- Lack of human control: Many AI algorithms are complex and non-intuitive hence hard to understand and control. Authority of objectivity, automation bias leads to too much delegation and no accountability.
Amazon uses algorithms with little intervention to fire Flex workers.
- Lack of safety: AI is often not robust and not safe when conditions change. AI is often embedded in physical systems which affect physical safety.
e.g. child pedestrians are less likely to be detected compared e.g. a STOP sign covered in snow may not be recognised
- Discrimination: could impact and harm specific groups due to lack of consideration during design, lack of data, lack of testing, and other application problems.
- Privacy invasion, surveillance: allows more pervasive privacy invasion, constant data gathering to get data for training or selling
- Environmental and societal impact: modern AIs require 100s of GPUs with 1000s of hours of training; relies on data centres being constructed (mineral mining) and cooled (heat and power pollution)
Ethical principles for AI
Many institutions have adopted guidelines for ethical development and deployment:
| Principle | Description |
|---|---|
| human oversight and agency | upholding fundamental rights, informed decisions, correct trust, human-in-the-loop |
| safety and robustness | minimisation of errors, functioning for range of settings, resilience to attack |
| privacy | consent, anonymity, access, de-anonimisation |
| transparency | documentation and logging of data / decisions, explainability, informing when AI is used |
| fairness | non-discrimination, accessibility, stakeholder participation |
| societal & environmental wellbeing | impact to environment, social relationships, democracy, politics |
| accountability | auditability, documenting how dilemmas and trade-offs resolved, responsibility for harm |