–Boldizsár Szentgáli-Tóth, senior research fellow of the ELTE Centre for Social Sciences, Institute for Legal Studies (Budapest) and lawyer at the K-Monitor Public Benefit Association (Budapest), and Rebeka Kiss, junior research fellow of the ELTE Centre for Social Sciences, Institute for Legal Studies (Budapest)
Every technological rupture eventually reaches the constitutional lawyer’s desk. The printing press, nuclear energy and the internet each raised the question whether existing constitutional categories could absorb the new phenomenon. Artificial intelligence (AI) is the latest arrival, and from Santiago to Athens the temptation to reach for the constitution is strong. If an algorithmic system decides who is flagged for a tax audit, matches a face in a crowd to a name in a database and shapes how public administration reasons about individual cases, many conclude that the highest law should say something about it. For the most part it already does, through general rights guarantees. However, almost everything we want from the “constitutional governance of AI” is best placed in instruments that can be revised as understanding improves, and only a narrow residue belongs in the constitutional text itself. That is why a constitutional drafting moment is exactly when that residue should be identified with precision.
The issue is no longer hypothetical: states around the world now operate AI-based decision-support systems in welfare, policing, migration and taxation. This is a global constitutional moment, and constitutional drafters have begun to respond in registers as different as a binding Council of Europe treaty; national amendments in Chile and Brazil; proposals in Greece, Mexico and Minnesota; and a steady stream of court rulings in the Netherlands and France extending old guarantees to new machines. Occasionally a fresh drafting process makes the stakes unusually concrete, as in Hungary. The party that won a constitutional majority on 12 April 2026 has announced a broader participatory process for an entirely new constitution. A country whose 2011 Fundamental Law became an instrument of the majority that wrote it must now decide what a constitution can do about a technology of state power. That makes Hungary a clear test of our claim, because it pulls apart the two things the debate continues to conflate: incorporating a guarantee, and having an institution willing to enforce it.
The instinct to reach upward
The instinct to constitutionalize the state’s use of AI rests on a genuine intuition. Constitutional norms enjoy a durability and a symbolic weight that statutes cannot match: they bind the legislature itself, they structure judicial review, and they express foundational commitments in a form that a transient majority cannot easily undo, whereas a statute on automated decision-making binds only the majority willing to be bound. That is why several scholars speak of an emerging “algorithmic constitutionalism” as a next frontier of constitutional law. As Pollicino and De Gregorio observe, the diffusion of AI reshapes how public and private power is exercised, and the constitutional response should be to reframe the existing apparatus of fundamental rights rather than to abandon it. That reframing has to reach beyond the state to the corporations that build and control the underlying models, whose command of data, compute and infrastructure can rival public power.
Comparative practice already shows drafters acting on that instinct, though for AI so far only in proposals. Chile’s 2021 constitutional reform on neurorights protected brain activity and the data derived from it before the technology matured, not as a self-executing right but as a mandate to the legislature, and the Chilean Supreme Court later applied it against a technology company. The move signaled that some threats to the person are grave enough for the founding document rather than a data-protection code. The case for new rights, made most influentially by Marcello Ienca and Roberto Andorno, invites the obvious parallel: if neurotechnology warrants constitutional protection of the mind, why should algorithmic power not warrant a comparable anchor?
Why entrenchment is the wrong tool
This is a familiar problem of constitutional design, not a new problem about technology. Rosalind Dixon and Tom Ginsburg call this deferral: vague language that leaves the work to courts, or express delegation to the legislature through a “by law” clause. Deferral suits high error costs, where information is thin and a decision taken now is likely to prove wrong, which describes AI exactly. Ginsburg also finds that specificity aids constitutional endurance, but his reason is that detail signals the drafters’ investment and helps constitutional subjects coordinate on the bargain, which a technical definition of a moving technology does not.
The first difficulty is that AI is not a stable object of regulation, a fast-moving bundle of techniques whose capabilities, risks and vocabulary – “machine learning”, “foundation model”, “agent” – shift on a timescale of months. Constitutions are deliberately hard to amend so that they remain stable across generations, and to write AI into one is to fix, in the most rigid available legal form, a concept almost certain to be obsolete before the ink is dry.
The second is that constitutional specificity tends to crowd out interpretive development. The strength of abstract guarantees such as human dignity, equal treatment and effective judicial protection is that courts can extend them to circumstances the drafters never imagined. The German Menschenwürde guarantee has been applied to automated police data analysis, a context its 1949 drafters never foresaw. A narrowly drafted AI clause risks displacing this generative abstraction with a brittle enumeration, inviting the expressio unius argument that whatever it does not mention falls outside constitutional protection.
The third is that entrenchment can produce a false sense of security. A clause requiring automated decisions to respect human dignity is easy to write and close to empty, because it restates a guarantee that already exists. The hard work is technical: when a human must remain in the loop, what counts as a meaningful explanation, how to allocate liability, how to audit a model. That work belongs in the EU AI Act and data-protection law such as Article 22 GDPR, and a constitutional slogan risks substituting for it.
Courts extending old guarantees
The algorithmic state is not a constitutional vacuum. When public administration deploys an automated system to decide individual cases, the ordinary apparatus of the Rechtsstaat already applies: legality requires a statutory basis, the duty to give reasons requires that the affected person be told why, and effective judicial protection requires that a court be able to review, which constrains systems whose outputs cannot be explained.
