A unified ontological and explainable framework for decoding AI risks from news data

Statistical analysis of various attributes

Analysis of event attributes

The risky AI events were rarely reported before 2000, but they gradually increased from 2000 to 2012, and experienced a spike after 2013. This trend can be attributed to two main factors. First, recent advances in the Internet and mobile communications have brought more relevant events to public attention and expanded the dataset. Second, as the popularity of AI products and services increases, so does the number of risk events and publications. As shown in Fig. 4a, the introduction of deep learning technology, particularly in 2012, marked a significant shift. Subsequent developments have triggered a wave of deep learning research and application, accelerating the overall progress of AI and fueling the rise of AI risk incidents.

The spatial distribution of AI risk is shown in Fig. 4b. Events with identified countries were most prevalent in the United States, followed by China, the United Kingdom and India. This finding may be associated with the prevalence of AI applications in communication platforms such as social media in recent years, as well as globalization trends driven by large AI companies involved in cross-border production and trade.

The distribution of technology providers is shown in Fig. 4c. Among all risk events, the top 5 most involved technology providers were “Facebook”, “Google”, “Tesla”, “OpenAI”, and “Amazon”, accounting for 49.86% of the risk events containing clear providers. There are 17 technology providers involved in the number of times more than twice. These 17 technology providers involved in the incident accounted for 70.03% of the risk events. Fig. 4c shows that AI products from large companies are more widely used and have more reported risks, suggesting that these large companies should take more responsibility for AI risk prevention.

Fig. 4d shows the distribution of involved AI technology of AI risk events. Based on it, the primary areas of AI technology involved were recognition algorithms, evaluation algorithms, autopilot, moderation algorithm and AI conversation programs which together accounted for 63.88%. Among these, identification algorithms, evaluation algorithms, and moderation algorithm carry inherent risks due to their susceptibility to human interference during redesign. As these algorithms are mature and widely used, they require increased security attention. On the other hand, autonomous driving and AI conversational programs, while posing more challenges due to their nascent nature, have significant future potential. Therefore, ensuring the security of these technologies is critical to prevent potential risks.

Fig. 4e represents the distribution of victimization type of AI risk events. The percentage of total risk events with a clear victimization type is 52.40%, with the most events involving discrimination or victimization of racially profiled people, vehicles, staff, people subjected to gender discrimination, and students with a total of 79.68%.

Harm of AI risk

The percentage of the four types of harm in AI risk events is shown in Fig. 5. 48.3% of all events involved human rights violations. Psychological harm was involved in 50.91% of risk events, physical harm in 22.38%, and economic harm in 37.70%.

Fig. 6a demonstrates that among all risk events with psychological harm, 17.54% involve vulnerable groups. 25.05% are reversible. 18.95% affect self-identity and values. 37.12% result permanent harm. It is evident that a significant proportion of risk events resulting in psychological harm involve vulnerable groups, suggesting an urgent need for targeted interventions to protect these populations. In addition, the high proportion of reversible harm events suggests the potential effectiveness of mitigation strategies in addressing such risks. However, the significant proportion of risk events that cause permanent harm underlines the importance of implementing robust preventive measures. Fig. 6b shows severity levels, with level 2 accounting for 12.0%, level 3 for 46.8%, level 4 for 28.0% and level 5 for 12.8%. Fig. 6c shows a concerning trend of increasing annual occurrences of psychological harm events, albeit with a decreasing proportion relative to all types of harm. This suggests a potential improvement in overall risk management strategies, although continued vigilance and proactive measures are required to address the evolving landscape of AI-related risks effectively.

For risk events involving physical harm as shown in Fig. 7a, 6.45% are reversible. 18.99% are detectable, and 11.72% cause permanent physical harm. A considerable percentage of these events are reversible, suggesting the potential effectiveness of intervention strategies. In addition, the detectability of a significant proportion of physical harm events implies opportunities for early detection and prevention. However, the presence of events resulting in permanent physical harm underscores the urgency of implementing rigorous preventive measures.

With regard to the severity of physiological harm as shown in Fig. 7b, the severity in level 2 accounts for 26.1%, level 3 for 30.6%, level 4 for 16.2%, and level 5 for 27.0%. As can be seen from Fig. 7c, although the number of events with physical harm fluctuates significantly over time, the overall proportion decreases. This indicates potential improvements in risk mitigation efforts, although continued vigilance and proactive measures are essential to effectively address emerging challenges.

