2026 年 5 巻 2 号 論文ID: cr.26-002
Artificial intelligence (AI) is increasingly being introduced into workplaces, raising expectations for improved efficiency and safety while also generating new challenges for occupational safety and health (OSH). However, empirical evidence on how AI is adopted, perceived, and governed in everyday workplace settings remains limited, particularly from the perspective of workers. This descriptive, cross-sectional, web-based survey aimed to describe the current state of AI adoption in Japanese workplaces and to examine participants’ perceptions of AI-related OSH risks, organizational responses, and accountability related to AI use. The survey was conducted in June 2025 among 1000 employed adults registered with an online survey panel in Japan. AI adoption was still limited in this sample: 5.8% of participants reported that AI had been fully implemented in their organizations, and 16.6% reported partial adoption. In addition, 78.9% reported no work experience related to AI. Across AI-related OSH risk items, approximately 39%–42% selected “Not applicable/AI not used,” and responses among those for whom at least 1 risk item was applicable were frequently concentrated in the “Neither” category. Organizational measures to address AI-related OSH risks appeared unevenly implemented, although human final review of AI outputs was among the more frequently reported measures. Regarding the perceived need for preventive measures, 36.5%–44.6% of participants considered measures such as transparency and explainability, workplace applicability verification, clear allocation of responsibilities, and collaboration with OSH experts to be necessary to some extent or very necessary. Responsibility for AI-related decisions was most frequently attributed to organizations and senior management that decided to adopt AI, and 59.5% of participants indicated that legal reform mandating accountability for AI decisions and actions was necessary or conditionally necessary. These findings suggest that AI use in Japanese workplaces remains at an early stage for many participants, with limited direct experience and uncertainty regarding AI-related OSH risks and organizational responses. Responsible AI use in OSH may require workplace-level guidance to identify AI-related risks, preserve appropriate human oversight, and clarify responsibility and accountability across multiple stakeholders.
artificial intelligence
OSHoccupational safety and health
AI technologies are increasingly being introduced into workplaces across industries, offering new opportunities for hazard detection, automation, decision support, and efficiency improvements in OSH practices (Shah and Mishra 2024; Trivedi and Alqahtani 2024). At the same time, AI use in worker monitoring, algorithmic decision-making, task allocation, and decision support raises operational and human–AI interaction challenges that may affect safety, well-being, and work processes (European Agency for Safety and Health at Work 2023; Fiegler-Rudol et al. 2025).
Previous research and policy-oriented reports have increasingly addressed the potential implications of AI and digitalization for OSH. Foresight reports on digitalization and work have identified emerging risks related to human–machine interaction, algorithmic decision-making, and workplace monitoring, while also emphasizing the limited availability of empirical evidence at the workplace level (Stacey et al. 2018). Similarly, global policy reports note that although AI is expected to contribute to improvements in OSH, the actual implementation, governance, and risk management practices vary widely and remain insufficiently examined in real-world work settings (International Labour Organization 2025). Evidence on measurable improvements in worker health outcomes resulting from AI-enabled OSH tools also remains limited (Jetha et al. 2025).
Empirical studies have provided more concrete evidence on the OSH implications of algorithmic management, particularly in platform work. Studies of gig and platform workers have shown that algorithmic control can coexist with worker autonomy while also contributing to excessive working hours, low pay, sleep deprivation, psychosocial stress, reduced autonomy, and mental health concerns (Wood et al. 2019; Mbare, Perkiö, Koivusalo 2024; Cheng et al. 2024). In contrast, empirical evidence from non-platform workplaces remains more limited, although recent studies in logistics suggest that algorithmic management may be associated with psychological distress, musculoskeletal pain, and occupational accidents, and that its health implications depend on how such systems are implemented within organizational practices (Nilsson et al. 2025a; Nilsson et al. 2025b). Recent employer survey evidence also indicates that algorithmic management is increasingly relevant beyond platform work and raises concerns related to worker well-being, monitoring, accountability, and organizational governance (Milanez, Lemmens, and Ruggiu 2025). In addition, review studies on algorithmic management identify psychosocial risks as an emerging OSH challenge, while indicating that workplace-level evidence on governance structures and risk management practices is still underdeveloped (Bowdler et al. 2026).
Taken together, these findings point to a need for descriptive, workplace-based research that integrates AI adoption status, perceived OSH risks, organizational responses, and governance considerations within ordinary workplace contexts. This gap is particularly relevant in country-level research, where AI adoption may still be partial, experimental, or unevenly distributed across sectors and organizations. The purpose of this study was therefore to describe the current state of AI adoption in Japanese workplaces and to examine how workers perceive OSH risks, organizational responses, and responsibility and accountability related to AI use.
This study was designed as a descriptive, cross-sectional, web-based survey using a commercial online survey panel in Japan. The survey aimed to describe the current status of AI adoption in workplaces and to examine workers’ perceptions of OSH risks, organizational responses, and accountability related to AI use.
Because the study used a registered online survey panel rather than probability-based sampling, the findings should be interpreted as exploratory and descriptive and should not be regarded as statistically representative of all workers in Japan.
The present survey formed part of a broader FY2025 Health and Labour Sciences Research Grant project on management challenges arising from AI use from a legal perspective and legal issues under the current Industrial Safety and Health Act, as documented in the project report (Mishiba and Japan Association of Occupational Health Law 2025).
