Page 67 - ISC PROCEEDINGS 21.4

Moreover, integrating data from multiple distributed systems may lead to risks of
data inconsistency or loss of integrity in the absence of proper standardization. Analytics
performed on non-uniform data can generate false positives or false negatives, thereby
undermining trust in the system. Therefore, data security and integrity must be regarded
as prerequisites for implementing an intelligent monitoring model.
3.4.2. Moral hazards and biases in AI
AI applications in PIE monitoring typically rely on ML models trained on historical
data. However, past data may contain biases arising from differences in local governance
capacity, uneven inspection intensity, or specific historical factors. According to a 2021
OECD study, more than 60% of surveyed public authorities acknowledged concerns about
algorithmic bias risks when deploying AI in the public sector.
If not carefully designed and validated, AI models may inadvertently label certain
groups of projects or localities as “high risk” based on historical patterns rather than
current conditions. This could result in imbalanced allocation of inspection resources and
raise concerns about institutional fairness.
Furthermore, many deep learning models function as “black boxes,” making their
decision-making processes difficult to interpret. Under this framework public governance
- where transparency and accountability are essential - the use of non-explainable AI may
undermine the legitimacy of decisions. Consequently, current international trends,
particularly within the European Union’s AI governance framework, emphasize
requirements for transparency, explainability, and independent auditing of AI systems
used in the public sector.
3.4.3. Accountability in the digital environment
When oversight decisions, inspection sample selection, or risk alerts are supported
by algorithms, a key issue is clearly identifying who bears responsibility in the event of
errors. In traditional models, accountability is typically attached to a specific individual or
agency. However, in a digital environment, the value chain may involve multiple actors:
data providers, the agency operating the system, the algorithm development team, and
the final decision-making authority.
According to a World Bank survey on public sector digital transformation, more than
50% of developing countries face difficulties in defining accountability mechanisms when
implementing automated decision-making systems. Without a clear legal framework, the
risk of “diffused responsibility” may arise, weakening oversight effectiveness and public
trust.
Accountability also entails an obligation to disclose information at an appropriate
level, enabling oversight bodies and citizens to understand the basis of alerts or decisions.
This requires the establishment of algorithmic audit procedures, data logging mechanisms,
and independent oversight arrangements.
Implementing a BDA and AI-based public investment monitoring model is therefore
not merely a technological issue, but a comprehensive governance challenge involving
data security, algorithmic ethics, and accountability. Only when these risks are
systematically identified and managed can an intelligent public financial governance
model operate sustainably and strengthen public trust.
3.5. Directions for Vietnam in building a smart model for monitoring public
investment expenditures
Under this framework Vietnam’s accelerated implementation of the National Digital
Transformation Program under the Government of Vietnam, the application of BDA and

66