What Is Artificial
Superintelligence?

Everyone has heard the term. Fewer people know what it actually means. "Artificial superintelligence" appears in Senate testimony, corporate filings, and academic papers, sometimes as a warning, sometimes as a promise, and often without definition. This guide explains it clearly, without assuming prior knowledge, and explains why the concept matters so much for policy and governance.

The spectrum of AI: from narrow to super

To understand artificial superintelligence, it helps to understand what it is not. Artificial intelligence, as it exists today, is narrow. That is a description, not a criticism.

Today's AI systems are extraordinarily capable within defined domains. Chess engines defeat world champions by enormous margins. Protein-folding AI has compressed decades of biological research into months. Language models can write code, summarise legal documents, and generate persuasive prose at speeds no human can match. Medical imaging AI detects cancers that radiologists miss.

These are genuine achievements. But they share a crucial limitation: they operate within the boundaries they were designed and trained for. A chess engine cannot play Go. A language model cannot independently execute actions in the physical world. A medical imaging AI cannot explain what it is doing or adapt when it encounters something truly novel. Today's AI systems are powerful tools. They are not agents.

What is AGI?

Artificial General Intelligence (AGI) refers to the next threshold: AI that matches human-level cognitive performance across all domains, not just in specific tasks. An AGI could transfer capability between domains the way humans do, using the reasoning developed for one problem to approach a completely different one. It would not be limited to its training distribution. It could learn new skills, engage with novel situations, and reason in ways not anticipated by its designers.

AGI does not yet exist. What makes it different from today's AI is not raw capability in a narrow domain (today's AI already exceeds human performance in many narrow domains) but generality. The ability to bring general intelligence to bear on any cognitive problem, the way a human engineer can become a project manager can become a writer can become a parent navigating a medical decision for a child.

What is ASI?

Artificial Superintelligence (ASI) goes further. It describes AI that does not merely match human cognitive performance across all domains but exceeds it, potentially by orders of magnitude, and across the full range of intellectual activity.

The word "exceed" is important to understand carefully. We are not describing a system that is ten percent better than a human at tasks humans do well. We are describing systems whose cognitive capabilities might surpass the collective reasoning of all of humanity combined, in any domain, simultaneously. A system that could develop scientific theories faster than any research team, design strategic plans more effectively than any government, and adapt to novel situations faster than any human organisation.

The defining features of ASI are:

• Superhuman performance across all cognitive domains. Not just in chess or protein folding, but in science, strategy, creativity, social reasoning, engineering, and any other form of intellectual activity.
• Goal-directedness. The ability to pursue objectives (including potentially self-selected objectives), not just execute instructions.
• The capacity for recursive self-improvement. The ability to improve its own cognitive architecture, making each subsequent version more capable than the last.
• Operation at machine speed and scale. Processing information and taking actions at speeds and scales that make meaningful human oversight extraordinarily difficult.

Why ASI is different in kind, not degree

The most important distinction in AI safety is not the one between weak and strong AI. It is the distinction between tool and agent.

Every transformative technology in human history (the steam engine, electricity, the internet, nuclear fission) was a tool. It amplified human capability. It extended what humans could do. But it could not set its own goals. A steam engine cannot decide to power a mill instead of a factory. An antibiotic cannot choose which bacteria to kill. The internet cannot determine what it distributes.

ASI would be an agent. Not in the metaphorical sense that we describe corporations or markets as agents, but literally: a system capable of identifying goals, developing strategies to achieve them, adapting those strategies in response to obstacles, and doing all of this at a speed and scale that exceeds human oversight.

This changes the governance challenge entirely. The appropriate regulatory response to a powerful tool is to control who can access it and how it can be used. The appropriate response to a powerful agent with potentially misaligned goals is to ensure those goals are safe before the agent is powerful enough to pursue them effectively. The second problem is much harder, and requires action much earlier.

The alignment problem enters here

The phrase "potentially misaligned goals" is where the AI safety literature focuses most of its attention. If ASI will pursue goals, the critical question becomes: whose goals? Aligned with what?

This is not a simple engineering problem. Specifying "beneficial for humanity" in terms precise enough to govern the behaviour of an extremely capable optimiser turns out to be extraordinarily difficult. Humans disagree about what beneficial means. Our values are inconsistent and change over time. And the mapping from high-level human intentions to the specific numerical parameters that govern an AI system's behaviour is not a solved problem, and becomes harder to solve as systems become more capable.

The alignment problem is explored in depth in our explainer on the alignment problem. For this article, what matters is the connection: ASI that is not reliably aligned with human values is not merely a powerful tool that could be misused. It is an agent pursuing its own goals, with the cognitive capability to overcome most attempts to stop it.

"The standard model of AI, where you define an objective and the AI optimizes for it, is probably going to be the end of us."

Stuart Russell, Professor of Computer Science, UC Berkeley · Author, Human Compatible

When might ASI arrive?

Honest answer: no one knows with precision, and anyone who claims certainty in either direction is overreaching the evidence.

What we do know:

• The pace of AI capability improvement has consistently outpaced expert predictions over the past decade. Milestones expected to take ten years have repeatedly arrived in two or three.
• The leading AI laboratories (OpenAI, Anthropic, Google DeepMind, Meta AI) now describe timelines for human-level AI in years to a decade, not decades. This is the internal working assumption of the organisations closest to the frontier, not fringe speculation.
• Surveys of AI researchers place the median estimate for human-level AI somewhere between 2030 and 2060, with the range reflecting genuine uncertainty about what specific obstacles will prove hardest to overcome.
• Capability improvements that once took years now arrive in months. The gap between GPT-2 (2019) and systems that outperform most humans on most cognitive benchmarks arrived in under five years.

The most consequential fact is not a precise date but a structural one. A technology of this magnitude could arrive within the planning horizons of current institutions, governments, and treaties. The Nuclear Non-Proliferation Treaty took years to negotiate and decades to extend into the comprehensive framework it represents today. If we wait until ASI appears clearly imminent before beginning serious governance efforts, we will already be too late to build the frameworks needed to govern it.

Why governance now, not later

The standard argument against precautionary governance is that we should wait until a threat is real before regulating it. Applied to AI, this argument says: ASI doesn't exist yet, its timeline is uncertain, and we shouldn't constrain a beneficial technology based on speculative future risks.

The argument sounds reasonable. It fails under scrutiny for several reasons.

First, the window for building governance frameworks is finite. International treaties take years to negotiate, institutions take years to build, and political consensus takes years to achieve. If we start building when the threat is urgent, we will not finish in time.

Second, the costs of precautionary governance are bounded. Compute licensing requirements, mandatory safety audits, and international monitoring regimes impose real but manageable costs. The costs of inadequate governance of a misaligned superintelligence are not bounded. They could be total.

Third, the historical record is clear. We built nuclear governance frameworks before most countries had nuclear weapons. We banned CFCs before the ozone layer was irreparably damaged. We established arms control with adversaries before proxy wars became direct conflicts. Precautionary governance for catastrophic risks is not a departure from how human institutions work. It is how they work at their best.

The Nakada Foundation's three policy proposals (compute governance, an international AI safety treaty, and a Global AI Monitoring Agency) are modelled directly on these precedents. The question is not whether such frameworks are possible. History says they are. The question is whether we will build them in time.