AI is changing the world. But who gets to control it?

AI is changing the world. But who gets to control it?

By Shriyaa Agarwal

In the rapidly evolving world of artificial intelligence, a new battle for control is beginning to emerge – and it is happening between the companies building AI and the governments trying to regulate it. In February 2026, Anthropic (the company behind the creation of Claude) refused Pentagon’s demands for unrestricted military use of its AI. The Trump administration responded by ordering federal agencies to stop using Anthropic’s technology and the Pentagon designated the company a national-security “supply-chain risk”. However, on 28th August, a federal judge rules that the Pentagon’s actions were unlawful and blocked the designation. So, who gets to decide what AI can be used for: the government (who is accountable to the public), or the company (who understands the technology the best)? And if neither should have complete control, who should?

This question matters because the AI considered here is not simply the internal software that a company uses. It is global AI: highly capable, general-purpose AI whose development and deployment can affect people, economies and societies beyond the organisation that created it. Throughout this article, “regulation” refers to the formal rules and enforcement mechanisms used to constrain or direct the development and deployment of these systems, rather than the broader concept of AI governance, which can include organizational practices, technical tools, and ethical principles as well as law.

Should AI companies have the final say

At first glance, there is a strong argument for AI companies to have substantial control over the systems they develop. The researcher building the frontier models are likely to have the most understanding of their capabilities, limits and risks. This is crucial as AI development is happing extremely quickly, and traditional legislation can take up to years to design and implement. Companies are therefore better positioned to respond to new safety risks much faster than governments as well as possessing the technical ability to test models and put in safeguards directly. This expertise is clearly visible in the leading companies of the AI industry, more informally referred to as MANGOS (Meta, Anthropic, Nvidia, Google, OpenAI and SpaceX) – these companies occupy different parts of the AI ecosystem, from frontier models and computing hardware to infrastructure and distribution. To regulate global AI effectively, ignoring the expertise concentrated within these companies would make little sense.

However, technical expertise cannot equal legitimate authority. An AI company may be better qualified to determine whether a model has a particular capability, but this does not mean that it has the right to determine whether society ought to accept the risks associated with that capability. There is also an inevitable conflict of incentives: a company while being responsible for the technology’s safety are also the ones who will benefit the most by releasing a more capable system quickly. This conflict is further deepened by the ongoing AI rat race. The question is not whether these companies are trustworthy or not; rather it is whether it makes sense for them to hold unilateral authority over technology whose consequences extend far beyond their shareholders. Recent events such as the EU’s AI Act demonstrates why this distinction matters. This act now gives the AI Office specific enforcement powers over general-purpose AI models, which emphasises that technical expertise alone is no longer being treated as sufficient as governments are beginning to exercise this form of authority.

Should governments have the final say?

Governments have a different advantage: democratic legitimacy. Governments are, at least in principle, accountable to the populations affected by their decisions. This matters because some AI governance decisions are not just technical questions but questions about what risks and uses society should permit. Especially as AI can affect areas such as employment, education, healthcare, elections, public services, etc. The EU serves as an important example of this approach. Its AI Act is already being enforced for general-purpose AI, while further rules for high-risk AI systems are scheduled to apply in December 2027 and August 2028, depending on the type of system. This shows how governments are attempting to establish boundaries around AI before the consequences are dire.

Yet, while regulators may have the authority to decide whether an AI system should be permitted, they may not understand the technology well enough to evaluate its risks accurately. There is an information asymmetry where companies know considerably more about the systems they develop than the regulators attempting to oversee them. Poorly designed regulation could become outdated, restrict beneficial innovation, allow more dangerous competitors to get ahead or focus on the wrong risks. This is why simply transferring all authority from companies to governments would not be an adequate solution to this problem. The solution should therefore not choose between expertise and democratic legitimacy, rather to find a way of combining them.

The solution

Perhaps, then, the answer is to create an independent AI regulatory body, such as the role that the FDA plays in the medical sector The FDA brings scientific and technical expertise into a system of government regulation; a similar institution for global AI could combine the two forms of authority that are currently separated. The body would include representatives from AI companies and independent technical experts, who could explain how systems work, assess capabilities and identify technical risks. Alongside them would be government representatives who could represent societal interests, establish legal boundaries and determine what level of risk is acceptable. Importantly, neither side would have complete control. Companies would provide the expertise necessary to make informed decisions, while democratic representatives would ensure that technical knowledge does not become a substitute for public accountability. This model could also adapt as AI regulation develops internationally (perhaps even play a role in creating that regulation). The EU's current system is already moving towards a mixture of government enforcement and technical expertise: its Scientific Panel provides independent scientific and technical expertise to support enforcement by the AI Office. Future regulation could thus build on this principle.

The strongest model, therefore, may be one in which expertise informs authority rather than replacing it. No single actor should control global AI simply because it possesses either the most knowledge or the most power. As AI continues to change the world, the challenge is not to decide who controls it, but to ensure that those with the knowledge, authority and responsibility to shape it work together to make it safe for everyone.