Accuracy and AI Bias
The question then arises: What is the error rate of AI detection?
To enable the system to accurately detect weapons in real-world scenarios, ZeroEyes records footage from CCTV and live video in various environments, including daytime, nighttime, and low-light areas such as parking lots or dark buildings. It then performs detailed annotation on hundreds of thousands of images to train the AI model to effectively recognize the characteristics of visibly brandished guns.

ZeroEyes conducted a detailed study analyzing over 300,000 real images in 2021–2022 and found that the false positive rate was only 0.0656%.

Another issue the development team considered is AI Bias. ZeroEyes designed its system to focus on "object-based detection," specifically firearms, rather than detecting "person" characteristics such as face, gender, or skin colour, which often cause bias in other AI systems. This approach allows ZeroEyes to further avoid both privacy concerns and bias.

However, ZeroEyes acknowledges that bias can unintentionally seep into AI systems. Therefore, in the past year, the company conducted an in-depth study in collaboration with a major U.S. retailer, publishing a Fairness Report. This report aimed to analyze the rate and nature of false positive alerts, especially cases where the AI mistakenly detected a weapon when the image was actually of individuals of diverse skin tones. This analysis helped identify the level of demographic bias that might be embedded in the model. The results of this study are therefore crucial for developing a fair, transparent, and suitable AI weapon detection system for use in public spaces, especially in schools, shopping malls, and public transportation, which require a high level of public trust.

Furthermore, recent research in 2025 published on arXiv proposes a new approach to enhance firearm detection accuracy using Human Pose Estimation, which analyses
human posture in conjunction with weapon recognition. AI not only looks for the shape of a gun but also assesses whether a "person's posture" indicates preparation to fire. This helps the system reduce unnecessary false alerts.

A research team from the UK has also developed the CCTV-Gun Dataset, a repository of gun images from real-world CCTV cameras in complex environments such as shops, public gatherings, or transportation hubs, with factors like low light, blur, or oblique camera angles. This dataset is used to train AI to operate in conditions as close to real life as possible. Comparison with various models shows that using real CCTV footage is crucial for significantly enhancing AI's ability to accurately see what humans might miss.

Lessons for Thailand
While Thailand has not faced school shootings as frequently as the United States, it has experienced shocking incidents, such as the event at a kindergarten in Nong Bua Lamphu Province, and mass shootings in shopping malls in Bangkok and Nakhon Ratchasima Province.

These incidents reflect that public safety is a crucial issue that should receive attention from the government, educational sector, and local administrators. Applying AI systems for firearm detection in Thailand, especially in large schools, shopping malls, airports, train stations, stadiums, or large event venues, could be a significant step towards developing a truly Safe City.

This is especially true if combined with real-time emergency alert systems and continuous joint drills with authorities.

Artificial intelligence for firearm detection demonstrates that technology can genuinely save lives, not by "responding after an event," but by "preventing before an event occurs." AI technology does not replace humans; rather, it is a tool that helps humans detect danger faster than ever before. This could be the beginning of a transition from traditional security systems to creating a smart society that prioritizes everyone's life and safety.