Risks have been at the centre of AI debates for a long time. They reflect our hopes, fears, and uncertainties about a technology that is already changing how we work, learn, and communicate.

Today, AI risks can be understood through three broad categories: existential risks, existing risks, and exclusion risks. Each category operates on a different timeline. Existential risks remain more distant and uncertain, but they continue to dominate public imagination. Exclusion risks are likely to deepen over the coming years. Existing risks are already visible. By following how AI risks are discussed, we can see how policy priorities are shifting.

Existential risks: The fear of losing control

Existential risks are the most dramatic and most attractive in public debate. They concern the possibility that future super-intelligent AI systems could move from being human tools to autonomous powers beyond human control. In this scenario, AI would no longer serve humanity; it could become a force that shapes or even endangers human survival.

These risks are often compared to nuclear war, pandemics, or climate catastrophe. As they remain ‘unknown unknowns’, it is difficult to envisage what form they might take, when they might emerge, or how likely they are. This uncertainty makes governance especially difficult.

Yet uncertainty is not a reason for inaction. Climate governance shows that societies can prepare for deep uncertainty through scenario-building, precautionary thinking, early-warning systems, and international cooperation. The same approach is needed for AI. Existential risk should not paralyse policy, but it should encourage humility, preparedness, and long-term thinking.

Existing risks: The harms already here

Existing risks are part of our reality: job displacement, data protection breaches, misuse of copyrighted material, algorithmic discrimination, loss of human agency, cybersecurity threats, and the mass production of synthetic text, audio, images, and video.

AI is also creating new challenges in education. Students and teachers are struggling to define what counts as learning, authorship, cheating, or legitimate assistance. In public communication, synthetic content can flood the information space, increasing confusion and weakening trust.

The good news is that many of these risks do not require entirely new legal systems. Existing instruments can already do much of the work: privacy law, consumer protection, anti-discrimination rules, labour regulation, intellectual-property law, and cybersecurity frameworks.

The challenge is implementation. In some cases, new AI-specific mechanisms will be needed, especially for transparency, auditing, accountability, and risk assessment. But the starting point should be practical: use what already exists, identify the gaps, and build targeted instruments where necessary.

Exclusion risks: The danger of AI monopolies

Between immediate harms and distant existential threats lies a third category: exclusion risk. This is the risk that AI will concentrate economic, technological, and knowledge power in the hands of a few companies and countries.

Advanced AI depends on three scarce resources: data, computing power, and specialised know-how. Today, these resources are heavily concentrated. A few companies control the leading models, the infrastructure needed to train them, and many of the data streams that improve them. This concentration could shape markets, public policy, education, culture, and everyday life.

The danger is not only economic monopoly. It is also a knowledge monopoly. If a few actors control the tools through which societies search, write, translate, analyse, and decide, they will gain influence over how knowledge itself is produced and distributed.

This risk, which will sharpen in the coming years, will create sharper divisions between a few AI-makers and many AI-takers, dependent on systems designed elsewhere, trained on data they do not control, and aligned with values they did not help define.

The policy tools are already known: competition law, antitrust enforcement, data governance, open standards, public-interest infrastructure, intellectual property safeguards, and investment in local AI capacity. What is missing is often a political priority. Exclusion risk must be recognised before concentration becomes irreversible.

From Bletchley to New Delhi: The evolution of AI risk debates

The evolution of AI risks can be followed through the agenda of AI summits held over the last few years. The first AI Safety Summit at Bletchley Park placed strong emphasis on frontier AI and catastrophic or existential risks. This focus was understandable. Governments needed to acknowledge that advanced AI could pose risks beyond those addressed by ordinary technology regulation.

Subsequent summits broadened the agenda. Seoul placed greater emphasis on safe, innovative, and inclusive AI. Paris shifted the language from safety to action, highlighting implementation, inclusion, sustainability, and economic opportunity. New Delhi continued this movement by focusing on the impact of AI, especially on inclusion, development, and practical deployment.

This does not mean that risks have disappeared from the agenda. Rather, the discussion has become more balanced. The initial fear of existential danger has been joined by a more practical concern with present harms and future inequalities.

Towards a holistic approach to AI risks

Existential, existing, and exclusion risks must be addressed simultaneously with an adjusted focus as per technological developments.

For existing risks, governments should deploy and update current legal instruments. Data protection authorities, consumer-protection agencies, courts, regulators, schools, and employers already have roles to play. The main task is to make these institutions AI-ready.

For exclusion risks, governments should act before AI monopolies become structurally embedded. This means stronger competition policy, better data governance, support for open and interoperable AI ecosystems, and investment in public and local AI capacities.

For existential risks, governments and societies need precautionary governance. This includes scenario planning, scientific assessment, international cooperation, testing regimes, and mechanisms to pause or constrain systems that may pose unacceptable risks.