Products for the public, war machines and IT weapons, digital mentoring in school contexts, straightforward systems in logistics: these are examples of deployment opportunities for AI systems. All of them draw from the same capability frontier, which is currently being expanded by researchers across the globe. This frontier is, in itself, neutral — it simply describes what AI systems can technically achieve.
Naturally, this raises a question: is unregulated AI research a good idea?
Before discussing regulation, an important distinction must be made between AI research and the deployment of AI systems.
Research primarily takes place in universities, research institutes and laboratories. Its outputs are scientific knowledge, algorithms, datasets and models. Deployment, by contrast, occurs when these outputs are integrated into products, infrastructures or decision-making systems. Most current AI regulation, such as the EU AI Act, primarily targets deployment rather than research.
The boundary between research and deployment is not perfectly sharp — releasing open model weights, for instance, can enable independent users to build products upon powerful models. Nevertheless, the distinction matters because different types of risks, and therefore different regulatory strategies, emerge at each stage.
Many of the arguments that appear in debates about regulating AI research are not unique to artificial intelligence. They reflect longstanding questions about the governance of scientific research in general. In the following, these broader considerations are applied to the specific context of AI.
For this discussion to matter, one thing must be true: research itself can produce risks, not just deployment. And it can.
The Dual-Use Nature of AI Research. A fundamental concern is the dual-use nature of AI research. The same scientific capabilities can serve both beneficial and harmful purposes. Advances in machine learning may enable medical diagnostics, scientific discovery and improved logistics. At the same time, similar capabilities can contribute to cyber attacks, large-scale surveillance or autonomous weapon systems — including military applications that raise serious ethical and legal concerns.
Alignment and Control Risks. A second concern is not about misuse but about unintended behavior. When AI systems optimize objectives that are incorrectly specified, they may pursue those objectives with extreme efficiency while producing harmful outcomes. Russell describes this as the "King Midas problem": an intelligent system achieves its stated objective perfectly while failing to capture what humans actually intended. From this perspective, regulating certain forms of research could be seen as a precautionary measure against the emergence of highly capable systems whose behavior becomes difficult to predict or control.
Externalities and Public Interest. Finally, AI research does not occur in a vacuum. Large-scale training processes depend on data centers with significant energy consumption and environmental footprints. If research activities impose substantial environmental, economic or social costs, it seems reasonable that such externalities should be subject to oversight and governance. In these cases, regulation would not target knowledge itself, but the broader societal costs associated with its production.
The arguments above address risks that arise from AI research. The arguments against regulation, by contrast, target the proposed remedy — and highlight the costs it may introduce.
First, complex regulatory frameworks may introduce bureaucratic barriers. Large corporations typically have the resources to comply with extensive requirements, while universities and independent researchers may not. This pattern is visible even within government: empirical assessments of AI governance in U.S. federal agencies show that implementation gaps are particularly pronounced in organizations with fewer resources. This suggests that regulatory capacity constraints would likely be even more severe for smaller private research institutions.
Second, research is inherently global. If strict regulations are introduced in one region, research may simply continue unregulated in jurisdictions with fewer restrictions. The technological development then proceeds, while the regulating region loses influence over it.
Third, regulation may slow down experimentation and innovation. Delayed scientific progress could postpone beneficial applications, including those addressing major global challenges such as climate change or disease.
Taken together, these arguments point toward an important asymmetry between regulating research and regulating deployment.
Regulating the deployment of AI systems — whether regionally or globally — appears both feasible and desirable. Many societal risks, including discrimination, disinformation, surveillance and autonomous weapons, emerge primarily when technologies are embedded into real-world institutions. In these contexts, legal oversight and democratic accountability are both necessary and practical.
Regulating the production of scientific knowledge, however, is considerably harder. Regional restrictions on research may be justified where they address local externalities, such as the environmental costs of large-scale computation. But for global risks — such as the emergence of misaligned or uncontrollable systems — regional regulation offers little protection: such systems could be developed in unregulated jurisdictions and still pose risks worldwide. In these cases, the regulating region limits its own research capacity without reducing the underlying danger.
Globally coordinated regulation could, in principle, address some of these challenges. In practice, however, such agreements are difficult to establish and enforce in a competitive geopolitical environment.
For this reason, a more realistic approach may focus less on formal restrictions on research and more on strengthening professional norms and ethical responsibilities within the research community itself.
A recurring theme across the literature is a call for awareness — for upholding sound ethical values within the research community and raising broader public attention to the implications of AI development. Given the practical limitations of regulating knowledge production, this may be the most important lever available.
Regulating AI research in a way that is both effective and fair would require an enormous sustained effort — one that, given current geopolitical realities and the many open conflicts around the world, may not be realistic. It may therefore be more productive to invest in other approaches: creating incentives for responsible development, strengthening international dialogue, and persuading the institutions that fund and direct AI research that embedding these technologies in ways that promote peace and human flourishing is ultimately more valuable than pursuing narrow strategic advantages.