Page 267 - Kaleidoscope Academic Conference Proceedings 2024
P. 267
Innovation and Digital Transformation for a Sustainable World
Table 7 – Top seven deployers of the AI systems in AIID Table 9 – Sharing of incident data by AIAAIC and AIID
Deployer of AI system incidents %age Data AIAAIC AIID
tesla 39 6% sharing
facebook 36 6% Format Available as a Weekly snapshots
google 28 4% Google Sheet. of the database in
unknown 23 4% JSON, MongoDB,
amazon 21 3% and CSV format
openai 20 3% Information Contributor details -
cruise 12 2% not are not public.
accessible Harm data is only
accessible to
Table 8 – Top seven countries of the incidents in AIAAIC
premium members.
APIs None None
Countries Incidents %age
USA 424 46.9% exchange. Secondly, the captured data is generally insufficient
UK 59 6.5% for assessing the severity and proper categorization of the
China 53 5.9% incidents.
USA; Global 26 2.9%
Global 21 2.3% Recommendation 3: Standardise AI-incident database
India 21 2.3% structures: Standardising the fields of AI-incident databases
Canada 18 2.0% will ensure that the collected data has sufficient granularity
required for analysis. It will also facilitate interoperability,
data exchange, and ease of aggregating data from multiple
risks and hinder efforts to develop inclusive and equitable
databases.
solutions.
5.4 Inadequate motive to report incidents
Inference: The AI-incident databases may suffer from the
biases of the submitters or the reviewers related to attributes
Observation: As indicated in Table 1, incident reporting
such as their political leanings, gender, minority groups,
in both databases is voluntary and lacks incentives.
countries, and so on. Further, different individuals classify
Without legal mandates or rewards, reporting relies on
the incidents and their harms in distinct ways, depending on
reporters’ discretion and motivation, potentially resulting in
their exposure, capabilities, and understanding, which may
underreporting.
lead to inconsistencies and misclassification.
Inference: Fears of data privacy breaches may discourage
Recommendation 2: Define guidelines for AI-incident reporting, leading to incomplete or underreported AI
database quality audits: Formulate procedures to regularly incidents. Without transparent and privacy-protective
audit the AI-incident databases for consistency, checking for reporting mechanisms, stakeholders may hesitate to disclose
misreporting, misclassification, reported incidents meeting incidents, hampering the effectiveness of incident databases.
the defined criteria, and so on. Additionally, fragmentation among databases complicates
data collection and analysis, impeding comprehensive risk
5.3 Insufficient and incompatible data fields understanding and response.
Observation: Table 3 compares the columns available Recommendation 4: Develop regulatory and policy
in the two databases, showing that only six fields are frameworks for AI-incident reporting: Make sector-specific
compatible between the two datasets, while the remaining legal provisions to mandate or encourage AI-incident
are incompatible. Secondly, these databases do not have reporting. Global standards organizations such as ITU should
enough detailed data fields needed for thorough analysis, develop standardized regulatory and policy frameworks for
like identifying the causes, context, and impact of reported AI-incident reporting to enable consistency across nations.
incidents. AIID does not have fields to capture impacted
sectors (Table 6), impacted countries (Table 8), and so on. 5.5 Narrow base of the incidents reported
On the other hand, AIAAIC does not capture details of the
harmed (or nearly harmed) parties the way AIID does, such Observation: Though the incident reporting is open to the
as Facebook users, minority groups, patients, and so on. public, only a few individuals report the incidents. Table 4
indicates that just four individuals, excluding the anonymous
Inference: Different and incompatible structures of AI ones, have reported half of the incidents in AIID. Further,
incident databases make aggregating data from multiple the top sources of the reports submitted to AIID are from
databases difficult, limit interoperability, and restrict data American or European newspapers, as detailed in Table 5.
– 223 –