Five pillars for guiding AI
AI is no longer a laboratory project: it now permeates our public services, our production chains, and even our daily interactions. Its rapid spread necessitates that we start thinking now about the governance that will accompany it; otherwise, society risks suffering from a patchwork of ineffective rules or, worse, a race to the bottom in ethical standards.
Faced with this responsibility, public authorities are called upon to work with civil society and practitioners to clarify five priority areas: explainability, fairness, safety, human-AI collaboration, and legal accountability. These areas form the basis of a strategy capable of supporting innovation while protecting citizens.
Explainability is the first pillar. Understanding why a system suggests a medical diagnosis or denies a loan strengthens trust and allows users to challenge the algorithm if necessary. However, the "right" explanation varies depending on the audience, the context, and the technical constraints; there is no single model.
A balance is essential: the explanation must be clear, sufficient, and proportionate. It must provide concrete reasons without overwhelming the user with details or hindering the adoption of otherwise beneficial solutions. Collections of examples of "good explanations" and graduated scales would help teams calibrate their explainability requirements according to the sector and the risk involved.
When a detailed explanation is impossible or unnecessary, other accountability mechanisms take over: reporting channels, human appeals processes, adversarial testing, or independent audits. These safety nets allow for model correction and reassure users without requiring the "black box" to be opened with every request.
Equity, the second area of focus, confronts organizations with sometimes contradictory definitions: equal treatment, equal outcomes, compensation for historical injustices… A model that is “fair” according to one criterion may be discriminatory according to another. It is therefore necessary to consciously decide which notion of justice to prioritize for each use case.
Regulators can contribute by indicating how to arbitrate these tensions and by identifying the perverse effects of existing rules: prohibiting any inference of ethnicity, for example, sometimes prevents verification of the absence of racial bias and produces the opposite of the intended goal.
Safety forms the third pillar. No system is infallible; risk must be assessed in light of potential damage and the cost of preventive measures. Common testing and documentation standards, and where appropriate, certification labels, could standardize the demonstration of due diligence in critical areas.
On the ground, public-private collaboration could lead to conformity marks similar to CE marking: for example, a smart lock with biometric recognition would only be marketed after validation on a representative test set attesting to a predefined maximum error rate.
The question of "human in the loop" goes beyond simple surveillance. Humans and machines possess distinct strengths; choosing their respective roles requires assessing the task, the ethical implications, and the psychology of the operators. In some cases, the pairing is more effective than each individual part; in others, introducing humans increases errors or dilutes accountability.
We must also avoid the illusion of security: entrusting the final control of a flood of alerts to a human agent often leads, through decision fatigue or automation bias, to the system's recommendations being validated without real verification. Designing escalation thresholds, audit quotas, and incentives for vigilance then becomes essential.
Finally, from a legal standpoint, the idea of granting legal personhood to machines has been rejected: the general consensus is that moral and financial responsibility must remain with humans. Nevertheless, the complex chain of suppliers and integrators can complicate the attribution of fault. Sector-specific regimes—liability caps, “safe harbors,” mandatory insurance—could cover high-risk applications without stifling innovation.
Ultimately, the most promising framework combines responsible self-regulation with targeted government intervention. Standards must remain adaptable, based on empirical risk assessments and ongoing dialogue between researchers, industry, and citizens. This flexible approach, already widely adopted in other emerging technologies, offers the best guarantee of inclusive and safe digital progress for all.
To go further: these five pillars resonate with our article on transparency and trust and with our analysis on generative AI and the GDPR.