2-2. Survey procedureThe survey was conducted by Neo Marketing Inc., a commercial research company in Japan, using its registered online survey panel, “i-Research.” Invitation requests were sent via the survey platform and by email to registered panel members whose profile information matched the predefined eligibility criteria. Panel members were invited to participate voluntarily, and those who chose to cooperate proceeded to the web-based questionnaire.
No formal quota allocation or stratified sampling was applied in the present survey. Instead, invitations were distributed to registered panel members whose profile attributes matched the eligibility criteria. To obtain 1000 valid responses after data quality checks, invitations were sent to approximately 5000 eligible panel members, and responses were collected in excess of the target number. Data collection was conducted over a 4-day period from June 13 to June 16, 2025.
2-3. ParticipantsEligible participants were registered online panel members aged 20–69 years who resided in Japan and were engaged in 1 of the predefined occupational categories: company executives or officers, company employees or staff members, full-time employees, non-regular employees, individual business owners or self-employed workers, public servants, or teachers.
Individuals who were aged 19 years or younger, aged 70 years or older, resided outside Japan, or were students, homemakers, part-time or casual workers, unemployed individuals, retired individuals, self-employed professionals such as physicians or lawyers, or individuals in other non-eligible occupational categories were excluded at the screening stage. After screening and data quality checks, 1000 participants were included in the final analytic sample.
2-4. Participant flow and data quality controlA total of 4820 responses were initially collected. Among these, 1367 respondents completed the questionnaire and met the eligibility criteria. After data quality checks, 367 responses were excluded, and the final dataset delivered to the research team consisted of 1000 valid responses.
Data quality control and cleaning were conducted by the survey company before data delivery. Responses were flagged and excluded based on several checks: meaningless free-text entries, such as repeated single-character strings, blank spaces, or 1-character words; numerical free-text responses in the upper and lower 0.5% of the distribution; response times in the longest and shortest 1%; and inconsistencies between questionnaire responses and registered panel profile information regarding sex, age, and residence.
Because the survey was conducted for academic research, additional item-specific checks were applied. These included checks for responses inconsistent with routing logic, meaningless or irrelevant free-text responses, and comments about the questionnaire entered in required response fields. Duplicate-response detection was not included in the data cleaning procedures reported by the survey company.
2-5. Survey measuresThe questionnaire was developed based on preliminary interviews, relevant OSH literature, and a review of relevant international reports. Preliminary interviews on AI use and the work environment were conducted between April 10 and May 11, 2025, with 7 company personnel from 4 companies. The interviewees were involved in AI implementation, digital transformation, production technology, design solutions, civil engineering management, and production information management. One interview was conducted in person at a company headquarters, and 3 interviews were conducted online via Zoom. These interviews provided qualitative insights into current AI usage in workplace settings, technical challenges, organizational issues, and perceived psychological effects of AI use.
Drawing on these findings, the research team developed survey items through internal discussion. The development of the questionnaire was also informed by literature discussing OSH regulation in the era of Industry 4.0, including regulatory challenges associated with emerging technologies (Mishiba 2024). The items were further refined with reference to reports published by the UK Health and Safety Executive and the AI Safety Institute (Health and Safety Executive 2025; Bengio et al. 2025) and were adapted to the Japanese OSH context. Although no formal pilot test was conducted, the draft questionnaire was reviewed repeatedly within the multidisciplinary research team to assess clarity, relevance, and coverage of the survey domains. The questionnaire was finalized after multiple rounds of discussion within the research team.
The final questionnaire covered the following domains: participant and organizational characteristics; self-reported AI knowledge and AI-related work experience; AI adoption status in the participant’s organization; types of AI systems adopted or planned in the organization and main purposes of AI use; perceived OSH risks related to AI use; measures implemented to address such risks; perceived necessity of preventive measures; perceived effects of AI adoption on workplace safety and physical and mental health; psychosocial responses to AI adoption; attitudes toward delegating work-related decisions to AI; organizational AI-related risk assessment, guidelines, and workplace governance; perceived future task replacement by AI; perceived need for social policy responses to increased automation; perceived responsibility of stakeholders; perceived need for legal accountability for AI decisions and actions; and free-text comments on AI-related OSH issues and possible legal reforms.
The questionnaire included both items administered to all participants and items administered only to relevant subgroups according to routing logic. Items on participant and organizational characteristics, AI knowledge and experience, AI adoption status, perceived AI-related OSH risks, perceived necessity of preventive measures, psychosocial responses, attitudes toward AI decision-making, organizational AI-related risk assessment, guidelines, and workplace governance, future task replacement, social policy responses, stakeholder responsibility, legal accountability, and free-text comments were administered to all participants unless otherwise specified.
Items on the types of AI systems adopted or planned in the organization and the main purposes of AI use were administered only to participants who reported that AI had been fully or partially implemented in their organization. Items on measures implemented to address AI-related OSH risks were administered only to participants who selected a response other than “Not applicable/AI not used” for at least 1 AI-related risk item.
Items on perceived effects of AI adoption on workplace safety and physical and mental health were administered only to participants who reported full or partial AI adoption in their organization. Items on which stakeholders should bear legal accountability for AI decisions and actions were administered only to participants who indicated that legal reform mandating accountability for AI decisions and actions was necessary or conditionally necessary.
