Why Safety Matters

  • Prevent harm: Sets unacceptable behaviors and blocks hazardous outputs or actions.
  • Test for hazards: Uses red teaming, stress tests, and fail safes before exposure to users.
  • Kill switch and fallback: Provides escalation, rate limits, and shutdown paths when risk rises.
  • Post-incident learning: Captures lessons and updates safeguards so failures do not repeat.

When Safety Is Missed

Cruise operated driverless robotaxis on public roads in California. After several safety incidents, including a pedestrian being struck and dragged by a vehicle, regulators reviewed the company’s practices. In 2023 California suspended Cruise’s permits. The action underscores the need for thorough hazard analysis, robust safeguards, and clear escalation paths before and during wide deployment.

Safety Inter-Driver Relationship List

The following table summarizes the 14 safety related, inter-driver relationships. The full 105 relationships can be viewed here:

Note: The convention when displaying drivers Ds vs. Dt, is to display the first driver alphabetically as Ds.

Drivers Relationship Explanation Example
Inter-Pillar Relationships
Pillar: Operational Integrity
Governance vs. Safety Reinforcing Governance frameworks enhance safety by establishing standards ensuring AI operational integrity and harm prevention (Bullock et al., 2024 ) AI governance mandates safety protocols in autonomous vehicles to prevent malfunctions and accidents (Fjeld et al., 2020 )
Robustness vs. Safety Reinforcing Robustness ensures AI operates safely under challenging conditions, enhancing overall safety (Leslie, 2019 ) In medicine, robust AI systems reliably identify anomalies despite data distribution changes, promoting safety (Leslie, 2019 )
Interpretability vs. Safety Reinforcing Interpretability aids safety by enhancing understandability and identifying system flaws (Leslie, 2019 ) In medical AI, interpretable models allow doctors to verify predictions, improving safety (Leslie, 2019 )
Explainability vs. Safety Reinforcing Explainability enhances safety by making AI decision processes transparent and aiding risk assessment (Dubber et al., 2020 ) Explainable models in autonomous vehicles help identify decision-making flaws, promoting operational safety (Dubber et al., 2020 )
Safety vs. Security Tensioned Adversarial robustness efforts enhance security but may reduce safety by increasing complexity (Braiek & Khomh, 2024 ) Autonomous vehicle safety protocols might focus on preventing adversarial attacks at the expense of real-world robustness (Leslie, 2019 )
Cross-Pillar Relationships
Pillar: Ethical Safeguards vs. Operational Integrity
Fairness vs. Safety Tensioned Fairness can conflict with safety since safety may require restrictive measures that impact equitable access (Leslie, 2019 ) Self-driving algorithms balanced between passenger safety and fair pedestrian detection can lead to safety and fairness trade-offs (Cath, 2018 )
Inclusiveness vs. Safety Reinforcing Inclusiveness motivates safety by ensuring diverse needs are considered, enhancing overall safety standards (Fosch-Villaronga & Poulsen, 2022 ) Inclusive AI health tools accommodate diverse groups, improving overall safety and personalized healthcare (World Health Organization, 2021 )
Bias Mitigation vs. Safety Reinforcing Bias mitigation increases safety by addressing discrimination risks, central to safe AI deployment (Ferrara, 2024 ) Ensuring fair training data mitigates bias-related risks in AI models, enhancing safety in autonomous vehicles (Ferrara, 2024 )
Accountability vs. Safety Reinforcing Accountability ensures AI safety by demanding human oversight and system verification, enhancing procedural safeguards (Leslie, 2019 ) In autonomous vehicles, accountability through rigorous standards enhances safety measures ensuring fail-safe operations (Leslie, 2019 )
Privacy vs. Safety Tensioned Balancing privacy protection with ensuring safety can cause ethical dilemmas in AI systems (Bullock et al., 2024 ) AI in autonomous vehicles must handle data privacy while addressing safety features (Bullock et al., 2024 )
Pillar: Operational Integrity vs. Societal Empowerment
Safety vs. Sustainability Reinforcing Safety measures contribute to the responsible lifecycle management, essential for sustainability in AI projects (van Wynsberghe, 2021 ) Applying safety protocols in AI reduces environmental risks, contributing to sustainable management practices (van Wynsberghe, 2021 )
Human Oversight vs. Safety Reinforcing Human oversight improves safety by providing necessary monitoring and intervention capabilities in AI operations (Bullock et al., 2024 ) In aviation, human oversight actively ensures safety by intervening during unexpected autonomous system failures (Williams & Yampolskiy, 2024 )
Safety vs. Transparency Reinforcing Transparency reinforces safety by enabling detection and mitigation of risks effectively (Leslie, 2019 ) Clear documentation of AI processes ensures safety, enabling effective oversight and risk management (Leslie, 2019 )
Safety vs. Trustworthiness Reinforcing Safety measures enhance AI systems’ trustworthiness by ensuring reliability and robust risk management (Leslie, 2019 ) Safety protocols in autonomous vehicles improve trustworthiness, ensuring public confidence and acceptance of the technology (Leslie, 2019 )