İşyeri Görsellerinden Güvenlik Değerlendirmesi: Beş Mesleki Bağlamda Yapay Zeka Asistanlarının Karşılaştırılması

Öz

Bu çalışma, üç genel amaçlı yapay zeka asistanının beş işyeri bağlamındaki görüntü temelli güvenlik değerlendirmelerini karşılaştırmıştır. İnşaat, atölye/üretim, bakım/teknik servis, depo/lojistik ve laboratuvar ortamlarını temsil eden 15 görsel ChatGPT GPT-5.5, Claude Sonnet 5 ve Gemini 3.5 Flash tarafından değerlendirilmiş; 45 tekil görsel ve 15 bağlam içi sıralama çıktısı elde edilmiştir. İki iş sağlığı ve güvenliği uzmanı çıktıları beş boyutta bağımsız olarak puanlamış ve altı hata örüntüsünü kodlamıştır. Değerlendiriciler sıralı puanların %81,8’inde aynı veya yalnızca bir puan farklı değerlendirme yapmış; Cohen kappa katsayısı orta düzeyde bulunmuştur (κ = 0,479; %95 güven aralığı [0,340–0,595]). ChatGPT görünür kanıt kapsamı, görsele sadakat, önceliklendirme ve belirsizlik disiplini boyutlarında en yüksek ortalamalara ulaşırken, Claude kontrol önerilerinin niteliğinde en yüksek betimsel ortalamayı elde etmiştir. Modeller arasındaki farklar kontrol önerileri dışındaki dört boyutta anlamlıdır. Üç model de beş bağlamın dördünü doğru sıralamış ve 15 ikili ilişkinin 14’ünü doğru kurmuştur. Ortak hata, yüksekte çalışan kişinin bulunduğu sahnenin daha düşük riskli sıralanmasıdır. Bulgular, yapay zeka asistanlarının gözetimli ilk değerlendirmeyi destekleyebileceğini, ancak bağımsız güvenlik değerlendirme aracı olarak kullanılmaması gerektiğini göstermektedir.

Anahtar Kelimeler

• Yapay Zeka Asistanları
• İşyeri Güvenliği Değerlendirmesi
• Görsel Kanıt
• Risk Önceliklendirme
• İş Sağlığı ve Güvenliği

Workplace Safety Judgment from Images: Comparing AI Assistants Across Five Occupational Contexts

Öz

This study compared the image-based workplace safety judgments of three general-purpose AI assistants across five occupational contexts. Fifteen workplace images representing construction, workshop/manufacturing, maintenance/technical service, warehouse/logistics, and laboratory settings were evaluated by ChatGPT GPT-5.5, Claude Sonnet 5, and Gemini 3.5 Flash. Each model completed 15 single-image assessments and five within-context relative-ranking tasks, yielding 60 outputs. Two occupational safety specialists independently rated the outputs on five 0–5 dimensions and six binary error patterns. The raters showed exact or near-exact score agreement in 81.8% of the ordinal ratings; pooled quadratic weighted agreement was moderate, κ = 0.479, 95% CI [0.340, 0.595]. Agreement was complete for pairwise ranking accuracy and exact final ordering. Based on the two-rater means, ChatGPT led in visible evidence coverage (4.37), visual fidelity (4.43), prioritization (3.73), and uncertainty discipline (4.70), whereas Claude obtained the highest descriptive control-guidance mean (4.40). Between-model differences were statistically significant in the other four dimensions but not in control guidance. No rater-confirmed problematic control recommendation was observed. All three models achieved the same final-match performance: four of five context rankings were fully correct and 14 of 15 pairwise relations were correct. The shared failure occurred in the workshop/manufacturing context, where the image containing a visible person at height was ranked below a dusty but static scene. The findings support supervised first-pass use of AI assistants but not autonomous workplace safety evaluation.