In each results table, the table note specified the denominator used to calculate percentages, clarifying whether percentages were based on all participants or on a relevant subgroup. Questionnaire domains, response formats, routing conditions, and denominators are summarized in Supplementary Appendix 1. The full questionnaire, including item wording and response options, is provided in Supplementary Appendix 2.
2-6. Ethical considerationsThis study did not undergo formal ethics committee review. The study was conducted as an anonymous, non-interventional, web-based survey of adult participants. Before proceeding to the questionnaire, participants were provided with information on the survey administrator, the purpose of data use, and the handling of survey information. The information stated that responses would be statistically processed and used in a form that would not identify individuals, and that private information would not be publicly disclosed. Participation was voluntary; invitation messages asked registered panel members to cooperate with the survey, and participants proceeded to the questionnaire only if they chose to participate.
The dataset delivered to the research team contained reassigned response identification numbers and did not include personally identifiable information. The collected information was analyzed in aggregate form and used only for the stated research purposes.
Table 1 summarizes the characteristics of the survey participants and their organizations. The mean age of the participants was 53.6 years (SD = 9.8; range = 21–69 years), and the majority was male (81.8%). In terms of employment status, most participants were full-time employees (62.0%), followed by individual business owners or self-employed workers (14.3%) and non-regular employees (11.3%). Regarding position within the organization, nearly half of the participants were non-managerial staff (49.2%), while 18.9% were line managers and 15.4% held top executive or senior management positions. Participants were distributed across organizations of varying sizes, with 25.3% working in organizations with 1–9 employees and 25.2% in organizations with 1000 or more employees. With respect to industry, services not elsewhere classified (26.6%), manufacturing (20.2%), and wholesale and retail trade (13.4%) accounted for the largest proportions.
| Characteristic/Category | n | % |
|---|---|---|
| Gender | ||
| Male | 818 | 81.8 |
| Female | 182 | 18.2 |
| Age, years | ||
| Mean (SD): 53.6 (9.8), range: 21–69 | ||
| Employment status/occupational category | ||
| Company executive, officer, or organization officer | 75 | 7.5 |
| Full-time employee | 620 | 62.0 |
| Non-regular employee | 113 | 11.3 |
| Individual business owner/self-employed | 143 | 14.3 |
| Public Servant (non-teaching) | 30 | 3.0 |
| Teacher | 19 | 1.9 |
| Position within the organization | ||
| Top executive | 93 | 9.3 |
| Senior management | 61 | 6.1 |
| Line manager | 189 | 18.9 |
| Non-managerial staff | 492 | 49.2 |
| Not applicable | 165 | 16.5 |
| Primary department/division | ||
| Corporate planning/strategy | 101 | 10.1 |
| Human resources | 32 | 3.2 |
| General affairs/administration | 136 | 13.6 |
| Sales | 229 | 22.9 |
| Production management | 98 | 9.8 |
| Occupational safety and health | 9 | 0.9 |
| Occupational health (industrial physician, occupational health nurse, etc.) | 8 | 0.8 |
| Other | 387 | 38.7 |
| Number of employees (including part-time workers) | ||
| 1–9 employees | 253 | 25.3 |
| 10–49 employees | 147 | 14.7 |
| 50–99 employees | 93 | 9.3 |
| 100–299 employees | 134 | 13.4 |
| 300–999 employees | 121 | 12.1 |
| 1000 or more employees | 252 | 25.2 |
| Industry | ||
| Manufacturing (e.g., food products, chemicals, electrical machinery) | 202 | 20.2 |
| Construction (e.g., general construction, equipment installation) | 72 | 7.2 |
| Information and communications | 79 | 7.9 |
| Wholesale and retail trade | 134 | 13.4 |
| Health care and social services | 65 | 6.5 |
| Education and learning support services | 55 | 5.5 |
| Services (not elsewhere classified) | 266 | 26.6 |
| Other | 127 | 12.7 |
Percentages are based on all participants (n = 1000). Age is presented as mean (SD) and range. Percentages may not sum to 100.0 because of rounding.
Table 2 presents participants’ self-reported levels of AI knowledge and work experience. Most participants reported familiarity with basic AI-related terms such as generative AI and machine learning (74.3%), while 12.3% indicated that they were not familiar with what AI refers to. A smaller proportion reported understanding advanced AI concepts and techniques (9.7%), and only 3.7% reported being able to design, develop, and operate AI systems.
| Variable/Category | n | % |
|---|---|---|
| Self-reported AI knowledge | ||
| Not familiar with what AI refers to | 123 | 12.3 |
| Familiar with basic AI-related terms (e.g., generative AI, machine learning) | 743 | 74.3 |
| Understand advanced AI concepts and techniques (e.g., supervised learning, reinforcement learning) | 97 | 9.7 |
| Able to design, develop, and operate AI systems (e.g., implementation, tuning, and deployment) | 37 | 3.7 |
| AI-related work experience | ||
| No work experience related to AI | 789 | 78.9 |
| Have used AI in work tasks (e.g., analytical support) | 139 | 13.9 |
| Have been involved in the selection or implementation of AI systems | 45 | 4.5 |
| Involved in the design, development, or operation of AI systems | 27 | 2.7 |
| AI adoption status in the organization | ||
| AI has been adopted and is fully implemented | 58 | 5.8 |
| AI has been partially adopted (pilot or limited use) | 166 | 16.6 |
| Considering the adoption of AI | 83 | 8.3 |
| No AI adoption and no plans for adoption | 484 | 48.4 |
| Not sure | 209 | 20.9 |
Percentages are based on all participants (n = 1000).
