is critical. Use case #2 (Automated Post-Accident Analysis) and #3 (Automated Generation of Safety Guidelines and Checklists) were excluded as no conversation takes place between the user and GPT in those use cases.

2.3.2. Evidence-Based Reasoning

Evidence-based reasoning refers to the ability of GPT to process information and generate conclusions that are logically sound and rooted in factual data. This capability is crucial for GPT, as it ensures that the decisions or recommendations made by it are reliable and grounded in evidence. These criteria are essential because they underpin the trustworthiness of the outputs GPT generates, which is particularly critical when it is applied in domains requiring high accuracy and reliability.

○

GPT Assessment Procedure: Evidence-based reasoning was assessed by comparing the generated information to (1) established regulations and standards, (2) relevant safety documents, and (3) a similar analysis conducted by a human expert, whenever applicable. The following criteria were used for assessment:

▪

Using Factual Information: GPT’s ability to generate answers based on reliable, factual information was evaluated, with a focus on minimizing hallucination.

▪

Detecting Incorrect Inputs: GPT’s capability to identify and flag user mistakes or incorrect input information was assessed.

▪

Extracting Correct Information from User’s Given Documents: GPT’s effectiveness in extracting accurate and relevant information from documents and instructions provided by the user was examined.

○

Assessed Use Cases: this assessment was conducted for all use cases as evidence-based reasoning is equally important for all application areas.

2.3.3. Critical Analysis and Extrapolation

Critical analysis of available information and data involves the capacity to analyze textual data, discern pertinent information, and make connections between different data points. This analytical capability is essential for LLMs like GPT to contextualize, prioritize, and infer based on the data in order to provide nuanced and informed responses or make decisions. The extrapolation aspect involves going a step further and using this analyzed data to make educated inferences or hypothesize about underlying reasons and potential future outcomes.

• GPT Assessment Procedure: GPT’s performance was qualitatively evaluated by two researchers and also benchmarked against the analysis conducted by human experts (from NIOSH and OSHA) for the same task, providing a comparative perspective on their ability to critically understand and apply safety-related information. Existing safety regulations and documents were also used for comparison. The following criteria were used for assessment:

  ○

  Depth of Analysis: GPT’s ability to thoroughly analyze textual data, identifying key information and relevant details that are crucial for understanding safety contexts, was assessed. This analysis was compared to the depth achieved by human experts from NIOSH and OSHA wherever applicable.

  ○

  Connection and Contextualization: GPT’s ability to link disparate pieces of information to form a cohesive understanding of safety scenarios was evaluated. The assessment includes determining whether GPT can contextualize safety information in a manner comparable to insights from existing safety regulations and documents.

  ○

  Inferencing and Extrapolation: GPT’s ability to infer and extrapolate from analyzed data to predict potential safety risks or recommend preventive measures was examined.

• Assessed Use Cases: This assessment was conducted for use cases #2 (Automated Post-Accident Analysis), #3 (Automated Generation of Safety Guidelines and Checklists), and #4 (Real-time Safety Management Assistance) where critical analysis and extrapolation from given information is important for GPT to deliver appropriate results.

2.3.4. Knowledge and Comprehension

The expansive knowledge base of GPT is one of its most notable attributes as it determines its effectiveness in interpreting and responding to complex queries. Knowledge refers to the breadth of information GPT can access, while comprehension involves the depth of understanding of this information and of user inputs. These elements are crucial for GPT to provide accurate, relevant, and contextually appropriate responses.

• GPT Assessment Procedure: GPT’s performance was qualitatively evaluated by comparing the generated information to (1) regulations and standards, (2) relevant safety documents, and (3) human responses whenever applicable. The following criteria were used for this assessment.

  ○

  Breadth of Knowledge: GPT’s ability to cover a wide range of topics was assessed by assessing its responses to diverse queries from different domains of safety management to evaluate its capacity to retrieve and accurately present information across a broad spectrum of subjects.

  ○

  Depth of Comprehension: GPT’s depth of understanding was evaluated, focusing on its ability to grasp complex concepts and contextual nuances in its responses.

• Assessed Use Cases: this assessment was conducted for all use cases.

3. Results of the Safety Use Case Analysis

This section discussed the analysis results for each use case based on the four criteria of (1) Interactivity, (2) Evidence-based Reasoning, (3) Critical Analysis and Extrapolation, and (4) Knowledge and Comprehension.

3.1. Analysis Results for Use Case #1 (Personalized Safety Training and Education)

In this use case, GPT provided safety training experience about fall-related hazards and working on heights. The training was provided through the following phases. Figure 2 depicts these interaction phases:

• Introduction: GPT introduced itself as a virtual trainer, establishing a welcoming and engaging tone similar to that of a human trainer. This initial interaction was designed to build rapport and ease the trainee into the learning environment, setting the stage for effective communication and learning.
• Training Content Delivery: GPT delivered educational content focused on fall-related hazards and safe practices for working at height. After presenting the material, GPT invited questions from the user to ensure clarity and understanding. This interaction allowed the trainee to engage actively with the content, moving the training forward based on the user’s responses or expressed understanding.
• Safety Scenario: GPT presented customized safety scenarios along with questions to evaluate the user’s understanding. These scenarios were crafted to simulate real-life situations, thereby testing the practical application of the learned safety measures and GPT’s ability to assess critical understanding.
• Provide Feedback: Based on the user’s answers to scenario questions, GPT provided personalized feedback, identifying areas of strength and those needing improvement. This step was crucial for evaluating GPT’s instructional effectiveness and enhancing the trainee’s learning experience. The process progressed with the user either asking further questions or confirming their understanding of the feedback.
• Adjust Scenario Complexity and Introduce New Scenarios: Depending on the user’s level of understanding, GPT adjusted the complexity of the safety scenarios. This iterative process of scenario presentation and feedback continued until the user consistently demonstrated satisfactory responses, ensuring a deep and practical understanding of the safety topics covered.
• Closure: The training session concluded with GPT summarizing the key points and confirming that the user had achieved a comprehensive understanding of the material, akin to the closure provided by a human trainer. This final interaction was designed to reinforce the training outcomes and provide the trainee with confidence in their new knowledge.

