Fairness: Machine learning (ML) uses decision-making, which might be biased. For instance, the dataset is biased due to human prejudices.
Transparency: In many cases, it is difficult to understand how ML systems take decisions. Especially when the ML system includes a neural network, there is a lack of explainability of the decisions made by the system.
Misuse: The algorithms can be maliciously used by people.
Security: AI, like every software, is vulnerable to malicious attacks. This might result in unintended actions of the initial design purpose.
Policy: AI has an increasing impact on products and society.
Ethics: AI needs to act under certain ethical standards. Human values are one of the broader goals to limit functionalities.
Control/alignment: AI must be aligned with the values of the designer so that no misinterpretation can happen.
Verily, ML-based systems cannot fully satisfy the current safety standards.
From 'A Review on AI Safety in Highly Automated Driving', Frontiers