Every few months, an AI capability benchmark falls. Systems that could not pass professional licensing exams two years ago now score in the top percentiles. The executives of the leading frontier AI laboratories publicly describe artificial general intelligence as a matter of years, not decades. One of the most prominent AI safety researchers, Nick Bostrom, published a 2026 working paper asking not whether to develop superintelligence, but when.

Against this backdrop, the question of whether we are ready for superintelligence is not a theoretical exercise. It is an urgent practical assessment, and the answer is clearly no.

What readiness would actually require

It is worth being specific about what "readiness" means, because the word is often used loosely to mean "we have thought about the risks" or "responsible labs are working on safety." Those things are not readiness in the relevant sense.

Genuine readiness for superintelligence would require all of the following to be in place before a system at ASI capability is deployed anywhere:

  • A reliable technical method for verifying AI alignment before deployment, not inferring it from behavioral evaluations
  • International governance with enough reach to prevent any single actor (company or nation) from deploying a system before verification passes
  • Monitoring infrastructure operated independently of the organizations building the systems
  • Legal frameworks in major AI-developing countries that give oversight bodies actual authority
  • Established international norms with enough political weight to survive competitive pressure

Assessed against these requirements, here is where we actually are.

Technical alignment
Unsolved
Interpretability research is advancing but cannot reliably verify the internal goals of frontier-scale systems. Deceptive alignment has been documented in current models. Standard safety training failed to remove it in the Anthropic Sleeper Agents study.
International governance
Early stage
The Bletchley Park Declaration (2023) established international consensus on the existence of the risk. No binding treaty exists. The 2025 Paris Action Summit shifted the focus from safety to deployment and economic competition.
Independent monitoring
Does not exist
No international body has authority to audit frontier AI systems. Safety evaluations are conducted by the organizations developing the systems, with results that are often not independently verifiable.
National legislation
Fragmented
The EU AI Act establishes risk-tiered regulation for AI systems but does not specifically address AGI or ASI. The US has executive orders but no binding federal AI legislation. China, the UK, and other major AI countries each have partial frameworks.

The governance conversation has moved in the wrong direction

In November 2023, the UK government hosted the Bletchley Park AI Safety Summit, the first major international gathering specifically focused on frontier AI risk. The framing was explicitly safety-first. The resulting Bletchley Declaration was signed by 28 countries, including the US, China, and the EU, acknowledging "potentially catastrophic" risks from advanced AI.

By February 2025, France hosted what it called an AI Action Summit, deliberately renaming the format to signal a shift in priorities. The emphasis moved from identifying and managing risk to enabling deployment and capturing economic advantage. This was not a neutral change in terminology. It reflected genuine political pressure from governments and industries that see AI competition with China as the primary frame, and safety requirements as potential competitive disadvantages.

"The question is no longer whether to develop superintelligence, but when it is optimal to do so, and for whom the benefits of that timing decision accrue."

Nick Bostrom, "On the Timing of Superintelligence Development" · Working paper, 2026

Bostrom's framing is notable because his earlier work (Superintelligence, 2014) established many of the foundational arguments for why superintelligence poses existential risk. His 2026 shift toward timing and distribution questions reflects a broader movement in the field: accepting that development will happen and focusing on who controls the process. This is a legitimate and important question. It is not the same as ensuring the process is safe.

The AGI-to-ASI transition problem

Governance discussions often treat AGI as the relevant threshold, on the assumption that AGI-level systems would be comprehensible to human overseers and subject to meaningful evaluation. This is probably true at the AGI threshold. The problem is what comes next.

A system at AGI capability that can direct its own research and improvement could transition from human-level to significantly superhuman performance in a period far shorter than it took to reach human level. The technical term for this is a fast takeoff scenario. In a fast takeoff, a system moves from a level where human governance is possible to a level where it may not be, in a time window shorter than the response time of any governance institution.

Readiness for superintelligence means readiness before the AGI threshold, not a plan to respond after it. Governance that kicks in at AGI arrival may be too late to prevent a rapid transition to capability levels that defeat it.

Why the window is still open

The readiness gap is real and serious. Describing it honestly is not the same as concluding that nothing can be done. The window for building adequate governance is narrowing, but it has not closed.

The leading frontier AI labs are still at human-level or near-human-level performance on most cognitive tasks, not significantly beyond it. The transition from this capability level to the one that matters most has not yet happened. This is the time when governance frameworks can be built with the cooperation of the actors who will be subject to them, not after those actors have developed systems that make cooperation less necessary.

The political cost of building governance now is real: competitive friction, slower deployment, potential loss of economic advantage relative to less safety-conscious actors. These costs need to be weighed honestly against the cost of not building governance now and being in the situation described above. The asymmetry in that comparison is not subtle. The Foundation's three proposals are designed for exactly this window: before the threshold is crossed, while the cooperation of all major actors can still be sought and obtained.

Common questions.

Are we ready for superintelligence?

No. Readiness would require a verified technical solution to the alignment problem, binding international governance with enforcement mechanisms, and independent monitoring infrastructure. None of these exist at the required level. The alignment problem is unsolved for frontier systems. International frameworks are at the declaration stage, not the binding treaty stage. Safety evaluations are conducted by the organizations developing the systems without independent verification.

How long would it take to be ready?

Building genuine readiness from the current state would require at minimum five to ten years of serious effort. International treaty negotiations with binding enforcement mechanisms take years from initiation to signature and more years to implement. Technical alignment verification tools do not yet exist at the required level. The problem is that this timeline may already be shorter than the time remaining before superintelligence arrives, which makes beginning the process now — without waiting for timelines to become certain — the only defensible approach.

Why did the 2025 Paris Summit shift away from safety?

Because of genuine political pressure from governments and industries that see AI competition — particularly with China — as the primary geopolitical frame, and safety requirements as potential competitive disadvantages. The renaming from "safety summit" to "action summit" was deliberate. It reflected a real shift in how many governments are prioritizing AI: as an economic and strategic asset first, a safety challenge second. This is concerning precisely because the safety challenge does not become smaller because governments have decided to focus elsewhere.

What would closing the readiness gap actually require?

Three things working together: technical alignment verification methods reliable enough to be used as deployment prerequisites, an international treaty with binding authority over frontier AI deployment by all major actors, and monitoring infrastructure operated independently of the organizations building the systems. The Foundation's three policy proposals address each of these. None of them is easy. All of them are possible within the current governance window if sufficient political will exists.