A trusted AI (or Trustworthy AI is an approach to artificial intelligence that aims to develop systems reliable, ethical, transparent and respectful of human values. It is part of a framework designed to ensure that AI technologies operate in a way that safe, fair and responsibleby aligning their decisions with social, legal and ethical standards.
Its concept is similar to that of a Responsible AI.
Key principles of trusted AI
- Ethics :
- Respect for human rights, dignity and justice.
- Avoid bias discriminatory (e.g. discrimination based on gender or ethnic origin).
- Transparency (Explainability) :
- Ability to explain AI decisions (Interpretable AI).
- Clear documentation of the algorithms and data used.
- Robustness and safety :
- Resistance to errors, malicious attacks (e.g. adverse disturbances) and noisy data.
- Guaranteed reliable operation in real-life conditions.
- Liability (Accountability) :
- Clear definition of legal responsibilities in the event of error or harm caused by AI.
- Implementation of monitoring and audit mechanisms.
- Respect for privacy :
- Protection of personal data (e.g. compliance with RGPD in Europe).
- Use of techniques such as differential confidentiality or thefederated learning.
- Equity (Fairness) :
- Elimination of systemic bias in data or algorithms.
- Guarantee of equal treatment for all users.
- Human control :
- Maintaining human supervision of critical decisions (e.g. medical, judicial).
- Principle of "human-in-the-loop (humans in the loop).
Critical areas of application
- Health medical diagnostics, surgical robots.
- Justice Support for judicial decision-making (assessing the risk of re-offending).
- Finance : granting credit, detecting fraud.
- Autonomous transport Safety of driverless vehicles.
- Recruitment unbiased selection of candidates.
Regulatory frameworks and initiatives
- European Regulation on AI (AI Act) : Classifies AI systems according to their risk (prohibiting "high-risk" uses).
- EU guidelines for ethical AI 7 key requirements, including transparency and diversity.
- OECD Principles on AI promoting innovative and trustworthy AI.
- IEEE Ethically Aligned Design Technical standards for responsible AI
Challenges
- Algorithmic biases Reproduction of social inequalities (e.g. AI recruitment that disadvantages women).
- Black box Complexity of models such as deep neural networks, which are difficult to interpret.
- Security : Vulnerability attacks (e.g. alteration of training data).
- Balancing innovation and regulation Risk of slowing down technological progress.
Examples
- IBM Fairness 360 A tool for detecting and correcting biases in AI models.
- Google What-If Tool Data analysis: analyses the impact of data on predictions.
- Explainable AI models (Explainable AI - XAI) : Methods such as LIME or SHAP for interpreting decisions
