Artificial Superintelligence
Artificial superintelligence is a hypothetical AI system whose capabilities exceed the best human performance across essentially every intellectual domain — not merely matching people, which defines artificial general intelligence, but surpassing them by a margin that would make its judgement better than any person's or institution's. It does not exist; whether it could is contested; and the idea shapes the labs' missions, safety research and funding.
- Also known as
- ASI, Superintelligence, Superhuman AI
- Used at
- the mission statements of OpenAI, Anthropic and DeepMind; the alignment and safety research agendas; policy debates on AI risk; and the investment theses of the funds that back the labs — and nowhere in a fintech's product roadmap
- Binding
- no law defines it; regulation (the EU AI Act, the US executive orders, the UK's institute) addresses capable general-purpose models by compute thresholds and evaluations rather than by the term
- Typical arguments for taking it seriously
- capability has scaled with compute; systems already exceed humans in narrow domains; an AI that can do AI research could accelerate its own improvement; the downside of a misaligned superintelligence is unbounded. Arguments against: no evidence that scaling continues indefinitely; general intelligence may need what current methods lack; the term encourages speculation over engineering
- Typical practical consequence
- the labs' safety teams, red-teaming and "responsible scaling" policies that tie deployment to evaluations; the governance structures (a nonprofit board, a public-benefit charter, a long-term benefit trust) meant to keep the mission ahead of commercial pressure — and the governance crises when the two collide
- Parties
- the labs and their researchers, the investors, governments, the academic and effective-altruism communities that framed the risk, and the sceptics who think the debate distracts from present harms
How the idea works in practice
Nobody is building a superintelligence on purpose; what the labs are building are models that improve with scale, and the question is where that stops. The optimistic version: models that automate research, medicine and engineering, with the wealth to match; the pessimistic: a system pursuing objectives its makers did not intend, at a capability no one can correct — the alignment problem at its limit. Both versions are why the labs fund safety research, write scaling policies that require evaluations before the next model is deployed, and set up governance meant to let a board stop a launch. The 2023 board crisis at OpenAI was that governance being tested; Anthropic's structure — a public-benefit corporation with a long-term benefit trust — is another attempt.
For the AI economy the idea sets the stakes and the valuations: the labs raise tens of billions on the claim that they are building the most important technology in history, and the compute those billions buy is what produces the models everyone else uses. For policy it sets the frame: compute thresholds, evaluation requirements, safety institutes, and an argument between those who see existential risk and those who see present harms — bias, fraud, labour displacement — being crowded out by science fiction.
For a fintech the term is context, not a plan: the models it uses are products of labs whose missions mention it, whose safety policies constrain what those models will do, and whose governance decides what the next model is.
Example
A fintech's board asks its CTO, after reading a lab's announcement, whether the company's AI strategy should account for superintelligence. The CTO's answer is a short memo: the fintech uses two vendors' models over APIs, chosen on quality, cost, data terms and the vendors' safety practices; those practices include evaluations and staged deployment tied to the vendors' own scaling policies; the fintech's exposure is to the vendors' decisions and prices, and its mitigation is not to depend on one; and nothing in its fraud, credit or support products would change if the models were twice as capable next year except that the guardrails would matter more. The board files the memo and adds vendor concentration to the risk register, which is the useful outcome.
Common mistakes
- Reading a lab's mission statement as a product roadmap. - Building a business strategy around a capability that does not exist and a date nobody can give. - Dismissing safety practices that constrain today's models because the term that motivates them sounds speculative. - Treating present-day harms and long-term risk as rivals rather than as both real. - Boards asking about superintelligence and not about which vendor's terms let it train on the company's data.