AI, artificial intelligence
With respect to AI-related work experience, the majority of participants reported no work experience related to AI (78.9%). A total of 13.9% reported having used AI in work tasks, such as analytical support, while 4.5% had been involved in the selection or implementation of AI systems. Only 2.7% reported involvement in the design, development, or operation of AI systems.
Regarding the status of AI adoption in participants’ organizations, only 5.8% reported that AI had been fully implemented, while 16.6% reported partial adoption on a pilot or limited basis. A total of 8.3% indicated that their organizations were considering the adoption of AI. Nearly half of the participants reported that their organizations had not adopted AI and had no plans for adoption (48.4%), and 20.9% reported that they were not sure about the AI adoption status of their organizations.
3-3. Perceived AI-related OSH risksTable 3 summarizes participants’ perceptions of OSH risks in the workplace. Across all items, approximately 39%–42% of participants selected “Not applicable/AI not used,” indicating that the relevant AI was not used in their workplaces or that the item was not applicable to their current workplace situation. Among participants who selected a response other than “Not applicable/AI not used,” responses were most frequently concentrated in the “Neither” category. Overall, “Do not feel much” was the next most frequent response category, although “Feel somewhat strongly” was relatively more frequent for some items. Relatively higher proportions of participants reported perceiving risks related to lack of explainability in AI decision-making, reduced human attentiveness due to over-reliance on AI, and bias and inaccuracy in AI decisions.
| AI-related OSH risk | Not applicable/AI not used | Do not feel at all | Do not feel much | Neither | Feel somewhat strongly | Feel very strongly |
|---|---|---|---|---|---|---|
| Work disruption due to false alarms or false detections | 40.8 | 5.1 | 12.9 | 33.4 | 5.7 | 2.1 |
| Risk of AI making incorrect judgments about safety procedures | 41.7 | 3.4 | 12.6 | 31.0 | 8.8 | 2.5 |
| Insufficient alignment with work environment data (e.g., temperature, vibration) | 40.5 | 5.2 | 13.1 | 30.3 | 7.1 | 3.8 |
| Lack of explainability in AI decision-making (black-box problem) | 39.6 | 5.1 | 10.4 | 30.2 | 11.7 | 3.0 |
| Psychological stress caused by increased AI-based monitoring | 41.6 | 6.9 | 12.0 | 29.9 | 6.9 | 2.7 |
| Reduced human attentiveness due to over-reliance on AI | 38.9 | 4.9 | 10.9 | 29.0 | 12.1 | 4.2 |
| Tendency to neglect human verification due to overtrust in AI | 41.6 | 6.9 | 12.0 | 29.9 | 6.9 | 2.7 |
| Bias and inaccuracy in AI decisions | 38.7 | 4.0 | 9.2 | 31.3 | 13.1 | 3.7 |
Values are percentages. Percentages are based on all participants (n = 1000).
AI, artificial intelligence; OSH, occupational safety and health
Table 4 summarizes organizational measures implemented to address AI-related OSH risks and participants’ perceptions of the necessity of preventive measures. Panel A presents measures implemented among participants who selected a response other than “Not applicable/AI not used” for at least 1 of the AI-related OSH risk items shown in Table 3 (n = 680). Across all measures, a large proportion of participants selected “Neither” (44.7%–51.0%), indicating that the implementation status of risk mitigation measures was often unclear.
| Panel A. Measures implemented to address AI-related OSH risks | |||||
|---|---|---|---|---|---|
| Measure | Not implemented at all | Not implemented much | Neither | Implemented to some extent | Fully implemented |
| Ensuring human final review of AI outputs | 6.2 | 8.7 | 44.7 | 27.8 | 12.6 |
| Configuration and maintenance by specialized staff | 10.3 | 8.7 | 51.0 | 21.2 | 8.8 |
| Education and training for on-site workers | 10.6 | 10.3 | 48.2 | 21.3 | 9.6 |
| Defining tasks in which decisions are not delegated to AI | 9.0 | 8.8 | 50.0 | 20.6 | 11.6 |
| Establishing response procedures for risk incidents | 10.4 | 9.9 | 47.9 | 21.8 | 10.0 |
| Audits or reviews conducted by external experts | 14.6 | 13.5 | 49.9 | 14.0 | 8.1 |
| Panel B. Necessity of preventive measures for AI-related OSH incidents | |||||
| Measure | Not necessary at all | Not very necessary | Neither | Necessary to some extent | Very necessary |
| Transparency and explainability of AI decision-making | 13.2 | 8.3 | 42.0 | 25.1 | 11.4 |
| Verification of workplace applicability (e.g., trials or pilot testing) | 12.5 | 6.5 | 42.0 | 24.2 | 14.8 |
| Clear allocation of responsibilities and rule-setting | 12.4 | 5.7 | 37.3 | 27.3 | 17.3 |
| Collaboration with OSH experts | 12.1 | 6.2 | 41.1 | 25.8 | 14.8 |
Values are percentages. For Panel A, percentages are based on participants who selected a response other than “Not applicable/AI not used” for at least 1 AI-related OSH risk item in Table 3 (n = 680). For Panel B, percentages are based on all participants (n = 1000).