According to Fig. 8a, 28.02% of risk events with economic loss contain persistent loss, highlighting the enduring impact on affected entities. In Fig. 8b, the severity of economic loss is categorized as level 2 in 29.9% of economic loss risk events, level 3 in 35.8%, level 4 in 12.8% and level 5 in 21.4%. According to Fig. 8c, the percentage of risk events with economic loss has remained relatively stable over time at approximately 40%. This indicates an ongoing challenge in mitigating the risk of economic loss and the need for sustained efforts to address underlying vulnerabilities. Overall, these findings underscore the importance of implementing effective risk management strategies to minimize the economic impact of such events on affected companies and the broader ecosystem.

human rights violations are categorized as privacy violations and equal rights violations.

Among the risk events involving privacy violations shown in Fig. 9a, 49.02% of them involved sensitive privacy. That is, nearly half of these events involve sensitive data breaches, indicating the severity of the compromised information. In Fig. 9b, the severity of privacy violation is categorized as level 2 in 22.5%, level 3 in 32.4%, level 4 in 18.6%, and level 5 in 26.5%. According to Fig. 9c, risk events involving privacy violations are consistently around 20% over time. This persistent level underscores the ongoing challenge of protecting privacy in the face of evolving threats and technologies. Overall, these findings underscore the critical importance of robust privacy policies and regulatory frameworks to mitigate the risks of data breaches and to uphold individuals’ rights to privacy and data protection.

Among the risk events involving equal rights violations shown in Fig. 10a, 15.32% contain vulnerable groups in terms of equal rights. In Fig. 10b, the severity of equal rights violations is categorized as level 2 in 11.9%, level 3 in 46.2%, level 4 in 30.8%, and level 5 in 11.2%. According to Fig. 10c, risk events involving equal rights violations consistently accounted for approximately 30% of total risk events. This persistent level underscores the ongoing challenge of addressing systemic inequalities and upholding principles of equality and justice. Overall, these findings highlight the critical need for comprehensive measures and policies to prevent and address equality violations, promote inclusion and ensure equitable treatment for all individuals and communities.

Impact attributes

According to Fig. 11, in terms of impact scope of AI risk events, 6.5% of the events had an undefined scope. Of the risk events with a defined scope, 8.8% are global, 28% are individual, and 63.1% are local. Risk events affecting individuals and localized populations have increased significantly from in number, although their percentage of the total has remained relatively stable over time. This suggests a consistent pattern of localized impacts despite an overall increase in the number of AI risk events. These findings underscore the need for targeted interventions and tailored risk management strategies to address the specific needs and vulnerabilities of affected individuals and communities. They also highlight the importance of understanding the localized nature of many AI-related risks and the potential implications for policy development and implementation. Overall, these findings enhance our understanding of how AI risk events are distributed and evolve over time, thereby supporting more effective mitigation strategies and policy planning.

Fig. 11 reveals significant insights into the transmission potential of AI risk events. Just over half, 50.20%, have the capacity for worldwide dissemination, facilitated by the pervasive influence of social media platforms. This highlights the interconnected nature of global technological ecosystems and the role of social media in amplifying the spread of AI risks. Moreover, the persistent proportion of risk events with global transmission capabilities, which remains consistently around 50% after 2015, underlines the ongoing challenge of mitigating and controlling such risks. Effective prevention and control strategies should therefore incorporate measures to enhance social media governance and AI-related products and services provided by globally influential corporations. By addressing both the technological and social dimensions of AI risk transmission, stakeholders can work towards mitigating the potential harms associated with global dissemination and promoting responsible AI development and deployment. Overall, these findings reveal the need for collaborative efforts across sectors to address the complex challenges posed by the global transmission of AI risks and ensure the responsible and ethical advancement of AI technologies.

AI characteristic attributes

Fig. 12 provides critical insights into the factors contributing to AI risk events and their trends over time. Among the various reasons identified, the limitations of traditional oversight methods emerge as the predominant cause, accounting for 55.85% of risk events. This highlights the inadequacy of existing regulatory frameworks and oversight mechanisms to effectively manage AI-related risks. In addition, untimely maintenance of training data emerges as another significant contributor, accounting for 54.64% of risk events. This underscores the importance of ongoing data quality assurance and maintenance practices in ensuring the reliability and integrity of AI systems. Opacity or poor reproducibility is also a significant factor, contributing to 42.34% of risk events. The increasing prevalence of risk events attributed to large AI models (18.75% of the total) shows the unique challenges posed by the complexity and scale of such systems.

The trends observed in these risk factors reveal notable patterns. Risk events involving large models have shown a gradual increase in recent years, reflecting the growing adoption and complexity of AI technologies. Similarly, the percentage of risk events attributed to untimely maintenance of training data has consistently hovered around 50%, indicating an ongoing challenge in data management practices. In addition, the percentage of risk events related to opacity or weak reproducibility is steadily increasing year over year, highlighting the need for greater transparency and accountability in AI development and deployment processes. Of particular concern is the escalating trend of risk events resulting from the limitations of traditional oversight methods, which have surpassed the 50% threshold in recent years. This underscores the urgency of enhancing regulatory frameworks and oversight mechanisms to address effectively the evolving landscape of AI-related risks.