The analysis results for use case #1 (Personalized Safety Training and Education) are as follows:

Interactivity: During the training experiment, GPT demonstrated responsiveness and engagement in conversations with a trainee. Overall, GPT’s interactive ability was found to be acceptable for this use case based on the following observations.

• Comprehensive Language Processing: GPT exhibited the capability to process and understand the user’s questions correctly. It effectively comprehended the nuances and variations in the user’s language, successfully interpreting complex queries. For instance, when presented with a slightly jumbled question, “What’s my next step if I notice the safety harness seems wrecked just before I’m supposed to start working up high?”, GPT accurately deciphered the underlying safety concern despite the informal phrasing and potential ambiguity around “seems wrecked”. It responded appropriately: “Cease use immediately and report the issue to ensure a replacement harness is provided.” Additionally, GPT handled a follow-up question accurately: “And if there’s no replacement available immediately?”, to which it replied, “Utilize alternative safety measures such as a scaffold with guardrails or delay the task until the proper equipment is available.” These responses were relevant, but more importantly for this evaluation criteria, contextually appropriate, demonstrating a high level of linguistic accuracy and detail.
• Contextual Understanding: GPT’s ability to maintain continuity and context throughout the conversation was notably demonstrated during a session. At the beginning of the training, the user asked, “What steps should I take if I notice loose scaffolding at a worksite?” GPT advised to immediately cordon off the area and report the issue to the site supervisor for urgent repairs. Much later in the session, during the safety scenario phase of the training, the user inquired, “During my inspection, should I re-check previously reported issues like the scaffolding?” Remembering the earlier conversation, GPT responded, “Yes, verify that the previously loose scaffolding has been securely fixed as reported. Do not proceed with work in that area until all necessary repairs are confirmed.” This response shows how GPT successfully linked the initial identification of a hazard to subsequent safety verification procedures.
• Human-like Responses: The session began with GPT introducing itself and engaging in some initial small talk to set a friendly tone, saying, “Nice to meet you, Ali! How are you doing today? Are you ready to dive into some safety training for working at heights?” This helped to ease the trainee into the learning environment. During the training, GPT exhibited mixed performance in its human-like interactions. For instance, when a trainee expressed concern about feeling unsafe while working at heights, saying, “I always get nervous on a ladder, even with all the gear” GPT responded empathetically, “It’s completely normal to feel nervous, and it’s good that you’re cautious. Remember to double-check your gear before ascending and keep your focus on the task. I’m here to help you through this.” GPT also offered positive reinforcement, such as “Good job” or “Well done, that’s exactly right,” which effectively supported the trainee’s learning progress.

However, there were instances where GPT’s responses lacked the necessary human subtlety. For example, when delivering training material, other than the initial interaction which was “Absolutely, Ali! Let’s dive deeper into each area of safety training for working at heights,” the rest of the content was mere technical information delivery, same as any that can be read in a book or a research paper which lacks human subtlety. The overall performance of GPT in providing human-like responses was mixed. While it successfully demonstrated empathy and engagement in certain interactions, it occasionally failed to appropriately address trainee emotions or simplify complex information when needed. These moments highlighted areas where further refinement in GPT’s model could enhance its ability to more consistently mirror the nuanced and supportive communication style of an effective human trainer.

Evidence-based Reasoning: The evidence-based reasoning was evaluated using the feedback that GPT provided to the trainee and the training content generated by GPT. For this use case, GPT demonstrated acceptable evidence-based reasoning based on the following observations:

• Using Factual Information: The material and information that were provided to the trainee as part of the training aligned with OSHA guidelines and standards and established training practices. For each different part of training, GPT referenced the correct OSHA standard number; for example, for hazard management, it correctly used “OSHA Standard: 29 CFR 1926.502—Fall protection systems criteria and practices,” which is aligned with the context of the training.
• Detecting Incorrect Inputs: GPT demonstrated the ability to detect incorrect answers and provide targeted feedback based on user responses. Additionally, it could recognize partially correct answers, identify areas of misunderstanding, and guide the trainee toward the correct information. For instance, in the first scenario, when asked, “You are working on a construction site, and you notice that a coworker is about to use a ladder that appears to be damaged. The ladder has a split side rail and a couple of bent rungs. What steps would you take to address this situation?”, the user responded partially correctly by stating, “I will choose a new ladder to use.” While this response was on the right track, it missed key steps such as stopping the coworker, reporting the damage, and tagging the ladder. GPT emphasized these missing elements to improve the user’s understanding. To reinforce the learning, GPT introduced a follow-up scenario addressing the gaps identified in the previous response. It asked, “You are assigned to work on a rooftop that is 20 feet above the ground. The rooftop has no guardrails installed. What fall protection measures should you implement to ensure your safety while working on the rooftop?” This allowed the user to apply the knowledge gained from the earlier feedback and demonstrate improved understanding.
• Extracting Correct Information from User’s Given Documents: in this scenario, the user did not provide any document to GPT, so this criterion was not assessed.

Knowledge and Comprehension: During the training session, it was observed that GPT produced training content based on accurate OSHA standards and covered the necessary aspects for the given subject. However, certain issues were observed, as discussed in detail below.