AI, artificial intelligence; OSH, occupational safety and health
Regarding specific measures, ensuring human final review of AI outputs was reported as being implemented to some extent or fully implemented by 40.4% of participants (27.8% and 12.6%, respectively). Similarly, defining tasks in which decisions are not delegated to AI was reported as being implemented to some extent or fully implemented by 32.2% of participants (20.6% and 11.6%, respectively). For the remaining measures, audits or reviews conducted by external experts showed the lowest level of implementation, with 22.1% of participants reporting implementation to some extent or full implementation.
Panel B summarizes perceptions of the necessity of measures to prevent AI-related OSH incidents among all participants (n = 1000). For all 4 items, the largest proportion of participants selected “Neither,” ranging from 37.3% to 42.0%. Nevertheless, 36.5%–44.6% of participants indicated that these measures were necessary to some extent or very necessary. In particular, clear allocation of responsibilities and rule-setting was perceived as necessary to some extent or very necessary by 44.6% of participants, followed by collaboration with OSH experts (40.6%), verification of workplace applicability (39.0%), and transparency and explainability of AI decision-making (36.5%).
3-5. Responsibility and legal accountabilityTable 5 summarizes participants’ perceptions of stakeholder responsibility and legal accountability for AI decisions and actions. Panel A shows participants’ views on the degree of responsibility that various stakeholders should bear in AI-enabled workplaces. Across all stakeholder categories, responses were most frequently concentrated in the “Neither” category (42.4%–46.7%); however, 43.7%–50.4% of participants indicated that these stakeholders should bear some or very heavy responsibility. Organizations and senior management that decided to adopt AI were most frequently perceived as bearing responsibility, with 50.4% of participants indicating some or very heavy responsibility, followed by providers of training data for AI, primarily on the manufacturer side (49.1%). On-site supervisors who actually used AI (45.3%), developers and programmers (44.8%), and providers of learning data for AI, primarily on the user side (43.7%), were also perceived as bearing responsibility by more than 40% of participants.
| Panel A. Responsibility of stakeholders in AI-enabled workplaces | |||||
|---|---|---|---|---|---|
| Stakeholder | No responsibility at all | Little responsibility | Neither | Some responsibility | Very heavy responsibility |
| Developers and programmers | 4.0 | 5.8 | 45.4 | 31.2 | 13.6 |
| Providers of training data for AI (primarily on the manufacturer side) | 3.4 | 4.8 | 42.7 | 33.0 | 16.1 |
| Providers of learning data for AI (primarily on the user side) | 4.1 | 5.5 | 46.7 | 31.2 | 12.5 |
| Organizations and senior management that decided to adopt AI | 3.1 | 4.1 | 42.4 | 32.7 | 17.7 |
| On-site supervisors who actually used AI | 3.0 | 6.1 | 45.6 | 35.3 | 10.0 |
| Panel B. Perceived need for legal reform to mandate accountability for AI decisions and actions | |||||
| Response | n | % | |||
| Yes, it is necessary | 228 | 22.8 | |||
| Necessary depending on circumstances | 367 | 36.7 | |||
| Not necessary | 122 | 12.2 | |||
| Not sure | 283 | 28.3 | |||
| Panel C. Stakeholders who should bear legal accountability for AI decisions and actions | |||||
| Stakeholder | n | % | |||
| Developers and programmers | 222 | 37.3 | |||
| Providers of training data for AI (primarily on the manufacturer side) | 200 | 33.6 | |||
| Providers of learning data for AI (primarily on the user side) | 181 | 30.4 | |||
| Organizations and senior management that decided to adopt AI | 249 | 41.8 | |||
| On-site supervisors who actually used AI | 139 | 23.4 | |||
| Other | 3 | 0.5 | |||
| Cannot be specified/joint responsibility is appropriate | 161 | 27.1 | |||
Values in Panel A are percentages. For Panels B and C, values are n and percentages. For Panels A and B, percentages are based on all participants (n = 1000). For Panel C, multiple responses were allowed, and percentages are based on participants who selected “Yes, it is necessary” or “Necessary depending on circumstances” in Panel B (n = 595). In this study, “training data for AI” referred primarily to data provided on the manufacturer or developer side, whereas “learning data for AI” referred primarily to data provided on the user side.
AI, artificial intelligence
Panel B shows participants’ perceived need for legal reform to mandate accountability for AI decisions and actions. A total of 22.8% of participants indicated that such legal reform was necessary, and 36.7% indicated that it was necessary depending on the circumstances. In contrast, 12.2% indicated that such reform was not necessary, and 28.3% were not sure.
Panel C shows the stakeholders who participants thought should bear accountability for AI decisions and actions among those who selected “Yes, it is necessary” or “Necessary depending on circumstances” in Panel B (n = 595). The most frequently selected stakeholder was organizations and senior management that decided to adopt AI (41.8%), followed by developers and programmers (37.3%), providers of training data for AI (33.6%), providers of learning data for AI (30.4%), and on-site supervisors who actually used AI (23.4%). In addition, 27.1% indicated that accountability could not be specified or that joint responsibility would be appropriate.