These findings show the need for comprehensive risk management strategies that encompass regulatory, technical, and ethical dimensions. By addressing the root causes identified in Fig. 12 and adapting to emerging trends, stakeholders can work to promote the responsible and ethical development and deployment of AI technologies.

Lifecycle attributes

Fig. 13 shows the distribution of stages in AI product lifecycle where these events occur can be categorized as 29.03% in plan and design stage, 42.42% in data acquisition and preprocessing stage, 60.28% in model building stage, 70.77% in verification and validation, 52.02% in deployment, 73.79% in operation and supervision and 37.30% in user usage and its influencing factors (corresponding to the “user experience and interaction” stage defined in the ontological model).

Fig. 13 provides valuable insight into the distribution of AI risk events across the different stages of the product lifecycle. The largest proportion of risk events, 73.79% of the total, occur during the operation and monitoring phase, indicating the ongoing challenges of monitoring and managing AI systems in real-world environments. Because operational environments may introduce unpredictable user behaviors, dynamic data shifts, and evolving threats, flaws not evident during controlled testing could emerge post-deployment. The complexity of real-world ecosystems likely amplifies risks through increased pathways for errors and vulnerabilities. In contrast, model-building occurs under more stable and controlled conditions, where issues are more easily detected internally. This disparity highlights the need for robust post-deployment governance and adaptive risk management. The second largest proportion, 70.77% of the total, occurs during the verification and validation phase. This highlights the critical importance of rigorous testing and validation processes to identify and mitigate potential risks prior to deployment. Additionally, 60.28% of events occurred during the model building phase, highlighting the need for careful consideration of algorithmic design and training processes to mitigate the risks associated with model biases and errors. Similarly, risk events during the deployment phase, accounting for 52.02% of the total, underscore the importance of robust deployment practices to ensure seamless integration and functionality of AI systems in production environments.

The percentage of risk events occurring in plan and design stage is decreasing year by year. The percentage of risk events occurring in data acquisition and preprocessing stage increase slightly in recent years, to just over 50%. The percentage of risk events occurring in model building stage is consistently around 70%. The percentage of risk events occurring in verification and validation stage is consistently around 80%. The percentage of risk events occurring in deployment stage is consistently around 50%. The proportion of risk events occurring in operation and supervision stage is increasing. The proportion of risk events occurring in user usage stage is consistently around 30%. The distribution of risk events across phases shows certain trends over time. While the percentage of risk events in the planning and design stage has decreased over time, the percentage of events in the data collection and preprocessing stage has increased slightly in recent years, exceeding 50%. This suggests a shift in focus toward addressing challenges related to data quality and preprocessing techniques. In addition, the consistency in the percentage of risk events during the model building and verification/validation phases highlights the ongoing challenges and complexities associated with these phases of the AI product lifecycle.

Overall, these results suggest that AI risk management requires comprehensive strategies that span the entire product lifecycle. By identifying and addressing risks at each stage, stakeholders can work to improve the safety, reliability, and trustworthiness of AI systems.

Correlation analysis of attributes

When investigating AI risk events, distinctions are made based on categorical attributes such as technology provider, AI technology, and type of victimization. These attributes are interrelated and jointly shape the patterns and distribution of AI risk events. Based on the cases collected in this research, the correlations between the attributes were analyzed and some preliminary interpretive insights were drawn to better understand these relationships. Fig. 14 shows a correlation matrix of harm attributes, AI characteristic attributes and lifecycle attributes. The correlation between “Model Building” and “Verification and Validation” is 0.62, suggesting that there is a strong correlation between these attributes. This association is indicative of the interdependence between the quality of the constructed model and the efficacy of the verification and validation processes. A meticulously crafted model typically undergoes testing and validation to ascertain its accuracy and reliability. Conversely, shortcomings or deficiencies in the model-building phase may surface during verification and validation, establishing a correlation between these attributes.

The correlation between “Model Building” and “Untimely Maintenance of Training Data” is 0.41, indicating that there is a moderate positive correlation. This is likely attributed to the consequential effect of timely maintenance of training data on the process of model building. Delays or lapses in updating and maintaining the training data utilized for model construction can impact the quality and precision of the resultant models. Thus, the correlation suggests that challenges in training data maintenance may influence the overall model building process.

