Who Owns the Law? Can AI Give Citizens Back Their Rights?

Everyone is expected to obey the law. Very few can understand it without help. That is not merely a linguistic problem. It is a problem of time, money, education and power. Artificial intelligence promises to change this imbalance—but it could also give us a new class of invisible intermediaries. What would it take to put genuinely reliable legal intelligence in citizens’ hands?

The paradox of a law addressed to everybody

Legal language seeks precision. It must account for exceptions, changes over time, conflicts between rules and the unpredictable diversity of human circumstances. Yet those same qualities can make the law practically inaccessible. A citizen who receives a refusal of benefits, a notice from a landlord or a complex administrative decision may not know what law applies, whether the decision is correct or how to appeal.

The usual solution is delegation: hire a lawyer, seek an association’s help or rely on an official or political representative. Many of these intermediaries provide vital expertise and protection. But access is uneven. The United States Legal Services Corporation reported in 2022 that low-income Americans received no or insufficient legal help for 92% of their substantial civil legal problems. This figure concerns a specific US population and is not an estimate for Europe; it demonstrates the scale a justice gap can reach (LSC, 2022).

Now imagine approaching a legal problem in ordinary language: “The authority rejected my application. Is there something I can do?” An assistant could ask the necessary questions, locate applicable provisions, check the decision’s date, identify the competent body and explain possible next steps. Instead of simply offering information, it could make the citizen more capable of acting.

Four barriers, not one

The last barrier matters most. Explaining a rule is useful; enabling someone to challenge its incorrect application is potentially transformative. It is also far more demanding.

Stanford’s experiment: legal design before legal automation

At Stanford’s Legal Design Lab, directed by Margaret Hagan, lawyers, designers, students and technologists have worked since 2013 to make civil justice more usable. Rather than begin with a model and search for customers, the lab starts from actual problems: confusing court notices, eviction processes, people attending hearings without a lawyer, or failures to identify a legal issue.

Its tools and studies include redesigned documents, guidance flows, visual communication and technology developed with courts and legal-aid organizations. Its current AI work asks not just whether a system can answer a question, but whether the answer helps a person safely accomplish something.

Two complementary initiatives are worth examining. JusticeBench is an evolving catalogue of datasets, benchmarks and projects intended to make legal-help AI measurable. The Legal Help Commons aims to share architectures, playbooks and infrastructure so that legal-aid groups do not each have to build and evaluate AI from scratch. JusticeBench explicitly says it is not a public legal-advice service; it is infrastructure for builders and evaluators.

This distinction is important: a polished chatbot is not yet an accessible justice system. It requires reliable sources, procedures, institutional connections, measures of success, and escalation when the model is uncertain.

Europe is working on different pieces of the puzzle

The landscape is less concentrated than at Stanford, but several organizations address closely related questions:

• DIKE, Vrije Universiteit Brussel (Belgium): an interdisciplinary group researching digital justice, algorithmic decisions, online dispute resolution and access to justice.
• Legal Tech Lab, University of Helsinki (Finland): research on legal digitalisation, algorithmic fairness, justice by design, institutions and digital access to justice. It explicitly puts end users at the centre of legal technology.
• HiiL, The Hague Institute for Innovation of Law (Netherlands): people-centred justice projects using evidence, collaboration and implementation to make everyday disputes easier to resolve.
• OpenFisca and Publicodes (France): tools for expressing calculable legal and administrative rules in executable forms. They are not legal chatbots, but could provide a more verifiable foundation for some of an AI assistant’s answers.

These efforts are not interchangeable. Research labs study and test, nonprofits implement justice services, and rules-as-code projects formalize computable portions of legislation. A citizen-centred assistant would have to integrate all three.

Three levels of ambition

1. Understand the law. Explain a text in accessible language, translate terminology and point to the primary source.
2. Navigate and exercise a right. Identify the relevant version of a rule, ask for missing facts, calculate what can legitimately be calculated, identify deadlines and prepare a document for review.
3. Understand and influence lawmaking. Compare legislative drafts, identify who is affected, distinguish evidence from political arguments and help citizens contribute to public consultations.

The third level goes beyond the conventional concept of access to justice. It would mean access to the processes by which law itself is developed and interpreted. Alkemata has previously explored citizen participation in drafting legislation and machine-readable law. Citizen-facing AI connects those ideas: understanding the rule, examining its consequences and being able to challenge or improve it.

Illustration: the architecture of a trustworthy legal assistant

Concept illustration by Alkemata. A robust system should not let a language model silently replace the official sources or make consequential decisions autonomously.