• Breadth of Knowledge: The training session addressed all the predefined objectives provided to GPT. Specifically, GPT covered various safety management areas, such as hazard control, by identifying and discussing hazards like falls from ladders, unprotected edges and openings, and scaffold stability. Additionally, GPT addressed hazard management by outlining preventive measures, including the use of guardrails, personal fall arrest systems (PFASs), and proper ladder usage. This coverage of a wide range of safety management topics demonstrates the breadth of knowledge imparted during the training.
• Depth of Comprehension: In some instances, the training lacked depth, focusing primarily on more obvious content. This limitation was evident in most of the training material provided by GPT. While the training covered all the predefined objectives, the deeper insights that a human trainer could offer were notably absent. Even when a user requested a more in-depth explanation by asking, “Can you dive a bit deeper into each subject?”, although GPT’s subsequent response was more detailed, it still lacked the necessary depth and specificity. For example, when addressing scaffolding guardrails and planking, GPT provided the instruction, “Ensure scaffolds have proper guardrails and fully planked walkways,” which omitted crucial details related to hazard recognition in this context.

This gap could potentially be resolved by fine-tuning the model with more specific training data or by specifying the desired depth of coverage during the initial setup. Moreover, this issue suggests that GPT’s training delivery was largely user-driven. If the user requested more detailed information, GPT provided it; however, this approach may be problematic for users with limited prior knowledge of the subject. In such cases, critical information could be omitted without the user realizing it, which may result in gaps in understanding and safety practices.

In conclusion, use case #1 (Personalized Safety Training and Education) highlights both the strengths and limitations of GPT in the context of interactive safety training. While GPT demonstrated strong capabilities in processing complex and informal queries, maintaining contextual continuity, and providing evidence-based reasoning aligned with OSHA standards, its performance in interactivity was mixed. GPT effectively responded to user inputs with relevant and detailed safety guidelines, reflecting a broad knowledge base. However, two key limitations were observed. First, GPT’s ability to deliver consistently human-like responses was less reliable, especially when addressing user emotions or providing empathetic feedback. Second, its depth of comprehension was inconsistent, often requiring explicit prompts to offer more nuanced insights. These observations suggest that while GPT can serve as a valuable tool in safety training, further refinement is necessary to enhance its interactive capabilities and ensure it meets the needs of users with varying levels of expertise. Furthermore, integrating GPT into formal safety training programs presents practical challenges, including alignment with certified curriculum standards, ensuring content accuracy, and preventing over-reliance on AI-driven instruction. Ethical considerations such as maintaining instructor accountability and protecting learners from misinformation must be addressed before widespread adoption.

3.2. Analysis Results for Use Case #2 (Automated Post-Accident Analysis)

In this use case, GPT analyzed a past accident report of a real construction accident from NIOSH. The following steps are the interaction between the user and GPT. Figure 3 depicts these interaction phases:

• Input: in this step, user provided GPT with the accident report from NIOSH.
• Extracting Key Information from Accident Reports: This task evaluated GPT’s ability to understand and pull relevant details from complex documents. The accuracy and relevance of extracted information from NIOSH reports are crucial for understanding the context and specifics of each accident.
• Utilizing Relevant Standards and Guidelines: This task gauged GPT’s ability to effectively use its extensive knowledge base. By retrieving and applying relevant OSHA standards and guidelines, GPT showcases its capability to provide informed, contextually appropriate recommendations and references.
• Analyzing Causes and Proposing Preventive Measures: this task assessed GPT’s reasoning skills. By analyzing the causes of accidents and suggesting preventive strategies, GPT demonstrates its capacity to interpret data, understand underlying issues, and apply problem-solving techniques.

The analysis results for use case #2 (Automated Post-Accident Analysis) are as follows:

• Evidence-based Reasoning: GPT showcased acceptable evidence-based reasoning ability. This ability was measured by comparing the GPT responses to relevant OSHA regulations and analysis completed by the safety experts at NIOSH for the same accident report. Below are the details of assessment results.

  ○

  Using Factual Information: GPT demonstrated strong evidence-based reasoning ability by identifying specific hazards and linking them correctly to the corresponding OSHA standards. This could be seen for all of the hazards that GPT detected. Moreover, when GPT was tasked to provide the name of the injured person, it replied with “The victim’s name is not provided in the report”, which indicates that the responses created by GPT are not hallucinations and it is drawing its answers from given factual information. The responses indicated that GPT could be utilized as a valuable tool for automating the process of extracting necessary information from accident report databases and analyzing it to draw meaningful inferences. This saves time and resources, allowing professionals to focus on other critical tasks.

  ○

  Detecting Incorrect Inputs: this criterion did not apply to this use case as there was no user interaction.

  ○

  Extracting Correct Information from User’s Given Documents: Based on the responses provided by GPT, it is evident that the model is proficient in extracting relevant information from the given accident report. The responses to questions focused on information extraction (see questions 4 to 9 in Appendix A) demonstrated complete alignment with the details described in the accident report. For example, when GPT was asked to name the location of the accident, it replied, “the accident took place at a residential house where roof replacement work was being carried out”, or when the type and severity of injury were asked of GPT, it replied “the victim suffered severe head injuries, resulting in her remaining unconscious and ultimately leading to her death”, which completely aligned with the facts included in the provided accident report. These highlight the model’s capability to effectively understand and interpret the information contained within such reports.

• Critical Analysis and Extrapolation: This use case indicated that although there are areas where GPT had acceptable performance in critically analyzing the information, certain limitations need to be addressed to effectively use it as a tool for accident analysis. Below are detailed observations for this criterion.

  ○

  Depth of Analysis: The analysis of responses revealed differences between GPT’s observations and the human analysis conducted by NIOSH. While the human analysis did not identify any misuse of the forklift, GPT highlighted the potential hazard of using a homemade box to transport construction materials and equipment. This observation by GPT, which was overlooked by human experts, appears to be relevant to this accident. Another important observation was that GPT did not recognize the worker’s young age and potential lack of experience as a contributory factor to the accident. While NIOSH analysis cited this as a possible contributing element, it was inferred through extrapolation of facts rather than being directly evidenced in the report. This strict reliance on concrete evidence underscored GPT’s capability to reduce the impact of biases and subjective interpretations that might influence human analysis. Such a characteristic is invaluable, especially in situations where objective analysis is crucial. However, to fully account for the range of potential risks in construction safety, future improvements should consider incorporating controlled extrapolation into GPT, enabling it to identify indirect or nuanced factors while retaining its unbiased approach.