The correlation between “Psychological Harm” and “Equal Rights Violations” is 0.52, indicating a significant relationship. This correlation is likely explained by the psychological consequences of violations of equal rights. Instances of unequal treatment or discrimination can result in psychological harm, such as distress, anxiety, or other negative emotions. Therefore, the correlation underscores the connection between individuals’ psychological well-being and the occurrence of equal rights violations, suggesting a cause-and-effect relationship.

Analysis based on explainable machine learning

Using datasets from the ontological risk model, we applied XGBoost with SHAP for interpretability analysis on five attributes: privacy violations, equal rights violations, psychological harm, physical harm, and economic loss. Fig. 15 shows the top five features and their SHAP values. In Table 3, we present the evaluation metrics of the five models.

For the results of privacy violation, the top five important features of the privacy violation prediction model and their mean SHAP values were first presented. Ranked by importance, Data Acquisition and Preprocessing emerged as the most significant feature, exerting the greatest influence on the model output. Then it was followed by Equal Rights Violations, Physical Harm, Plan and Design, and Persistence of Psychological Harm. Furthermore, for Data Acquisition and Preprocessing, Plan and Design, and Persistence of Psychological Harm, higher feature values corresponded to positive SHAP values. This indicates that issues in these areas are significant driving factors behind privacy violations, especially Data Acquisition and Preprocessing, showing its critical role in privacy breaches.

In the prediction model for Equal Rights Violations, the top five most important features and their mean SHAP values were extracted. Ranked by importance, they are: Severity of Psychological Harm, Characteristics of Vulnerable Groups of Psychological Harm, Influential Attributes of Self-identity and Values, Physical Harm, and Privacy Violation. Among these, Severity of Psychological Harm stands out as the most significant feature influencing the model output, highlighting its central role in predicting equal rights violations. For Severity of Psychological Harm, Characteristics of Vulnerable Groups of Psychological Harm, and Influential Attributes of Self-identity and Values, higher feature values correspond to positive SHAP values. This indicates that the presence of these features (with higher values) significantly drives the model’s prediction of equal rights violations.

In the Psychological Harm prediction model, five most influential features, ranked by their mean SHAP values, were identified as follows: Human Rights Violations, Severity of Equal Rights Violations, Caused by Untimely Maintenance of Training Data, Physical Harm, and Local Factors. Among these, Human Rights Violations stands out as the most impactful feature, with the highest mean SHAP value, emphasizing its critical role in shaping the model’s predictions. Both Human Rights Violations and Severity of Equal Rights Violations emerged as the most impactful features in the model, indicating that human rights and equal rights issues are key drivers of psychological harm. The feature Caused by Untimely Maintenance of Training Data highlights that risks associated with outdated training data can lead to psychological harm. Possibly, untimely maintenance of training data can cause AI systems to produce biased, misleading, or inappropriate outputs, which can erode users’ trust, provoke emotional distress, or reinforce harmful stereotypes, thereby inflicting psychological harm. Although Physical Harm and Local Factors have relatively lower mean SHAP values, they still contribute meaningfully to the prediction of psychological harm.

In the interpretability analysis of the Physical Harm prediction model, Human Rights Violations emerged as the most dominant feature. However, its influence on Physical Harm exhibits a suppressive effect. This phenomenon may stem from the tendency of news cases involving Physical Harm to overlook Human Rights Violations, and conversely, cases centered on Human Rights Violations are less likely to include instances of Physical Harm. Based on the SHAP results, it can also be concluded that Physical Harm primarily impacts individuals. Errors during the Model Building phase are identified as the most significant contributors to the occurrence of Physical Harm among all lifecycle attributes.

For modeling Economic Loss, SHAP analysis reveals that User Experience and Interaction is the lifecycle stage with the greatest impact on economic loss. In other words, issues arising during the User Experience and Interaction phase are the most likely to lead to economic losses. Additionally, Privacy Violation is identified as another significant contributing factor. The occurrence of Privacy Violations also contributes to Economic Loss, highlighting the interconnected nature of privacy breaches and their financial consequences.

Based on the above results, we are able to provide a more in-depth interpretation of the cases presented in Table 2. Based on the SHAP-based interpretability results, this event presents a high likelihood of equal rights violations and psychological harm. The presence of severe psychological impact, involvement of vulnerable groups, and influence on self-identity align with the top SHAP features identified in both the equal rights and psychological harm prediction models. Moreover, the fact that the harm was triggered by untimely maintenance of training data further strengthens the model’s prediction, as this feature consistently shows a strong positive SHAP value in the psychological harm model. The engagement of early AI lifecycle stages—especially data acquisition and preprocessing—also contributes significantly to both harm dimensions, supporting the interpretation that this event is a compound ethical issue centered on representational bias and identity-based harm.