Societies have learned to govern powerful technologies before. Cars got licenses and seatbelts, medicines got clinical trials, and food got inspectors. AI is harder to govern because it changes fast, spreads through software and serves countless purposes. Still, the main approaches are taking shape, and each rests on a view of what citizens are owed.
Europe's risk-based law
The European Union's AI Act, which entered into force in August 2024, is the most ambitious attempt so far. It sorts AI systems by level of risk, with stricter rules for higher risk. A few uses are banned outright as an unacceptable risk, including manipulation that exploits people's vulnerabilities and social scoring that leads to unfair treatment of people based on their behavior or traits. High-risk uses, such as screening job applicants, scoring credit, grading students or deciding access to public benefits, must meet requirements for data quality, documentation, human oversight and accuracy. Under the Act's transparency rules, a chatbot must tell people they are dealing with an AI, and realistic AI-made images and video must be labeled. Most other systems, like spam filters, face no new duties.
The Act applies in stages. Its bans took effect in February 2025 and its rules for general-purpose models in August 2025, while most high-risk rules were due later, and in late 2025 the European Commission proposed delaying some of them. Supporters see a model for protecting rights. Critics, including some European business leaders, argue that it adds cost and slows innovation in a region that already trails the United States and China in AI.
Older European law matters too. Under Article 22 of the General Data Protection Regulation, people have a right not to face significant decisions made solely by automated means, with some exceptions, and can ask for a human to review such a decision.
Other paths
The United States has no comprehensive federal AI law. President Biden's Executive Order 14110 of October 2023 set up safety reporting and testing requirements, but it was revoked in January 2025 by President Trump, whose administration put more weight on speed and competitiveness. Several states have passed their own laws, and existing rules on fraud, discrimination and consumer protection still apply to AI. China has issued detailed rules for generative AI since 2023, requiring security reviews, labeling and content in line with state-approved values, which combine safety measures with political control.
Internationally, 28 countries and the EU signed the Bletchley Declaration on AI safety at a summit in Britain in November 2023. In 2024 the Council of Europe opened for signature its Framework Convention on Artificial Intelligence, the first legally binding international treaty on AI, which commits its parties to protect human rights, democracy and the rule of law in how AI is designed and used.
Which risks matter most?
A real debate runs through all of this. In 2023 hundreds of AI researchers and executives signed a one-sentence statement saying, in paraphrase, that reducing the risk of human extinction from AI should be a global priority alongside pandemics and nuclear war. Its supporters argue that systems more capable than people could pursue goals nobody intended, and that the time to prepare is before such systems exist. Critics say that focusing on extinction risk distracts from present harms and may favor the largest firms, which can afford heavy rules that would crush smaller rivals. Many researchers hold that both sets of concerns deserve attention, and that safeguards such as testing, transparency and accountability serve both.
Dignity as the test
Beneath the legal detail sits an old idea. Immanuel Kant wrote in 1785, in Thomas Kingsmill Abbott's translation: "So act as to treat humanity, whether in thine own person or in that of any other, in every case as an end withal, never as means only." A person sorted by a system they cannot see or question is being treated partly as a data point. One practical form of Kant's principle is contestability: the ability of people affected by a decision to get an explanation and challenge it. Citizens can ask of any AI used by a government or a large firm: who is accountable, can its decisions be explained, and can I appeal?
Worked example: the benefits pilot
A state agency proposes AI to flag possible fraud in unemployment claims. At a public hearing, a resident asks four questions. What does the model predict, and what proxy does it use? Has it been tested for errors across neighborhoods, ages and language groups? Will anyone lose benefits on a flag alone, or will a trained worker review each case? And can people see why they were flagged and appeal quickly, with payments continuing in the meantime? She cites Robodebt and the Dutch childcare scandal as her reasons for asking. The agency agrees to publish an audit, to use flags only to decide which cases people review first and to keep payments flowing during appeals. The model may still prove useful, but now it serves people instead of sorting them.