  ○

  Connection and Contextualization: This can be seen when GPT was tasked to provide reasons and analysis when determining hazards in the provided scenario. For example, when GPT recognizes inadequate training and supervision, the analysis implies the following: “The owners verbally reminded the male laborer and the victim to be careful while working, but there is no mention of formal training or supervision. OSHA standard number: 1926.21—Safety Training and Education. This standard mandates employer to provide proper training and supervision to employees to ensure their safety”. The analysis indicates that there was not adequate training based on the fact that the owner was not present when the accident happened and the fact that there was no indication of adequate training. Further, this hazard was connected to proper safety regulations and documents. This indicates that GPT is capable of establishing an understanding of the presented scenario and contextualizing safety information based on the existing regulations.

  ○

  Inferencing and Extrapolation: GPT provided a detailed analysis of each hazard, such as fall risks, inadequate training, unsafe equipment use, and ladder safety concerns. However, the depth of GPT’s analysis, while detailed, does not always capture the broader, systemic considerations of safety management. For example, the human response included broader, more generalized recommendations for preventive measures, such as developing a comprehensive safety and occupational health program. This reflects an understanding that preventing accidents often requires systemic changes and a holistic approach, rather than just addressing specific hazards. GPT, in contrast, provided more direct and specific solutions tied closely to the information presented in the report, which could have been seen by GPT’s recommendation to use fall hazard protection equipment or appropriate equipment.

• Knowledge and Comprehension: Evaluating the information referenced in GPT’s responses revealed that they were consistent with OSHA regulations and established standards of practice. This suggests that GPT was drawing information from accurate and factual sources, as indicated by the observations discussed below.

  ○

  Breadth of Knowledge: GPT demonstrated the ability to analyze the accident report and identify multiple issues contributing to the accident, as well as suggest relevant prevention strategies. Its knowledge was demonstrated through its ability to accurately reference and apply various sections of OSHA regulations, including fall protection, safety training and education, rigging equipment for material handling, and ladder safety, in the correct context of the details mentioned in the accident report. These references were consistent with OSHA standards. For example, GPT correctly identified the requirement for employers to provide fall protection systems and training for work at heights above 6 feet, aligning with OSHA requirements.

  ○

  Depth of Comprehension: GPT demonstrated a good depth of understanding in its responses, indicating that it comprehended the accident report thoroughly. For instance, it detected fall hazard as the primary cause of accident and noted, “The victim fell approximately 26 feet from the roof to the stone patio”. In its analysis, GPT stated, “The victim was not wearing any fall protection equipment while working on the roof, which exposed her to the risk of falling,” demonstrating its ability to consider the contextual nuances of the accident report. For preventive measures, GPT recommended, “Employers should provide and enforce the use of proper fall protection equipment, such as harnesses and guardrails, when working at heights. Adequate training on fall protection measures should also be provided,” which also suggests a deeper understanding of how an accident such as the one mentioned in the report could be prevented.

In use case #2 (Automated Post-Accident Analysis), the exploration of GPT’s capabilities and limitations within construction safety revealed mixed results. GPT demonstrated a robust ability to perform evidence-based reasoning, effectively extracting and analyzing information from accident reports. This capability suggests that GPT could serve as a valuable tool for automating data extraction and initial analysis, potentially saving considerable time and resources. However, the model exhibited limitations in critical analysis and extrapolation. While GPT identified hazards not noticed by human experts, it occasionally missed broader, systemic safety issues, such as the need for comprehensive safety programs that require more nuanced understanding. Moreover, GPT’s strict reliance on explicit information sometimes prevented it from recognizing less obvious, contributory factors such as experience or indirect risks. Addressing these limitations by enhancing GPT’s ability to perform controlled extrapolation could make it a more effective tool for comprehensive safety analysis. From a practical perspective, integrating GPT into existing incident analysis systems would require oversight protocols to verify the accuracy of extracted insights and to ensure compliance with legal and regulatory reporting standards. Additionally, ethical concerns such as attributing causation based on AI reasoning and the risk of overlooking contextual factors underscore the importance of human validation.

3.3. Analysis Results for Use Case #3 (Automated Generation of Safety Guidelines and Checklists)

In this use case, GPT generated safety guidelines and checklists for an arc-welding scenario that takes place at height. The following interaction between the user and GPT was used in this use case. Figure 4 demonstrates these interaction phases:

• Input: in this step, the user provided GPT with the construction activity scenario.
• Creating Guidelines: This task involved GPT generating guidelines that are specific to the construction scenario. This task assessed the knowledge, comprehension, and critical analysis abilities of GPT as it needed to understand the scenario, determine the accidents that can take place, and generate guidelines to prevent them.
• Creating Checklists: In this task, GPT had to generate a checklist based on the construction scenario and the guidelines it generated. This task targeted GPT’s ability to reason as it had to utilize the guidelines to generate appropriate questions for the checklist.
• Identifying relevant Standards: In this step, GPT was tasked to provide relevant standards from OSHA for the items in the guidelines and checklist that it generated. The objective was to assess knowledge and comprehension.

The qualitative analysis results for use case #3 (Automated Generation of Safety Guidelines and Checklists) are as follows:

• Evidence-based Reasoning: GPT’s response in generating safety guidelines and checklists for arc welding demonstrated robust evidence-based reasoning capabilities, particularly in its understanding and application of safety protocols and regulations.