3-6. Supplementary findingsAdditional findings on the types and main purposes of AI use, perceived effects of AI adoption on workplace safety and physical and mental health, psychosocial responses to AI adoption, attitudes toward AI decision-making, perceived future task replacement by AI, and social policy responses to increased automation are presented in Supplementary Tables S1–S6.
This descriptive, cross-sectional web-based survey examined AI adoption, AI-related OSH risk perceptions, organizational responses, and responsibility and accountability related to AI use among registered online panel participants in Japan. Several principal findings emerged. First, AI adoption and direct work experience with AI were still limited in this sample. Only 5.8% of participants reported that AI had been fully implemented in their organizations, and 16.6% reported partial adoption, while 78.9% reported no work experience related to AI.
Second, perceptions of AI-related OSH risks were characterized by frequent selection of “Not applicable/AI not used” and “Neither.” Across the AI-related OSH risk items, approximately 39%–42% of participants selected “Not applicable/AI not used,” suggesting that many participants did not regard the listed AI-related risks as applicable to their current workplace situation. Among participants who selected a response other than “Not applicable/AI not used,” responses were most frequently concentrated in the “Neither” category.
Third, organizational measures to address AI-related OSH risks appeared to be unevenly implemented. Among participants who selected a response other than “Not applicable/AI not used” for at least 1 AI-related OSH risk item, “Neither” responses were frequent for all risk mitigation measures, suggesting that the implementation status of such measures was often unclear. Human final review of AI outputs and defining tasks in which decisions are not delegated to AI were the most frequently reported measures, whereas audits or reviews by external experts showed the lowest level of implementation.
Fourth, participants expressed moderate support for preventive measures and accountability mechanisms. Although “Neither” was again the most frequent response for preventive needs, 36.5%–44.6% of participants indicated that measures such as clear allocation of responsibilities, workplace applicability verification, transparency and explainability, and collaboration with OSH experts were necessary to some extent or very necessary. In addition, responsibility for AI-related decisions was most frequently attributed to organizations and senior management that decided to adopt AI, and 59.5% of participants indicated that legal reform mandating accountability for AI decisions and actions was necessary or necessary depending on the circumstances.
4-2. Limited AI adoption and workplace exposureThe limited level of AI adoption and direct AI-related work experience provides an important context for interpreting participants’ risk perceptions and responses to governance-related items. In this sample, only 22.4% of participants reported that AI had been fully or partially implemented in their organizations, and nearly 80% reported no work experience related to AI. Although many participants reported familiarity with basic AI-related terms, relatively few reported advanced knowledge or direct involvement in AI system selection, implementation, design, development, or operation. These results suggest that, for many participants, AI was not yet embedded in everyday work processes.
This limited workplace exposure may partly explain the high proportions of “Not applicable/AI not used,” “Neither,” and “Not sure” responses across several items. Such responses should not be interpreted simply as a lack of concern or absence of risk perception. Rather, they may reflect limited opportunities to observe AI use in practice, uncertainty about how AI-related risks might emerge in specific workplace settings, or difficulty evaluating technologies that are still at an early stage of implementation.
Among participants whose organizations had fully or partially adopted AI, the supplementary findings indicate that AI was used or planned primarily for operational efficiency, data analysis and prediction, and generative AI, whereas safety management and health management were less frequently reported purposes of AI use. This pattern suggests that workplace AI adoption may currently be driven more by productivity, efficiency, and information-processing goals than by explicit OSH objectives, which may have implications for the timing and development of AI-related OSH governance.
4-3. AI-related OSH risk perceptions and organizational preparednessThe findings on AI-related OSH risk perceptions should be interpreted in the context of limited workplace AI exposure. Across the risk items, many participants selected “Not applicable/AI not used” or “Neither,” suggesting that AI-related OSH risks were not yet clearly recognized or directly experienced in many workplaces. However, this pattern should not be interpreted as evidence that such risks are absent. Rather, it may indicate that AI-related OSH risks remain difficult for workers to evaluate because AI systems are not yet widely embedded in work processes, and because concrete organizational procedures for identifying and managing such risks may still be underdeveloped.
Among the listed risks, relatively higher proportions of participants reported concerns regarding lack of explainability in AI decision-making, reduced human attentiveness due to over-reliance on AI, and bias and inaccuracy in AI decisions. These concerns align with commonly discussed issues in AI risk governance, including transparency, explainability, human oversight, accuracy, robustness, and bias mitigation (Organisation for Economic Co-operation and Development 2024; European Union 2024). In OSH contexts, these issues are particularly important because AI outputs may affect safety-related judgments, task allocation, monitoring, and intervention decisions.
The findings on organizational measures further suggest that workplace preparedness for AI-related OSH risks remains uneven. Among participants who selected a response other than “Not applicable/AI not used” for at least 1 AI-related OSH risk item, “Neither” responses were frequent across all risk mitigation measures, indicating that the implementation status of such measures may not have been clear to many participants. Measures involving human oversight, such as human final review of AI outputs and defining tasks in which decisions are not delegated to AI, were reported more frequently than audits or reviews by external experts. This pattern suggests that organizations may rely first on internal human control mechanisms, while more formalized or externally reviewed governance mechanisms remain less developed.
