Many organizations are rushing to adopt AI tools. Yet successful AI adoption depends on more than achieving strong performance metrics.

Systems that perform well during testing may fail when deployed in complex real-world environments. In sectors such as energy, water, telecommunications, and transportation, AI systems operate in dynamic and often adversarial environments.

Data can be incomplete, operating conditions can change unexpectedly, and malicious actors may actively attempt to manipulate system behavior. Under these circumstances, an AI model that performs exceptionally well during testing may fail when deployed in the real world.

Through our research on cybersecurity and resilience in smart grids, power-communication networks, water infrastructures, and other cyber-physical systems, we have observed a recurring pattern: Trust depends on far more than performance metrics such as accuracy, precision, recall, and F1 score.

Trustworthy AI must remain reliable when confronted with unexpected events, adapt to changing environments, understand the broader system context, and support human decision-makers rather than replace them.

These lessons are increasingly relevant for Lebanon as organizations across both the public and private sectors embrace digital transformation initiatives.

Whether deploying AI in healthcare, banking, telecommunications, education, or critical infrastructure, decision-makers must look beyond benchmark performance and consider the broader characteristics that enable AI systems to remain dependable in practice.

Drawing on lessons learned from cybersecurity research in critical infrastructures, this article highlights three practical considerations organizations should evaluate before trusting AI systems in operational settings.

Lesson 1: Strong Performance Does Not Guarantee Trust

Organizations often evaluate AI systems using performance metrics such as accuracy, precision, recall, and F1 score. While these metrics provide useful insights, they do not necessarily indicate whether an AI system can be trusted in real-world environments.

In smart power grids, we found that attackers can manipulate measurements in ways that closely resemble normal system behavior, making malicious activity significantly harder to detect and reducing the effectiveness of many existing AI-based detection systems (Takiddin et al., 2024).

Trust can also be undermined during training. When malicious samples are incorrectly labeled as benign and included in training datasets, the performance of AI-based detection systems can deteriorate significantly, with some models experiencing performance losses of up to 29% (Takiddin et al., 2023).

These findings highlight an important lesson: Strong benchmark performance does not always translate into dependable real-world behavior.

Trustworthy AI should not only perform well under expected conditions but also remain reliable when confronted with uncertainty, data-quality issues, and deliberate attempts at manipulation.

This lesson is particularly relevant for Lebanon as organizations increasingly adopt AI across energy, telecommunications, healthcare, banking, and public services.

Decision-makers should look beyond performance scores and ask a more important question: How well will the system perform when confronted with unexpected events?

Building trustworthy AI requires designing systems that remain reliable under uncertainty, not only when everything goes according to plan.

Lesson 2: Trust Requires Generalization

A common mistake in AI development is assuming that strong performance during testing guarantees strong performance in practice.

In reality, operational environments continuously evolve. Power grids, for example, undergo seasonal reconfigurations, communication networks change, and new devices are added or removed.

An AI system that performs well only under the exact conditions seen during training cannot be fully trusted.

Our research in smart grids showed that many AI-based intrusion-detection systems are trained and evaluated on a single network topology. While they may perform well in controlled environments, their effectiveness can decline when deployed in previously unseen configurations.

A similar challenge emerged in our research on UAV swarms, where communication networks continuously evolve as drones move and missions change.

We found that intrusion-detection systems must learn patterns that generalize across diverse operating conditions and previously unseen network topologies rather than memorizing a specific network structure (Mughal et al., 2025).

The broader lesson is clear: Trustworthy AI should be evaluated not only on how well it performs under familiar conditions but also on its ability to adapt to new and evolving environments.

This lesson is particularly relevant for Lebanon as organizations increasingly adopt AI in energy, telecommunications, transportation, and public services.

Trustworthy AI should be evaluated not only on historical data but also on its ability to adapt to changing environments and continue operating reliably when conditions evolve.