  ○

  Using Factual Information: GPT aligned its recommendations with correct OSHA standards such as OSHA 1910.252 for welding, cutting, and brazing and 1926.501 for fall protection. It also used the information from the user input about work being at height and considered this fact in creating guidelines and checklists. For example, for fall protection, GPT mentions that “Implement appropriate fall protection measures, such as guardrails, safety harnesses, or safety nets.”

  ○

  Detecting Incorrect Inputs: this criterion did not apply to this use case as there was no interaction between the user and GPT.

  ○

  Extracting Correct Information from User’s Given Documents: in this scenario, the user did not provide any document to GPT, so this criterion was not assessed.

• Critical Analysis and Extrapolation: GPT’s critical analysis in this case was acceptable. However, certain issues need attention before GPT can be further used for this use case.

  ○

  Depth of Analysis: GPT listed safety measures that are needed for the specific task. For example, it mentioned proper PPE, fall protection, fire prevention, ventilation, and electrical safety. However, certain regulations, like the prohibition against immersing the electrode in water, were omitted. This omission is understandable as the generated output focused on general guidelines and checklists, which typically do not include such detailed instructions. These specifics are more appropriately conveyed in safety training sessions designed to address potential hazards in greater depth. Also, the guidelines appeared to only consider the static scenario, without accounting for potential changes in the work environment. A more critically analyzed response would consider dynamic factors, such as changing conditions or the presence of other workers in the vicinity, which could affect safety measures. Lastly, GPT’s response would benefit from a more comprehensive risk assessment, accounting for the interplay of multiple hazards specific to the task, such as electrical risks, working at heights, and exposure to welding fumes, requiring a deeper understanding of these combined threats.

  ○

  Connection and Contextualization: When comparing GPT-generated guidelines and checklists for arc welding with OSHA regulations, it became clear that GPT covered most of the relevant safety protocols. For example, in the safety checklist for arc welding, GPT correctly included specific PPE requirements, such as “Have all workers been provided with appropriate PPE, including welding helmets, gloves, and safety glasses?” Additionally, GPT captured the environmental context by including fall protection measures, considering the task’s execution at significant heights, even though no explicit fall hazards were mentioned in the user input. However, it did not necessarily prioritize guidelines based on the specific risks or likelihood, which is important to provide a more targeted set of guidelines. This issue could be resolved by tailoring the instructions given to GPT for the particular task.

  ○

  Inferencing and Extrapolation: The analysis of guidelines and checklists indicates that these were tailored to the specific hazards present in the scenario. For example, GPT highlighted relevant concerns such as, “Is the work area adequately ventilated to control welding fumes?” and “Have all workers been provided with appropriate PPE, including welding helmets, gloves, and safety glasses?” as well as “Have fall protection measures been implemented for workers working at heights?” These examples illustrate GPT’s ability to identify and address key hazards in the scenario, such as fall risks and fire-related dangers. Additionally, GPT developed checklists aimed at mitigating these workplace hazards, demonstrating its potential to enhance safety protocols through targeted hazard identification and prevention strategies.

• Knowledge and Comprehension: The response included accurate safety protocols for welding from OSHA, indicating a solid base of knowledge in construction safety. While GPT displayed a general comprehension of safety protocols, its depth of knowledge might be limited, particularly in dealing with the specific complexities of arc welding at height.

  ○

  Breadth of Knowledge: GPT was able to reference various sections of OSHA relevant to the given task, such as welding, cutting, and brazing, gas welding and cutting, and fall protection. This demonstrates its capacity to align generated guidelines with established safety standards. However, while GPT did acknowledge the importance of safety education and training within the guidelines and checklists it produced, it notably did not reference the specific OSHA standard number related to safety training and education among the citations it utilized. Neither did it refer to OSHA or any standard training for welding work or working at height, which a human safety manager would have. This indicates that retraining or fine-tuning GPT on specific safety documents may be necessary for future research.

  ○

  Depth of Comprehension: Unlike other use cases, GPT demonstrated a good depth of understanding in this use case. The detailed references to OSHA regulations in GPT’s output are strong indicators of its depth of comprehension, particularly in understanding the contextual nuances of the given scenario. For example, when addressing Hot Work Permits, GPT advised, “Obtain a hot work permit before commencing welding activities, as required by OSHA regulations,” and further emphasized, “Follow the permit’s specific requirements and ensure it is displayed at the work location.” This level of detail showcases GPT’s attention to critical safety procedures and its thorough understanding of the regulations, reinforcing its capability to generate precise and contextually appropriate guidelines.

To summarize, in use case #3 (Automated Generation of Safety Guidelines and Checklists), GPT exhibited significant capabilities but also revealed some limitations. The model effectively compiled safety protocols and checklists that align with established standards, demonstrating its utility in streamlining the preparation of essential safety documentation. This capability ensures that users can quickly generate reliable and standardized safety materials, enhancing overall safety compliance and awareness. However, the limitations can be seen in the model’s inability to incorporate highly specific or uncommon safety rules without direct input and its challenges in prioritizing safety measures based on specific risks or operational contexts. This underscores the need for human oversight in verifying the output. To optimize its application, future improvements should focus on enhancing the model’s understanding of contextual and risk-based prioritization in safety guideline formulation by fine-tuning it with more safety documents and validated guidelines and checklists created by human experts. To be implemented in practice, GPT would need to be integrated with project-specific data sources and reviewed by safety professionals to avoid applying generic checklists to complex site conditions. Operational risks include the potential misapplication of AI-generated guidelines without sufficient context, which could undermine safety protocols or introduce liability concerns.

