AI Policy Then and Now: What Five Years of Rules Actually Changed

Archi3 min read

aipolicyregulation

Contents
  1. From Guidance to Rulebooks
  2. The three things that changed
  3. Where the Gaps Are
  4. What to watch next

From Guidance to Rulebooks#

In 2021, most AI policy was advisory. The OECD AI Principles, published in 2019, had been endorsed by dozens of countries, but they were recommendations. The EU's AI Act was still a draft. In the United States, the White House issued an executive order in 2020 directing federal agencies to apply existing laws to AI, but it created no new enforcement body. A company that wanted to ship a credit scoring model could do so with a legal review and a privacy notice.

That is not the world in 2026. The EU AI Act's obligations for general-purpose AI models took effect in August 2025, and the high-risk provisions are rolling out on a published schedule. The US has moved through the 2023 executive order, its rescission in January 2025, and a patchwork of state laws (Colorado's AI Act, California's transparency rules, and others). China's rules on generative AI took effect in 2023. The shift is not that every country now has one law. It is that AI systems increasingly have a compliance file.

The three things that changed#

Audits replaced checklists. In 2021, an internal review might mean a data protection impact assessment adapted for a model. Now, in the EU, providers of high-risk systems must maintain a risk management system, technical documentation, and post-market monitoring. Regulators can request that documentation. In the US, the National Institute of Standards and Technology's AI Risk Management Framework, released in 2023, is voluntary, but contracts and state laws increasingly cite it. The practical result: an AI product team now budgets for an audit trail the way a payments team budgets for a SOC 2 report.

Transparency requirements hit the model layer. Five years ago, disclosure focused on the user: tell people they are talking to a bot. Today, the EU AI Act requires providers of general-purpose models to publish a training data summary and respect copyright opt-outs. Several US states require disclosure when AI generates political content. The burden has moved upstream, to the companies that train and fine-tune models, not only to the apps built on top.

Fines got specific. In 2021, the largest AI-related penalties came from privacy regulators applying existing data laws (the Italian Garante's 2023 fine against a chatbot maker, for example, was a GDPR action). The EU AI Act sets penalties up to 35 million euros or 7 percent of global turnover for prohibited practices, and up to 15 million euros or 3 percent for other violations. Those numbers are not yet tested in court, but compliance officers read them as a price list.

Where the Gaps Are#

A rulebook is not the same as enforcement. The EU AI Act's high-risk obligations for most systems apply from August 2026, with some later. National supervisory authorities in many member states are still being staffed. The US has no federal AI statute; state attorneys general enforce state laws, and their capacity varies widely. India's guidance is still largely advisory. Brazil's bill has moved slowly through its legislature.

There is also a measurement problem. A 2025 study by researchers at Stanford's Center for Research on Foundation Models found that many model cards and transparency reports omit the information the rules ask for, such as data provenance and evaluation results on specific subgroups. The rules assume companies can produce evidence. In practice, the evidence is uneven.

What to watch next#

Three numbers are worth tracking. First, the count of formal AI Act investigations opened by EU authorities, expected to rise after August 2026. Second, the number of US states with AI laws that include private rights of action, which turn compliance failures into lawsuits. Third, the share of foundation model providers that publish training data summaries; the first deadlines for that requirement passed in 2025, and compliance is public.

The comparison to 2021 is not that AI policy failed or succeeded. It is that the question changed. Five years ago, the debate was whether to regulate. Now it is whether the rules can be checked, and by whom.

More articles