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 an urgent practical assessment, and the answer is clearly no, not a theoretical exercise.
What readiness would actually require
Being specific about what "readiness" means matters, 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.
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 ASI.
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 no 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. A legitimate question. Different from 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.