Courts in several countries have shown how this works. In 2020 a court in The Hague, applying Article 8 ECHR, halted the Dutch state’s SyRI welfare-fraud risk-scoring system, because the legislation gave insufficient insight into how the model worked, and those affected were never told they had been assessed. France reached the same conclusion from its own text: when its Conseil constitutionnel examined the Parcoursup admissions algorithm in 2020, it upheld the confidentiality of the local algorithms but required each institution to publish the criteria it had used once the procedure was over, deriving that duty from the accountability of public officials in Article 15 of the 1789 Declaration. It did not invent an AI-specific right; it extended an eighteenth-century accountability clause to twenty-first-century administrative technology.
Together these cases trace a shared principle of reviewability: the algorithmic exercise of public power is no exception to accountability and to the right to a remedy. None of these guarantees mentions AI, and none needs to, because their abstraction is what allows them to bite. The Netherlands has since faced further scandals of this kind, a reminder that entrenchment is no substitute for enforcement.
A global patchwork
Above any single state, the most significant development is the Council of Europe’s Framework Convention on Artificial Intelligence, opened for signature in 2024 as the first binding international AI treaty, now signed well beyond Europe. It governs AI used by public authorities, enshrining accountability, transparency and non-discrimination, and it shows that enduring commitments can be stated without naming a technique.
National drafters have gone further and named the technology itself. In May 2026 Greek Prime Minister Kyriakos Mitsotakis proposed incorporating AI into the Greek constitution, so that it serves individual freedom and social welfare while its risks are mitigated. Brazil entrenched a data-related right rather than the technology itself, adding personal data protection to the fundamental rights of Article 5 by Constitutional Amendment 115/2022. A 2025 reform proposal from Mexico names AI expressly, giving the federal Congress a mandate to regulate it.
Minnesota in the United States illustrates the opposite risk. A 2026 amendment proposal there would ask voters to declare that AI has no free-speech rights, but critics warn the wording might curtail the rights of people who express themselves through AI tools. Two strategies thus emerge: entrenching express digital rights or objectives (Greece, Chile, Brazil) and extending existing rights by interpretation (the French and Dutch routes), and a badly drafted clause can narrow protection rather than widen it.
The narrow residue
That most of the work belongs below the constitutional level does not mean that all of it does. As De Gregorio argues, the algorithmic society calls less for wholly new rights than for re-anchoring existing guarantees against the new distribution of public and private power. For any constituent assembly drafting a new text, three principles are worth entrenching, each framed at a level of abstraction that will survive the turnover of particular technologies. All three bind public power; private algorithmic power reaches a constitution only through horizontal effect.
First, the primacy of human judgment, meaning how far AI may go in a public decision about a person. AI is a tool, not a legal subject, and responsibility belongs to people. From that follows a guarantee that any decision with significant effect on an individual must involve genuine, accountable human deliberation, and that no one may be sanctioned or excluded from a service automatically. Aziz Huq has argued against precisely such a right, since none of the usual justifications supports a general entitlement to a human decision, and proposes a well-calibrated machine decision instead. That objection lands against a case built on the quality of the output, but not against this one, which is about democratic accountability and the non-delegability of core state functions rather than about AI as such, and that is why it is durable enough for the text.
Second, equitable access to AI as a state objective. Just as several constitutions and international bodies have debated a right to internet access or to digital education, exclusion from AI-based services may bear on equality of opportunity and equality before the law. A forward-looking constitution can treat this as a state objective: not an enforceable entitlement, but a value obliging the state to secure fair access and counter digital exclusion. In the age of the AI state, access is also a precondition for holding public power to account.
Third, transparency of public-sector AI as a principle. Here the algorithmic exercise of public power is no exception to freedom of information and accountability, which means the public must be able to know the existence, funding, developers, purpose and operating logic of every public-sector decision-support system. Public algorithm registers, pioneered by Amsterdam and Helsinki and now run nationally in the Netherlands, already turn that duty into practice.
Conclusion
The constitutionalization contest is really a debate about the right level of legal abstraction. Detailed, revisable rules, the substance of AI governance, belong in legislation and international instruments, where they can keep pace with the technology. Enduring commitments about the individual and public power belong in the constitution, and for the most part are already there, waiting to be applied rather than rewritten.
The worst outcome would be to invert this: to enshrine a fashionable but perishable definition of AI while leaving the hard governance choices to a slogan. Comparative experience shows that fundamental-rights frameworks can be applied to the algorithmic state, but only if a court or the constituent power enforces them. Where a new text is being written, in Santiago, Athens or Budapest, drafters should ask what they want the constitution to do that the existing guarantees, properly interpreted, cannot already do. The answer is narrower than the global enthusiasm suggests, and identifying that residue precisely, at the drafting stage, is the more useful task.
Suggested citation: Boldizsár Szentgáli-Tóth & Rebeka Kiss, Writing Artificial Intelligence Into Constitutions: A Global Constitutional Moment for the Algorithmic State, Int’l J. Const. L. Blog, Oct. 21, 2026, at: http:///www.iconnectblog.com/writing-artificial-intelligence-into-constitutions-a-global-constitutional-moment-for-the-algorithmic-state/
