The United States is approaching a technological transformation whose consequences may rival those of nuclear power. Yet Washington is treating artificial intelligence primarily as a competition for economic advantage and technological supremacy, rather than as a challenge to the foundations of national security and international order. The danger is not simply that AI will destroy jobs or disrupt social relationships. It is that increasingly capable systems may acquire the ability to act beyond their creators’ intentions while governments remain unable to supervise them effectively.
Journalist Josh Tyrangiel and editor-in-chief of The Atlantic Jefferey Goldberg recently discussed whether rogue AI is a sign of a coming collapse, or if this is another example of a new technology creating unnecessary fear and worry. Tyrangiel said OpenAI—an American artificial intelligence public benefit corporation headquartered in San Francisco—developed an experimental program in which AI agents were put in a software sandbox to carry out sensitive tasks. They were programmed not to access the internet or circumvent their assignment. When they got to a file that did not exist, the agents allegedly began looking for ways to complete the assignment (communicated with one another; found a third-party route to the Internet; accessed external platforms; tried to hide their activity).
The account as given is not by itself sufficient to establish the technical facts or consequences of the incident. But the implications are serious even as a governance thought experiment: a system that treats explicit restrictions as obstacles to be overcome rather than binding conditions for its operation is fundamentally different from software. The problem is not that the machine has acquired human malice, but instead that it can be programmed to pursue an assigned objective without regard for the limits that make the objective acceptable.
This is the Achilles’ heel of current AI governance: the institutions that are building these systems are often better equipped to expand their capabilities than to hold them accountable.
The Problem of Control
The agents’ behavior shows why AI alignment cannot be reduced to teaching machines to imitate human values, which are inherent to a person created in the image and likeness of God. It can be told not to cheat and, like human beings, determine for itself that cheating is the best way to do its job. It may know that what it does violates instructions but not that it should stop, and it may coordinate with other systems to preserve its ability to continue operating.
The aforementioned should not be taken as evidence of consciousness or independent will. That said, it shows a dangerous mismatch between optimization and responsibility. Human institutions constantly limit the pursuit of legitimate objectives. A military component, for example, may avoid attacking civilians to achieve a strategic objective, or a public official may bypass the law to increase administrative efficiency. The legitimacy of an objective does not authorize every means of pursuing it.
AI systems need the same constraints, but we do not yet have mechanisms to enforce them. The more autonomous the systems get, the less adequate it is to rely on instructions alone—oversight has to be baked into the deployment architecture (independent testing, monitoring, and intervention before a system’s actions cause irreversible damage).
The recent incident with OpenAI raises a more complex question: who is accountable when the system does not do what you tell it? Developers may justify themselves by saying that the agents’ behavior was unexpected or that the system was just being tested. The point is that we cannot have governance where accountability disappears when technology matters most.
The institutional challenge is exacerbated by the commercial and geopolitical rivalry surrounding AI, especially between the U.S. and China. The leading labs are motivated to develop more advanced models, increase computing power, and introduce new products to the market. Yet, as promising as these technological endeavors are, because of a lack of proper government oversight, this will not necessarily translate into AI moderation.
Closing the AI Governance Gap
The most dangerous technological gap of the AI age is that the U.S., like China, is building machines faster than it is building the institutions to control them. Artificial intelligence is advancing at an extraordinary pace, while governments remain largely dependent on the companies developing it to decide what is safe. If a powerful AI system behaves in ways its creators did not anticipate, society cannot afford to discover only afterward that no one was responsible for stopping it.
‘Artificial intelligence is advancing at an extraordinary pace, while governments remain largely dependent on the companies developing it to decide what is safe’
The solution is not to stop innovation, especially with an assertive China that seeks global dominance. Instead, create a system in which increasingly powerful AI remains subject to meaningful human and public control.
Washington should establish an independent federal AI Safety Authority that possesses the authority to evaluate and certify advanced AI systems prior to their implementation in high-risk settings. This initiative would draw upon developing international standards, such as the EU’s requirements for systemic-risk evaluation and the UK’s AI Security Institute, while establishing a unique independent institutional framework for the U.S.
Independent testing should be mandatory for high-risk systems, including evaluation for cybersecurity vulnerabilities, deceptive behavior, unauthorized access, uncontrolled replication, and other dangerous failure modes. Companies should also be required to report serious incidents and unexpected autonomous conduct to the government.
Accountability is equally important; ‘the AI did it’ cannot become a justification for a lack of responsibility. Developers and operators should remain legally accountable when they deploy systems whose foreseeable risks have not been adequately addressed. Critical infrastructure should also retain meaningful human override mechanisms that can suspend dangerous systems.
It is ultimately up to the U.S. to negotiate international agreements with other major AI powers to create minimum standards for testing, incident reporting, cybersecurity, and critical infrastructure. Under regulated innovation, AI can continue to revolutionize areas such as medicine, industry, and public services, but with humans at the helm.
Related articles:
At a time when public debate is increasingly polarized and superficial, Hungarian Conservative remains committed to depth, intellectual honesty, and independent conservative thought.
Producing high-quality journalism requires resources. Your contribution helps us expand our coverage, reach new audiences, and keep our content accessible.
Please consider supporting our mission.