At the same time, the findings suggest some recognition of the need for preventive measures. More than one-third of participants considered transparency and explainability, workplace applicability verification, clear allocation of responsibilities and rule-setting, and collaboration with OSH experts to be necessary to some extent or very necessary. These findings suggest that even when direct experience with AI-related OSH risks is limited, participants recognize the potential importance of procedural safeguards and expert involvement when AI is introduced into workplace settings.
4-4. Responsibility and legal accountability for AI useThe findings on responsibility and legal accountability indicate that participants tended to attribute responsibility for AI use to multiple stakeholders rather than to a single actor. Organizations and senior management that decided to adopt AI were most frequently perceived as bearing responsibility, followed by providers of training data for AI, primarily on the manufacturer side, on-site supervisors who actually used AI, developers and programmers, and providers of learning data for AI, primarily on the user side. This pattern suggests that participants may view AI-related workplace responsibility as extending across both the development and implementation stages, including organizational decision-making, system development, manufacturer-side data provision, user-side data provision, and workplace-level use.
The relatively high attribution of responsibility to organizations and senior management is particularly important in the OSH context. Even when AI systems are developed or supplied by external vendors, the decision to introduce AI into a workplace, determine its scope of use, and integrate it into work processes is generally made at the organizational level. From this perspective, responsibility for AI-related OSH risks may need to be understood not only as a technical issue concerning system developers but also as an organizational governance issue involving employers and senior decision-makers.
At the same time, participants also attributed responsibility to developers and programmers, providers of training data for AI on the manufacturer side, providers of learning data for AI on the user side, and on-site supervisors. This pattern is consistent with lifecycle-oriented approaches to AI accountability, in which risks and responsibilities may arise at multiple stages, including system design and development, data provision, deployment decisions, and actual use in workplace contexts (Organisation for Economic Co-operation and Development 2023).
Participants’ responses regarding legal accountability also suggest support for some form of legal or institutional framework. Overall, 59.5% of participants indicated that legal reform mandating accountability for AI decisions and actions was necessary or necessary depending on the circumstances. Among these participants, organizations and senior management were the most frequently selected stakeholders who should bear accountability. Developers and programmers, providers of training data for AI, and providers of learning data for AI were also selected by 30.4%–37.3% of participants, while on-site supervisors were selected by 23.4%. In addition, 27.1% indicated that accountability could not be specified or that joint responsibility would be appropriate. These findings suggest that, in workplace AI use, participants may expect accountability frameworks that recognize both organizational responsibility and shared responsibility across multiple actors.
4-5. Implications for OSH governance and policyThese findings have several implications for OSH governance and policy in workplaces where AI is introduced, used, or considered for future use. First, the results suggest that many workplaces may still be at a stage where AI-related OSH risks are difficult to recognize concretely. The frequent selection of “Not applicable/AI not used” and “Neither” indicates that, for many participants, AI systems were either not yet part of their work environment or their potential OSH implications were not yet sufficiently visible. Therefore, rather than assuming that AI-related OSH risks are already well understood at the workplace level, guidance may be needed to help organizations first identify where AI systems are used, what types of decisions or tasks they affect, and what kinds of safety, health, and psychosocial risks could plausibly arise.
Second, AI-related OSH governance should not be limited to technical validation of AI systems. In this study, 36.5%–44.6% of participants considered transparency and explainability, workplace applicability verification, clear allocation of responsibilities and rule-setting, and collaboration with OSH experts to be necessary to some extent or very necessary. These elements are consistent with broader AI governance principles emphasizing trustworthy AI, transparency and explainability, human oversight, accuracy, robustness, and accountability, as reflected in the OECD AI Principles and the EU Artificial Intelligence Act provisions on high-risk AI systems (Organisation for Economic Co-operation and Development 2024; European Union 2024). In the OSH context, this suggests that AI governance should combine technical assessment with organizational procedures, human oversight, worker communication, and expert involvement.
Third, the results point to the importance of preserving human involvement in high-stakes workplace decisions. Supplementary findings on attitudes toward AI decision-making also suggested a preference among participants for human final judgment in safety- and health-related decisions, human resource–related decisions, and operational control of autonomous equipment or systems. This pattern is consistent with the broader finding that human final review of AI outputs was among the more frequently reported organizational measures. In OSH policy terms, this suggests that rules or guidelines may need to specify which decisions can be supported by AI, which decisions require human confirmation, and which decisions should not be delegated to AI.
Finally, building on these findings, AI-related OSH governance may require clarification of accountability across multiple actors. Participants attributed responsibility not only to developers and providers of training or learning data for AI, but also to organizations and senior management that decided to adopt AI and to on-site supervisors who actually used AI. This suggests that policy discussions may need to consider how responsibilities are distributed among developers, vendors, employers, managers, workers, and other stakeholders across the AI lifecycle (Organisation for Economic Co-operation and Development 2023). Such clarification may be particularly important when AI outputs influence safety-related judgments, monitoring, task allocation, or employment-related decisions. These findings do not directly indicate specific legislative amendments, but they highlight areas in which existing OSH governance frameworks may need to be adapted to address AI-supported decision-making and shared accountability in the workplace.
