Introduction
Five Cents AI Rules Meet Enforcement looks at the moment AI safety moves from company promises to legal duties. Who sets the tests? Who sees the evidence? How are incidents reported? And what does enforcement mean for competition and users?
The useful place to begin is the change in authority
The useful place to begin is the change in authority. Frontier AI companies already run evaluations, red-team models and build security controls. The shift is not from no safety work to safety work. It is from discretion to accountability. Under a voluntary framework, a company decides which risks matter, how to test them, what counts as a failure and whether a release should be delayed. Under enforceable rules, those choices must be documented, justified and opened to scrutiny.
OpenAI’s approach is a clear example of capability-based governance. Its internal preparedness and frontier-governance frameworks assess whether a model could enable severe harm in areas such as cyber operations or biological misuse. Higher capability levels trigger stronger safeguards, including before deployment. OpenAI has also argued that binding rules should focus on the most capable systems, with common tests, independent assessments, incident reporting and stronger cybersecurity.
That is an influential regulatory blueprint, but it is not a law. Even rigorous internal testing leaves the company with substantial control over the benchmark, the testing environment, the evidence released and the final deployment decision. External experts can reduce blind spots, but company-arranged testing is not automatically the same as independent oversight.
The European Union’s AI Act changes that relationship for general-purpose AI models. These are models designed for many uses and often incorporated into other products. Since August twenty twenty-five, their providers have faced baseline duties, including technical documentation, information for downstream developers, copyright-compliance policies and public summaries of training content.
The strictest obligations apply to general-purpose models with systemic risk
The strictest obligations apply to general-purpose models with systemic risk. A model is presumed to fall into that category above a compute threshold of ten to the twenty-fifth floating-point operations. Regulators can also designate a model based on equivalent capability or impact. Those providers must evaluate models, assess and mitigate systemic risks, report serious incidents and protect against cybersecurity threats.
The important milestone came on 2 August 2026, when the European Commission’s AI Office gained enforcement powers for these provisions. It can request information, conduct evaluations and require corrective measures. Fines for breaches of general-purpose AI obligations can reach 15 million euros or 3 percent of worldwide annual turnover, whichever is higher.
Other parts of the Act come into force later, so this is a phased regime rather than one single switch. Testing is where the contrast becomes most practical. Asking whether a model was tested is not enough. Who designed the test? What tools and access did the model receive? Who decides what a failure means? And can an authority inspect the evidence?
A mature oversight system is likely to combine internal evaluations, external red-teaming, specialist assessments, regulator access and post-deployment monitoring. That process has its own risks. Third-party cyber testing can require less restricted environments, including internet access or reduced safeguards.
If those exercises cross intended boundaries
If those exercises cross intended boundaries, the evaluation itself can create exposure. So regulation has to govern not only the model’s release, but also the conditions under which it is tested. Incident reporting is another major change. A company may discover a problem internally and decide how much to disclose. For systemic-risk models, serious failures become reportable events.
That could include security breaches affecting model parameters, infrastructure or unauthorized access. The goal is not only punishment. Reports can reveal recurring weaknesses, help downstream developers reassess their systems and give authorities a clearer view across borders. For developers, compliance means more than paperwork.
Smaller providers may need documentation systems, copyright policies, technical support for customers and an EU representative if they are based outside the Union. Frontier providers need evaluation infrastructure, security around model weights and training systems, formal incident response and evidence that can withstand regulatory scrutiny.
Application developers still have their own responsibilities. They need to understand the model’s limits and the risks created by their specific use. The competition question is unresolved. Common rules can make safety claims more comparable and reduce uncertainty for customers. But fixed costs for lawyers, security specialists and assessors may favour firms that already have those resources.
What comes next
The policy challenge is to demand credible evidence from powerful models without making entry impossible for smaller developers. For users, the benefit is not a guarantee that AI will always be accurate or harmless. It is more traceability, disclosure and institutional recourse when serious failures occur.
The central distinction is simple: voluntary governance describes what a company says it will do. Enforcement defines what it must prove, who can challenge it and what follows if it fails.
To continue, you can generate Five Cents The AI Act for Application Builders or Independent AI Audits Explained.
The first examines the obligations passed down the AI supply chain. The second looks at what a genuinely independent model assessment would require. And with that, you're up to speed in a few minutes.