Page 268 - Kaleidoscope Academic Conference Proceedings 2024
P. 268
2024 ITU Kaleidoscope Academic Conference
Inference: Technological interventions and process reforms incidents reported in AIID predominantly relate to AI systems
are required to widen the base of incident reporting. developed by American companies, as evident from Table 7.
Further, the top sources of the reports submitted to AIID are
Recommendation 5: Develop standards for automated from American or European newspapers, as detailed in Table
incident reporting: Develop standards to enable automated 5.
AI-incident reporting through the AI applications to
supplement manual reporting. Inference: Existing AI-incident databases particularly
lack representation from developing and underdeveloped
5.6 Inadequate data-sharing protocols countries. Capturing AI incidents prevalent in these
underrepresented regions is crucial for developing mitigation
Observation: As indicated in Table 9, the two databases strategies. It is also essential in advancing the UN SDGs.
allow downloading data in different formats, and both do
not provide APIs for accessing data. Further, there is Recommendation 8: ITU-led inclusive AI incident
inconsistency related to the information accessible from the reporting: Encourage international collaboration facilitated
two databases (Table 9). The submitter names are accessible by UN organizations, such as ITU, to establish standardized
in AIID but not in AIAAIC. Similarly, AIID provides access protocols for AI-incident reporting, prioritizing inclusivity
to the details of the harmed parties, but in AIAAIC, harm from developing countries. This promotes comprehensive
data is only accessible to Premium Members. understanding and mitigation aligned with UN SDGs.
Inference: Therefore, standardized mechanisms for sharing
5.9 Lack of awareness:
incident data among stakeholders, including government
agencies, industry partners, researchers, and the public, are Observation: As mentioned in the previous paragraphs and
lacking. It impedes collaborative efforts to address emerging observed through Tables 4 and 5, the base of AI incident
trends, root causes, and mitigation strategies for AI incidents. reporting is narrow.
Recommendation 6: Standardise data sharing mechanisms: Inference: The key stakeholders, including industry,
Define protocols for data sharing, access controls, and academia, civil society, the general public, and policymakers,
privacy protection to ensure the confidentiality and security are largely unaware of AI-incident databases. Without
of incident data. Establish mechanisms for sharing incident active involvement from diverse perspectives, databases will
data among stakeholders, including government agencies, fail to capture the full spectrum of AI-related risks and
industry partners, research institutions, and civil society opportunities.
organizations.
Recommendation 9: Awareness programs: Hold
5.7 Sectoral underrepresentation: regular campaigns to enhance stakeholders’ awareness and
understanding of AI incident reporting standards and best
Observation: Existing AI-incident databases have
practices.
skewed representations of application sectors.
"Media/entertainment/sports/art" sector has the highest These standardization actions can enhance the effectiveness,
number of incidents reported in AIAAIC, followed by transparency, and accountability of AI-incident reporting
automotive and politics sectors, as illustrated in Table 6. processes, thereby contributing to the achievement of the UN
Table 7 indicates that the maximum incidents reported SDGs.
in AIID relate to self-driving cars (Tesla, Cruise), social
media (Facebook), search engines (Google), online shopping It is further recommended to include incident reporting as
(Amazon), and advanced AI models (OpenAI). an integral part of the AI lifecycle so that it gets appropriate
focus in the future. Figure 1 illustrates the conceptualized
Inference: While these databases predominantly report AI lifecycle stages to collect data for developing incident
consumer-oriented sectors, they underrepresent critical mitigation strategies.
infrastructure sectors such as telecom and electricity supply.
The AI incidents in such sectors may not be as frequent as
in the consumer-oriented sectors; however, it is still vital to 6. CONCLUSION
maintain a repository of their incidents.
In conclusion, this study highlights the critical need for
Recommendation 7: Sector-specific AI-incident databases: standardized AI-incident reporting to enable data gathering,
Develop sector-specific AI-incident databases to supplement research, and development of mitigation strategies for
the general purpose AI-incident databases. preventing future incidents. Through an analysis of existing
open-access AI-incident databases, it presents the key
5.8 Demographic underrepresentation: observations and gaps in standardization, underscoring the
importance of policy and standardization initiatives in this
Observation: Table 8 shows that just three countries account domain. Table 10 summarises the gaps observed and the
for 60% of the incidents reported in AIAAIC. Similarly, the recommendations to overcome them.
– 224 –