3.4. Analysis Results for Use Case #4 (Real-Time Safety Management Assistance)

In this use case, GPT played the role of a safety management assistant who answered the user’s questions. The following steps show the interaction between the user and GPT in this use case. Figure 5 demonstrates these interaction phases:

• Input: in this step, a safety-related question was asked from GPT by the user.
• Response: in this step, GPT provided a response to the user’s query based on the given instructions.
• Suggesting Follow-up Questions to the User: After answering the user’s question, GPT suggested follow-up questions to help the user better understand the concept. The user could agree or disagree with these questions if they wanted to engage further on this topic.
• Response to the Follow-up Question: in this step, GPT provided the user with an answer to the follow-up question.
• The process repeated for the next question.

The analysis results for use case #4 (Real-time Safety Management Assistance) are as follows:

• Interactivity: GPT demonstrated acceptable interactivity as it was able to comprehend the user’s questions and keep the discussion warm and friendly with the user.

  ○

  Comprehensive Language Processing: Like other use cases, GPT exhibited a robust capacity to process and comprehend user queries, effectively navigating the nuances and variations in language. It demonstrated an ability to interpret complex questions, even when presented in an informal or ambiguous manner. For instance, when confronted with the somewhat disorganized query, “Give me a checklist for equipment inspection before they start?”, GPT accurately discerned the underlying safety concern despite the non-standard phrasing. It provided a detailed and relevant response, highlighting critical components such as harness and lifeline inspection, ladder safety, and tool and material storage while offering thorough explanations for each element. This answer is acceptable, given that the previous question targeted a scenario involving workers working at height. This illustrates GPT’s proficiency in interpreting varied linguistic inputs and delivering responses that align with essential safety considerations.

  ○

  Contextual Understanding: This ability was seen when GPT provided follow-up questions, and the user agreed to engage more on the topic. For example, in one of the questions, “Our multistoried construction site is surrounded by pedestrian traffic. Can you please analyze the fall hazard risks here?”, GPT provided the necessary answer but followed up with asking the user, “Would you like some tips on effective barricading or signage strategies for such environments?” This indicates that GPT is able to decipher the context behind the given question and suggest relevant follow-up questions to enhance the user’s understanding.

  ○

  Human-like Responses: GPT initiated the interaction with an introduction and informal conversation, creating a more personable and approachable atmosphere. For example, it began by saying, “Hello! I’m here to assist with all your construction safety needs. How has your day been at the site so far?” Additionally, GPT inquired about the user’s expertise in tailoring its responses accordingly. After the user mentioned they were a student conducting research on safety, GPT replied, “That’s a fascinating area of study! Your research must be providing valuable insights into improving safety protocols on construction sites.” This illustrates GPT’s effort to establish a more conversational and human-like interaction. However, when the conversation shifted to technical topics (for example, when the user asked safety-related questions), GPT’s responses were largely focused on delivering accurate information, lacking emotional engagement, and offering strictly technical answers. While this limits the conversational aspect, it can be addressed by refining the instructional prompts or fine-tuning the model with real conversation transcripts between safety managers and users to improve its ability to balance technical accuracy with human-like interaction.

• Evidence-based Reasoning: The responses generated by GPT in this use case were predominantly evidence-based, consistently referencing existing safety standards and guidelines. However, its knowledge base, while extensive, still lacks coverage of certain critical safety literature. Expanding GPT’s database to include more comprehensive safety resources would be beneficial for future research, enhancing its ability to provide more complete and accurate guidance in safety-related contexts. The following are more detailed observations.

  ○

  Using Factual Information: Throughout this scenario, GPT demonstrated acceptable performance using factual information. For instance, when addressing fall hazards associated with skylights, GPT identified the risk regarding the described situation. Similarly, in discussing the impact of rainy weather on fall hazards, the responses adhered to standard OSHA guidelines, reflecting a commitment to providing information that is accurate and practically applicable. Yet, there were instances where the conversation could benefit from more diverse sources, such as recent case studies or emerging research in construction safety. This would enhance the usefulness of the information provided and ensure that the most current and comprehensive safety practices are being conveyed.

  ○

  Detecting Incorrect Inputs: this was not evaluated for this use case as there was no instance where the user provided any information to GPT other than the questions they asked.

  ○

  Extracting Correct Information from User’s Given Documents: in this scenario, the user did not provide any document to GPT, so this criterion was not assessed.

• Critical Analysis and Extrapolation: GPT showcased good critical analysis ability and proved to be effective in answering safety-related questions.

  ○

  Depth of Analysis: GPT demonstrated an ability to go beyond surface-level responses by addressing the complexities involved in construction safety, indicating a strong depth of analysis. For example, when asked, “Our multistoried construction site is surrounded by pedestrian traffic. Can you please analyze the fall hazard risks here?”, GPT’s recommendations went beyond fall hazards and included using high-visibility barriers and clear signage, showing an understanding of practical safety management in complex environments. This suggests that GPT effectively recognizes and incorporates critical details from the query to provide a relevant and actionable solution to the safety concern.

  ○

  Connection and Contextualization: In this use case, GPT’s ability to connect and contextualize safety information is user-dependent, as it requires detailed or follow-up questions from the user for GPT to provide more in-depth responses. However, even with brief, single-line queries, GPT’s performance remains adequate. For example, when asked, “I have a team of six crews who I would like to send to install the skylight. What other equipment do I need to provide for safety?”, GPT offered a comprehensive list, including hard hats, safety glasses, non-slip safety boots, gloves, visibility vests, and communication devices. It also explained the relevance of each item to the specific task. Interestingly, it went beyond regulatory requirements, tried to contextualize the scenario, and added scenario-specific advice. For instance, it suggested communication devices “For clear communication among the crew, especially in noisy environments or over distances,” which aligns well with the nature of the task, where workers may be spread out across different areas. This demonstrates GPT’s ability to contextualize safety information in a manner that aligns with regulatory standards and is practical for the scenario at hand. For this use case, GPT showed competence in recognizing key situational factors, making its recommendations relevant and applicable to real-world construction environments.