Lesson 3: Trust Requires Context and Human Judgment

AI systems excel at identifying patterns in data, but trustworthy decision-making requires more than pattern recognition.

In many real-world applications, decisions depend on understanding how information fits within a broader context and how actions affect interconnected systems.

Our research showed that many traditional machine-learning approaches analyze measurements in isolation, overlooking the relationships between interconnected components.

In power grids, for example, an abnormal reading at one location may be perfectly normal when viewed in the context of neighboring substations, transmission lines, and control systems.

By incorporating both cyber and physical relationships, we found that context-aware AI models significantly improved cyberattack-detection performance compared to approaches that relied on isolated cyber or physical measurements alone.

More recently, we showed that capturing higher-order relationships among interconnected components can further improve resilience against sophisticated attacks, highlighting the importance of understanding system context when deploying AI in complex environments (Aboelmagd et al., 2026).

However, understanding system context is only part of the equation. Human expertise remains essential, particularly in complex and high-stakes environments.

Our studies of interdependent power, communication, and water infrastructures showed that failures can propagate across systems in unexpected ways.

Determining how to respond often requires balancing technical, operational, economic, and societal considerations that extend beyond what AI systems can infer from data alone.

In our recent work on blockchain-enabled incident response for power grids, automated mechanisms were used to detect anomalies and trigger coordinated alarms.

However, effective mitigation still depended on collaboration among power-system providers responsible for managing their own infrastructures and implementing appropriate response actions (Islam et al., 2025).

The lesson for organizations is clear: AI should augment human decision-makers, not replace them.

The most trustworthy systems combine context-aware intelligence with meaningful human oversight, allowing experts to interpret recommendations, validate outcomes, and intervene when necessary.

Conclusion

As AI becomes embedded in critical sectors, trust will increasingly determine the success or failure of its adoption.

While performance metrics remain important, they provide only a partial picture of how an AI system will perform when faced with the uncertainties and complexities of real-world deployment.

Our research demonstrates that trustworthy AI must be robust against adversarial manipulation, resilient to unforeseen events, capable of generalizing beyond its training conditions, aware of the broader system context, and supported by meaningful human oversight.

These characteristics become especially important in critical infrastructures where failures can affect public safety, economic stability, and essential services.

As AI adoption accelerates across Lebanon, organizations should resist the temptation to evaluate systems solely on performance metrics.

Trustworthy AI requires robustness, adaptability, contextual understanding, and meaningful human oversight.

By incorporating these principles early in the design and deployment process, organizations can reduce risk, improve adoption, and build greater confidence in AI-driven innovation.

References

Aboelmagd, S., Atat, R., and Takiddin, A. (2026). “Simplicial Graph-Based Detection and Localization of Cyber Attacks Against Large-Scale Smart Grids.” IEEE Transactions on Smart Grid. Advance online publication.
https://doi.org/10.1109/TSG.2026.3682950

Islam, M. M., Atat, R., Ismail, M., Davis, K. R., and Serpedin, E. (2025). “Enhancing Power Grid Management and Incident Response Mechanisms Through Consortium Blockchain.” IET Smart Grid, 8(1), e12203.

Mughal, U. A., Elshazly, A., Atat, R., and Ismail, M. (2025). “Generalizable Topology-Aware GNN-Based Intrusion Detection System for UAV Swarms.” IEEE Internet of Things Journal, 13(1), 1569–1580.

Takiddin, A., Ismail, M., Atat, R., Davis, K. R., and Serpedin, E. (2023). “Robust Graph Autoencoder-Based Detection of False Data Injection Attacks Against Data Poisoning in Smart Grids.” IEEE Transactions on Artificial Intelligence, 5(3), 1287–1301.

Takiddin, A., Ismail, M., Atat, R., and Serpedin, E. (2024). “Spatio-Temporal Graph-Based Generation and Detection of Adversarial False Data Injection Evasion Attacks in Smart Grids.” IEEE Transactions on Artificial Intelligence, 5(12), 6601–6616.