Never trust, always verify : a roadmap for Trustworthy AI?

AI systems contain multiple built-in biases that are difficult to detect, including reproducibility bias, selection bias (affecting races, genders, and other characteristics), and reporting bias where results don't reflect reality—these biases can lead to unfair or inaccurate decisions affecting real people.

The "black-box" nature of AI makes accountability nearly impossible: AI systems are inherently difficult to audit and certify because their decision-making processes are opaque, creating challenges for organizations and regulators trying to understand why an AI system made a particular decision.

AI systems are vulnerable to manipulation through untrusted data: When AI systems receive corrupted or malicious input data, they can misbehave in unsafe and insecure ways, posing risks in critical applications like autonomous vehicles and healthcare systems.

Existing trustworthiness principles lack practical implementation: While governments and international organizations have proposed principles for trustworthy AI, these principles are inconsistently applied in practice and subject to different interpretations, limiting their effectiveness in ensuring AI safety and ethics.

Verification and validation of AI is fundamentally challenging: The stochastic (probabilistic) nature of machine learning models makes it difficult to thoroughly test and verify that AI systems will perform reliably and safely in real-world conditions.