Why this is hard: the law is not just an algorithm

Legal rules have dates, exceptions and hierarchies. An EU provision, a national law, a municipal rule and an administrative decision may interact. Some legal concepts—good faith, proportionality, reasonableness—require interpretation rather than simple computation. Two competent lawyers may disagree without either fabricating the law.

AI adds its own failure modes. In a peer-reviewed study published in 2025, Stanford researchers found hallucination rates of 17% to 33% in the specific legal research tools and tasks they tested. Those figures should not be generalized to every legal AI product, but they disprove the assumption that retrieval from legal databases makes fabricated statements disappear (Magesh et al., Journal of Empirical Legal Studies, 2025).

A dangerous mistake may sound quite plausible: the wrong appeal deadline, a repealed article, a missing exception or invented case law. Legal AI therefore needs versioned citations, jurisdiction checks, meaningful uncertainty indicators, tests with real users, accessible human escalation and protection of sensitive case information.

Who might resist? And whose concerns are justified?

Professional lawyers may fear competition in routine advice and document preparation. Their professional bodies also raise genuine concerns about competence, confidentiality and accountability. In Germany, individualized legal services for others are regulated under the Rechtsdienstleistungsgesetz (RDG); questions of permitted scope and responsibility cannot simply be ignored by calling a product a chatbot.

Public authorities may favour the efficiency of automated explanations while struggling with oversight and liability. More fundamentally, should an assistant helping a citizen contest an authority be controlled by that same authority? The conflict is not inevitable, but institutional independence and auditability become design requirements, not optional features.

Civil liberties advocates rightly ask whether opaque automated advice can exclude vulnerable groups, bias outcomes or collect too much private information. The European AI Act treats certain judicial AI uses as high-risk; this does not mean every general public legal-information chatbot is automatically in that category. The actual intended use matters.

There is also a political risk: an assistant could quietly emphasize one interpretation of disputed law, or frame a legislative debate in a way that favours one interest. An empowering system must distinguish what the law says, what judges or authorities have interpreted, and what people believe the law should become.

A more radical question: why is legislation so hard to understand in the first place?

Perhaps AI should not merely translate legal complexity after the fact. It could help lawmakers test the comprehensibility and consistency of legislation before adoption. Imagine publishing three connected representations of a rule: the legally authoritative text, a plain-language explanation with examples, and a machine-executable model for its genuinely calculable parts. Their differences could be checked systematically and publicly debated.

That would not abolish ambiguity or make every legal standard computable. It would create a stronger expectation that legal complexity be explained, justified and tested. The rules themselves would become objects citizens could inspect rather than mysteries mediated entirely by experts.

What would a meaningful experiment measure?

Start with a narrow, high-volume domain—for instance, a clearly bounded administrative-benefits procedure. Compare three groups: ordinary web guidance; a conversational assistant with verified source retrieval; and the same assistant with an explicit human-review option. Measure not only user satisfaction but the correctness of identified rules, missed deadlines, appeal quality, task completion, errors and whether users actually gain autonomy.

Report failures publicly and recruit people with different language, literacy and digital-access needs. A tool that impresses law students but misleads vulnerable claimants has not solved the problem it was built for.

Further reading and places to explore

• Samuel I. Becher and Benjamin Alarie, Superjustice: Law in the Age of Artificial Intelligence (Oxford University Press, 2026). A provocative framework for rethinking justice systems around AI, inequality and institutional design.
• Miso Kim, Dan Jackson and Jules Rochielle Sievert (eds.), Legal Design: Dignifying People in Legal Systems (Cambridge University Press, 2024). Human-centred methods to make legal systems usable.
• Richard Susskind, The Future of Law: Reflections and Predictions (30th anniversary edition, Oxford University Press, 2026). A longer view of technological disruption and resistance within legal institutions.
• Margaret Hagan, Law by Design. A freely accessible practical introduction to redesigning legal communication and procedures.
• Stanford JusticeBench and Helsinki Legal Tech Lab research for ongoing work rather than only books.

Explore → Understand → Shape

The decisive question may not be “Will AI replace lawyers?” but: Can citizens gain enough independent understanding of the rules governing them to exercise their rights and question how those rules are made?

That question crosses technology, law and democracy. Explore the legal sources and research. Understand which rules actually apply, where interpretations differ and where uncertainties remain. Shape better public services, better legislation—and perhaps an open, independent, citizen-centred legal infrastructure.

The outcome is not guaranteed. An AI interface could distribute legal power more fairly, or concentrate it in the hands of the organizations controlling the models, data and answers. Which direction prevails will depend less on chatbot fluency than on governance, verification, independence and the public’s ability to scrutinize the whole system.