  ○

  Inferencing and Extrapolation: Multiple questions on hazard identification were presented to evaluate GPT’s ability to infer and extrapolate safety hazards. For example, when asked, “I would like to install a skylight on top of the building. What safety hazards should I be concerned about?”, GPT accurately identified key risks such as fall hazards, roof load capacity, and concerns about skylight strength. These responses indicate that GPT can infer potential risks based on the contextual information provided. Additionally, GPT’s follow-up question, “Do you need details on specific fall protection measures for this task?” demonstrates its ability to anticipate further necessary details, highlighting an understanding of the need for preventive measures. It went on to recommend specific solutions, such as guardrails, personal fall arrest systems (PFASs), and skylight covers, explaining their relevance to the task. However, GPT’s extrapolation remains limited, as its responses focus primarily on the immediate hazards presented without fully considering more complex or indirect risks that could arise in a dynamic work environment. While GPT performs adequately in predicting and addressing obvious risks, expanding its capability to anticipate less apparent hazards or longer-term safety concerns would improve its overall performance. This suggests that while GPT can infer risks and propose preventive measures, its extrapolation could benefit from further refinement to capture a broader range of potential safety issues.

• Knowledge and Comprehension: GPT was able to draw from correct safety documents for the solutions it provided, indicating that it uses the training knowledge effectively. The following are some observations.

  ○

  Breadth of Knowledge: GPT was tested with multiple questions across various aspects of construction safety management, including hazard identification, evaluation, and control. These questions covered a diverse range of scenarios, such as working at heights, safety requirements during bad weather, equipment safety, and managing pedestrian traffic. In each case, GPT demonstrated its breadth of knowledge by providing accurate responses based on existing safety regulations and documents as mentioned before. This indicates GPT’s ability to draw from a wide range of safety management criteria and apply relevant guidelines across different situations, reflecting its comprehensive understanding of construction safety protocols.

  ○

  Depth of Comprehension: GPT demonstrated a certain level of depth in its responses, as seen in several instances. For example, when asked, “My crews need to transit between different heights frequently, what additional fall risks should be evaluated in this scenario?”, GPT identified the risks and also provided explanations. For instance, it mentioned, “Transition Points: Increased risk at points where workers move from ladders to roofs or scaffolds. Ensure secure handholds and stable platforms at these transition areas.” This response reflects a sound understanding of the specific risks associated with transitioning between different heights and is consistent with existing safety regulations and documents. However, the depth of comprehension remains somewhat limited, as GPT’s responses are constrained to surface-level issues and do not fully explore more complex or indirect hazards. For instance, while GPT recognized transition points, it did not address other nuanced factors, such as how weather conditions or worker fatigue might compound the risk at these points. Expanding its analysis to consider such variables would demonstrate a more advanced level of comprehension.

In conclusion, use case #4 (Real-time Safety Management Assistance) demonstrates GPT’s potential in providing safety assistance, with several notable strengths and some key limitations. GPT effectively processed and interpreted user queries, even when they were informal or ambiguous, and provided detailed, relevant safety advice. It was able to maintain a warm and friendly tone, making interactions approachable. However, the model’s interactivity was mixed; while it handled technical information well, it sometimes fell short in offering human-like responses, particularly in maintaining emotional engagement with the user when discussing technical topics. The model’s responses, grounded in existing safety standards, underscored its capacity for evidence-based reasoning. However, expanding its database to include more diverse and current sources could further enhance its effectiveness. While it demonstrated a degree of depth in analyzing safety concerns, its responses often lacked the ability to fully extrapolate or account for more complex or indirect hazards, such as dynamic environmental factors. Overall, while GPT can effectively utilize established knowledge, continuous updates and broader information access are essential for advancing its utility in construction safety. For real-world use, deploying GPT as an on-site assistant would require reliable data connectivity, integration with sensor or wearable systems, and clear protocols for human override. Ethical risks include the possibility of delayed or inappropriate responses in time-sensitive situations, reinforcing the need for human supervision in all safety-critical decision-making.

4. Conclusions

This study investigated the potential of LLMs, specifically GPT, in construction safety management. Given the critical need to improve safety on construction sites due to high injury rates, the study aimed to explore how GPT could contribute to reducing workplace accidents and enhancing safety performance. Structured use cases were employed to demonstrate GPT’s application in various safety management tasks and to identify its strengths and limitations. These use cases were selected based on areas where LLMs or their predecessors have been applied in safety or where future applications are promising. GPT’s performance was evaluated in four use cases: #1 Personalized Safety Training and Education, #2 Automated Post-Accident Analysis, #3 Automated Generation of Safety Guidelines and Checklists, and #4 Real-time Safety Management Assistance. For each use case, GPT was assessed using four fundamental evaluation criteria critical for its applications in safety management: interactivity, evidence-based reasoning, critical analysis and extrapolation, and knowledge and comprehension.

For use case #1 (Personalized Safety Training and Education), GPT demonstrated effective interactive communication and delivered safety training content aligned with OSHA guidelines. This suggests promising potential for using LLMs to scale personalized safety training, addressing a significant challenge in current training practices. However, the study found that the depth of the training content was somewhat limited. Therefore, it concludes that with more refined model training and fine-tuning to incorporate greater detail and nuanced topic coverage, LLMs could become highly effective tools for delivering comprehensive and personalized safety training.

For use case #2 (Automated Post-Accident Analysis), GPT showcased a strong ability to perform evidence-based reasoning, efficiently extracting and analyzing information from accident reports. It demonstrated an aptitude for identifying specific causal factors of accidents and recommending appropriate prevention strategies. Interestingly, it identified factors that were missed by human experts and strictly relied on the facts, and did not extrapolate. However, the model’s limitations became apparent in critical analysis and extrapolation. While GPT identified certain direct hazards and causal factors, it sometimes failed to recognize broader systemic issues, such as the need for comprehensive safety programs, and often missed subtler contributory factors like worker experience or indirect risks. Therefore, this study concludes that to fully leverage GPT’s capabilities for post-accident analysis, it is important to enhance its ability to conduct controlled extrapolation. However, GPT (as is) still holds significant potential for automating data extraction and initial analysis, which could lead to substantial time and resource savings. This analysis should, however, be supplemented with human oversight to ensure a thorough and nuanced understanding of accident causation.

