To Govern AI, Start With Human Rights

Recently, I participated in discussions at the Pontifical Academy of Sciences in the Vatican with scientists whose discoveries have shaped our understanding of life, matter, and the universe. They included Nobel laureates Svante Pääbo, Gary Ruvkun, Aaron Ciechanover, Roger Kornberg, Eric Betzig, John Mather, Andrea Ghez, Demis Hassabis, Emmanuelle Charpentier, Steven Chu, Donna Strickland, and William Phillips, along with Fields Medalist Cédric Villani. We also had the privilege of engaging with Pope Leo XIV.

Our conversations spanned artificial intelligence (AI), the life sciences, climate change, and human well-being. One idea kept returning: innovation earns its place when it serves people, protects dignity, and contributes to the common good. This test is urgent for AI, which now shapes how we work, learn, heal, and govern.

The world does not lack AI principles. It lacks a coherent foundation that connects them to practice. At the Vatican, I emphasized that the UN Charter and the Universal Declaration of Human Rights provide that basis. Ten commitments follow from them.

First, begin with human rights. Every moral tradition includes some form of reciprocity: do not do to others what you would not accept being done to yourself. If an opaque system that decides your access to credit, treatment, or liberty would be intolerable, it should be intolerable for anyone else. Rights must be real for the person affected by an algorithm, including a meaningful opportunity to challenge consequential decisions.

Second, understand what we govern. Regulators need technical competence, while developers need to understand the institutions and people their systems affect. As I argued in my essay on rational opacity, a decision may be statistically effective yet lack legitimacy if no one can explain or contest it.

Third, govern both the parts and the whole. Data, algorithms, computing infrastructure, and applications create different risks. Data can encode past exclusion; computing can concentrate power; deployment can place a capable model in an unsuitable setting. Scrutinizing each component matters, but so does examining the effect of the whole system.

Fourth, make the balance explicit. Innovation and precaution, privacy and utility, transparency and performance cannot always be maximized together. No formula resolves every tension. Human rights provide a compass when judgment is required: which choice best protects dignity, agency, and fundamental freedoms?

Fifth, coordinate across borders and sectors. AI crosses borders while legal authority remains largely national. Countries need room for their own institutions and traditions, alongside interoperable safeguards and cooperation to prevent dangerous practices from migrating to jurisdictions with the weakest oversight.

Sixth, govern failure as carefully as capability. Models can perform well on average yet still fail in rare cases, in underrepresented communities, or in situations unlike their training data. Scores and probabilities also translate into discrete decisions: a treatment is given or withheld; a person is detained or released. Oversight must examine uncertainty, distributional effects, routes of appeal, and the consequences of error for individuals.

Seventh, disclose the environmental footprint. AI depends on electricity, cooling water, land, and physical infrastructure. The UNU report on carbon, water, and land footprints informed the case for UN Secretary-General António Guterres’ call for environmental disclosure. Significant AI infrastructure should report comparable resource use across training and deployment so communities can assess its costs.

Eighth, build a system of governance. Values must shape conduct; incentives must reward responsible development; standards must translate principles into practice; laws must establish duties; and institutions must have the capacity to enforce them. A statute without expertise and implementation offers little protection.

Ninth, bridge the divides. Connectivity is only one divide. Countries and communities also differ in computing power, mathematical expertise, regulatory capacity, and access to AI’s benefits. My work on the algorithm divide argues that participation in the emerging economy increasingly depends on data, infrastructure, and the ability to shape the rules. Governance should broaden those capabilities.

Tenth, make AI literacy a civic and institutional capability. Citizens need to know when to question an output, verify a claim, and decline to use a system. Ministries, universities, companies, and religious institutions need the same capacity. In my Doubt and Cross-Validate Protocol, the central habit is to ask what happens if a model is wrong, how the error would be detected, and who bears its cost.

At the Vatican, scientists at the frontiers of discovery and leaders concerned with moral tradition found common ground on how knowledge affects human lives. That convergence offers a practical starting point. The UN Charter and the Universal Declaration arose from a determination to protect every person’s inherent dignity. AI tests that commitment.

Whether AI expands freedom, distributes opportunity, and strengthens human agency depends on the choices made now. Let those choices begin with human rights and let every governance measure be judged by how well it protects the people whose lives AI changes.