4-6. Contributions and limitationsThis study makes several contributions. It provides descriptive empirical evidence on AI adoption, AI-related OSH risk perceptions, organizational responses, and accountability in Japanese workplace settings. It also broadens the empirical focus by examining AI-related OSH perceptions in broader workplace settings, rather than focusing only on platform work or specific AI-enabled OSH tools. By combining questions on AI knowledge, work experience, adoption status, perceived risks, risk mitigation measures, preventive needs, stakeholder responsibility, and legal accountability, the survey offers a broad overview of how AI use is currently perceived in relation to OSH. The inclusion of supplementary materials describing questionnaire domains, routing logic, denominators, item wording, and response options also improves transparency and interpretability.
Several limitations should also be noted. First, the study used a registered commercial online survey panel rather than probability-based sampling. Although participants were screened according to predefined eligibility criteria, no formal quota allocation or stratified sampling was applied. Therefore, the findings should be interpreted as exploratory and descriptive and should not be regarded as statistically representative of all workers or workplaces in Japan.
Second, the data were based on self-reports from individual participants. Information on AI adoption status, organizational measures, risk assessment practices, and responsibility was reported from the participants’ perspectives and was not independently verified at the organizational level. Some participants may not have had sufficient knowledge of their organization’s AI systems, governance practices, or risk management procedures. This limitation is particularly relevant to items with high proportions of “Not sure,” “Neither,” or “Not applicable/AI not used” responses.
Third, because workplace AI adoption and direct AI-related work experience were limited in this sample, many participants may have evaluated AI-related OSH risks hypothetically or with limited direct exposure. As a result, risk perceptions may reflect uncertainty, limited familiarity, or general impressions of AI rather than direct experience with specific AI systems in actual work processes.
Fourth, the questionnaire was developed for this exploratory study and was informed by preliminary interviews, international reports, and multidisciplinary research team discussions. However, no formal pilot test was conducted, and the items were not based on validated psychometric scales for all domains. Therefore, measurement error, differences in interpretation of response categories, and limitations in construct validity cannot be excluded.
Finally, the cross-sectional design does not allow assessment of changes over time or causal relationships between AI adoption, organizational measures, risk perceptions, and accountability expectations. In addition, this study did not conduct stratified analyses by industry, organization size, occupation, job position, department, or level of AI adoption. Such analyses are needed to examine whether AI-related OSH risk perceptions and governance needs differ across workplace contexts and worker groups. Longitudinal research, organization-level case studies, stratified analyses, and studies involving multiple stakeholders within the same workplace would be useful for examining how AI-related OSH risks and governance practices evolve as AI adoption progresses.
This study provides descriptive evidence on AI adoption, AI-related OSH risk perceptions, organizational responses, and accountability expectations among registered online panel participants in Japan. The findings suggest that AI use in Japanese workplaces remains at an early stage for many participants, with limited direct work experience and frequent uncertainty or non-applicability regarding AI-related OSH risks. Organizational measures to address such risks appeared unevenly implemented, while participants showed some recognition of the need for preventive measures, human oversight, clear allocation of responsibilities, and collaboration with OSH experts.
Responsibility and legal accountability were perceived as extending across multiple stakeholders, including organizations and senior management, developers and programmers, providers of training or learning data for AI, and on-site supervisors. These findings suggest that responsible AI use in OSH may require not only technical validation of AI systems but also workplace-level risk identification, human-centered governance, and clarification of accountability across the AI lifecycle. Future research should examine differences across industries, organization sizes, job roles, and levels of AI adoption, and should investigate how AI-related OSH risks and governance practices evolve as workplace AI implementation progresses.
Questionnaire domains, item routing, and denominators.
Supplementary Appendix 2Full questionnaire used in the web-based survey.
Supplementary Table S1Types of AI systems adopted or planned and main purposes of AI use.
Supplementary Table S2Perceived effects of AI adoption on workplace safety and physical and mental health.
Supplementary Table S3Psychosocial responses to AI adoption.
Supplementary Table S4Attitudes toward delegating work-related decisions to AI.
Supplementary Table S5Perceived proportion of work tasks that may be replaced by AI within the next five years.
Supplementary Table S6Necessity of social policy responses to increased automation by AI.
Not applicable.
Consent for publicationNot applicable.
Availability of data and materialsNot applicable.
FundingThis article was supported by the 2025 Health and Labour Sciences Research Grant (Comprehensive Research on Occupational Safety and Health), Ministry of Health, Labour and Welfare, Japan: “A Study on Workplace Management Issues Arising from the Use of AI from a Legal Perspective and Regulatory Challenges under the Current Occupational Safety and Health Law” (PI: The Japan Association of Occupational Health Law) (25JA1004). The funder had no role in the study design, data collection, analysis, interpretation of data, writing of the manuscript, or the decision to submit the manuscript for publication.
Authors’ contributionsMinae Nishimoto contributed to data curation, formal analysis, and writing—original draft. Takenori Mishiba contributed to conceptualization, supervision, and project administration.
Competing interestsTakenori Mishiba, an author of this paper, is a member of the JOWHSR Editorial Board but was not involved in any part of the peer review process for this paper. The authors declare no other competing interests.
Use of artificial intelligence (AI)ChatGPT (version 5.2) was used for literature summarization, drafting, proofreading, and translation during the preparation of this manuscript.