For use case #3 (Automated Generation of Safety Guidelines and Checklists), GPT successfully produced guidelines that aligned with basic safety protocols. However, its capability to adapt these guidelines to construction sites’ dynamic and often unpredictable nature might be limited. This points to the importance of updating LLMs with real-time, site-specific data, thereby ensuring the relevance and applicability of the safety protocols in various construction scenarios. Therefore, the study concludes that while GPT can be used to generate basic and general-purpose guidelines and checklists, it needs to be integrated with systems that can provide real-time site information to leverage its capabilities to generate site-specific checklists and guidelines. In addition, GPT or LLMs, in general, should be fine-tuned with relevant safety literature to help prioritize hazards and their prevention while creating safety guidelines.

Use case #4 (Real-time Safety Management Assistance) demonstrated GPT’s proficiency in maintaining responsive and interactive dialogues. The responses were correct, in accordance with established guidelines and protocols, and relevant to the questions. Yet, the scope of safety solutions and advice offered by GPT was somewhat limited, indicating a need for broader and more current safety data integration into its training. This would enhance the depth and diversity of the safety management assistance provided by GPT. These results collectively suggest that while GPT has a significant potential role in construction safety management, its optimal utilization presently requires synergistic operation with human expertise. The current study’s findings advocate for future developments to focus on increasing the depth of GPT’s knowledge, its context-awareness, integration with real-time data, and adaptability to the dynamic construction environment.

Across the four use cases, GPT demonstrated promising utility in supporting construction safety tasks. It was effective in delivering personalized training aligned with OSHA standards, though the depth of content was limited. In accident analysis, GPT showed strong evidence-based reasoning and the ability to identify factual hazards, but it struggled with systemic insight and extrapolation. For guideline generation, it produced compliant safety checklists but lacked adaptability to dynamic, site-specific conditions. In real-time assistance, GPT maintained accurate and interactive communication but offered a limited range of solutions. These findings suggest that while GPT can support structured, rule-based safety tasks, its optimal performance requires integration with real-time data, contextual fine-tuning, and ongoing human oversight.

In translating these findings into practice, several integration challenges must be considered. These include ensuring compatibility between GPT and existing safety management systems, maintaining the reliability and relevance of underlying data, and aligning AI-generated outputs with current regulatory frameworks. Additionally, effective implementation would require user training to ensure that safety personnel interpret GPT responses appropriately and do not over-rely on automated outputs in safety-critical decisions. Addressing these challenges is essential for realizing the practical value of GPT in real-world construction safety management.

This study contributes to the growing body of research on the use of LLMs by systematically evaluating their performance across key safety tasks. By demonstrating GPT’s effectiveness in automating key safety tasks and identifying areas for improvement, the study provides a foundation for developing more efficient, scalable, and personalized safety solutions. Its findings could lead to safer construction environments, reduced accidents, and significant time and resource savings, ultimately driving innovation in safety protocols and management.

Study Limitations

While the study provided some valuable insights, there are certain limitations that need to be discussed. A primary constraint is the exclusive focus on GPT for evaluation. While GPT is amongst the most prominent and extensively used LLM models, a comprehensive assessment of other models is crucial for a more rounded understanding of the field. The results obtained from GPT might not wholly translate to other models, as each may have unique limitations and capabilities. Secondly, the approach of using structured use cases to assess the effectiveness of GPT presents a methodological limitation in itself. While this approach has allowed for a concentrated exploration of GPT’s capabilities in defined contexts, it inherently narrows the scope of the study. Relying solely on particular use cases may not capture the full spectrum of GPT’s potential applications in construction safety management. Future research should consider a wider range of use cases, offering a more comprehensive assessment of GPT’s utility across diverse scenarios. Moreover, while the study focused on qualitative analysis, it did not incorporate quantitative validation. This was intentional, as the qualitative findings were needed to first uncover how GPT performs in context-specific safety tasks. These findings now serve as a foundation for designing future quantitative evaluations with more appropriate performance metrics tailored to construction safety applications. The third limitation of this study is the development and reliance on a single scenario for each use case. This approach restricts the diversity of data and insights that could be garnered from varied scenarios. Different parameters, questions, and prompts in multiple scenarios could reveal a richer array of findings and highlight different facets of GPT’s capabilities and limitations. Another limitation is the absence of direct input from safety professionals across all use cases. While regulatory standards (e.g., OSHA) and expert analyses (e.g., NIOSH reports) were used for comparison, future studies should incorporate safety professionals’ feedback to validate and refine GPT’s outputs in each application area. Furthermore, while GPT shows promise in supporting safety-related tasks, its application in safety consultancy or evaluation raises important ethical considerations. These include the potential for misinformation, over-reliance on AI-generated content, and lack of accountability in decision-making. Given the high-risk nature of construction environments, GPT should be used strictly as an assistive tool under the supervision of qualified safety professionals. Final decisions must remain the responsibility of certified experts to ensure regulatory compliance, ethical responsibility, and worker safety. Future studies should aim to create and analyze multiple scenarios for each application area, thereby enhancing the depth and breadth of the research. However, it is important to recognize that the results obtained from the current approach still offer valuable, targeted insights into GPT’s application in construction safety management. Furthermore, to move beyond proof-of-concept, future research should include longitudinal studies and real-world implementation trials to evaluate the sustained utility, reliability, and impact of GPT-assisted systems in active construction environments.