Anahtar Kelimeler

• AI Assistants
• Workplace Safety Judgment
• Visual Evidence
• Risk Prioritization
• Occupational Health and Safety

Kaynakça

1. Chan, C.-F., Wong, P. K.-Y., Guo, X., Cheng, J. C. P., Chan, J. P.-C., Leung, P.-H., & Tao, X. (2025). Context-aware vision-language model agent enriched with domain-specific ontology for construction site safety monitoring. Automation in Construction, 177, 106305. https://doi.org/10.1016/j.autcon.2025.106305
2. de Marco, G., Niederwieser, E., & Siegele, D. (2025). Exploring a multimodal conversational agent for construction site safety: A low-code approach to hazard detection and compliance assessment. Buildings, 15(18), 3352. https://doi.org/10.3390/buildings15183352
3. El-Helaly, M. (2024). Artificial intelligence and occupational health and safety, benefits and drawbacks. La Medicina del Lavoro, 115(2), e2024014. https://doi.org/10.23749/mdl.v115i2.15835
4. Guan, T., Liu, F., Wu, X., Xian, R., Li, Z., Liu, X., Wang, X., Chen, L., Huang, F., Yacoob, Y., Manocha, D., & Zhou, T. (2024). HallusionBench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 14375–14385. https://doi.org/10.1109/CVPR52733.2024.01363
5. Guo, K. X., Wong, P. K.-Y., Cheng, J. C. P., Chan, C.-F., Leung, P.-H., & Tao, X. (2025). Enhancing visual-LLM for construction site safety compliance via prompt engineering and bi-stage retrieval-augmented generation. Automation in Construction, 179, 106490. https://doi.org/10.1016/j.autcon.2025.106490
6. Hussain, R., Lee, D., Abbas, M. S., Zaidi, S. F. A., Pedro, A., & Park, C. (2026). Vision-language model-based intelligent assistant for onsite construction safety inspection. Automation in Construction, 182, 106728. https://doi.org/10.1016/j.autcon.2025.106728
7. International Labour Organization. (2025). Revolutionizing health and safety: The role of AI and digitalization at work. International Labour Office. https://doi.org/10.54394/KNZE0733
8. Jetha, A., Bakhtari, H., Irvin, E., Biswas, A., Smith, M. J., Mustard, C., Arrandale, V. H., Dennerlein, J. T., & Smith, P. M. (2025). Do occupational health and safety tools that utilize artificial intelligence have a measurable impact on worker injury or illness? Findings from a systematic review. Systematic Reviews, 14, 146. https://doi.org/10.1186/s13643-025-02869-1

9. La Fata, C. M., Barone, G., & Cammarata, M. (2026). Improving construction site safety with large language models: A performance analysis. Information, 17(2), 210. https://doi.org/10.3390/info17020210
10. La Torre, G., Manai, M. V., Meucci, S., Lucente, A., Picerno, A., Ammirati, S., & De Sio, S. (2026). Artificial intelligence and occupational health and safety: A systematic review. Journal of Public Health. https://doi.org/10.1007/s10389-026-02738-8
11. Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50_30392
12. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W. X., & Wen, J.-R. (2023). Evaluating object hallucination in large vision-language models. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 292–305. https://doi.org/10.18653/v1/2023.emnlp-main.20
13. Lou, X., Xu, J., Yin, J., Wang, X., Kang, Z., Liao, Y., Wang, Y., Shi, X., Mo, F., Yao, S., & Huang, K. (2026). When helpers become hazards: A benchmark for analyzing multimodal LLM-powered safety in daily life. Findings of the Association for Computational Linguistics: ACL 2026, 28937–28963. https://doi.org/10.18653/v1/2026.findings-acl.1446
14. National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0) (NIST AI 100-1). https://doi.org/10.6028/NIST.AI.100-1
15. Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253. https://doi.org/10.1518/001872097778543886
16. Perlman, A., Sacks, R., & Barak, R. (2014). Hazard recognition and risk perception in construction. Safety Science, 64, 22–31. https://doi.org/10.1016/j.ssci.2013.11.019
17. Sun, J., Chang, F., Zhou, Z., Man, S.-S., & Chan, A. H. S. (2025). A systematic review of hazard recognition and risk perception research in the construction industry. Safety Science, 186, 106813. https://doi.org/10.1016/j.ssci.2025.106813
18. Wu, X., Guan, T., Li, D., Huang, S., Liu, X., Wang, X., Xian, R., Shrivastava, A., Huang, F., Boyd-Graber, J., Zhou, T., & Manocha, D. (2024). AutoHallusion: Automatic generation of hallucination benchmarks for vision-language models. Findings of the Association for Computational Linguistics: EMNLP 2024, 8395–8419. https://doi.org/10.18653/v1/2024.findings-emnlp.493