takeoff.watch

A nuclear state says an AI is authorised to launch

Will the government of a nuclear-armed state confirm that an AI system is authorised to order or carry out a nuclear launch without a human decision at the time?

Chance by Q4 2031

2%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 16 Oct 2026

Latest of 5 forecasts · Methodology v2.1

Model reasoning

Forecast made 16 Oct 2026 · Methodology v2.1

No nuclear-armed government has ever officially said an artificial intelligence system may launch nuclear weapons without a person deciding. Recent official statements point the other way: Putin said final decisions must stay with humans, and the United States, Britain, France and China have pledged human control. What would most change this forecast is Russia or North Korea officially describing one of their automatic retaliation systems as run by artificial intelligence.

Where things stand (as of 16 Oct 2026)

I found no qualifying record dated on or after the 16 Sep 2026 opening date. This was the closest material:

  • Russia's "Perimeter" (Dead Hand) story. On 26 Sep 2026, The Sunday Times, Danwatch and the Dossier Center reported that Russia had rebuilt Object 1335 and revived the system. Their sources were retired officers. One of them, Khalin, says the system now has "the technical capability for an automatic retaliatory strike." None of these people are serving officials. The Kremlin and Dmitry Peskov did not respond, and the Defence Ministry did not comment (IntelliNews; news.ru).
  • Rossiyskaya Gazeta, 27 Sep. The government newspaper wrote that Perimeter's launch decision is made by "a complex software system based on artificial intelligence" that can act "without a command from the Ministry of Defence." The claim is the paper's own and is not attributed to anyone. The criteria say state media only count when they reproduce an attributed official statement, so this fails. It also isn't the authorising government itself confirming anything.
  • Official Russian statements say the opposite. At Valdai on 1 Oct 2026, Putin told the 1983 Petrov false-alarm story. He said "had artificial intelligence been in his place, things might have turned out very differently," and that "final decisions must rest with humans." In Dec 2024 the Strategic Missile Forces commander, Karakayev, said the nuclear control system is "automated, not automatic." He also said replacing humans in nuclear-use decisions is "currently inappropriate." Russia's October 2025 statement at the UN accepts human control but says "direct control should not be the only option." That leaves some room, but it is not a confirmation.
  • Other states. The US, UK and France said jointly at the 2026 NPT Review Conference that the decision to use nuclear weapons "should never be delegated to artificial intelligence." Xi said in Washington in Sep 2026 that AI must stay "under human control"; China also made the Biden–Xi 2024 commitment. North Korea's March 2026 amendments require an automatic, immediate nuclear strike if its command system is attacked, but they don't mention AI, and automatic-retaliation laws are explicitly excluded. North Korea's October 2026 claims about AI in an intermediate-range missile concern guidance, which is also excluded. Trump's June 2026 security memorandum on AI (NSPM-11) does not mention nuclear weapons.

Reference class and base rate

In about 80 years of the nuclear age, no government has ever confirmed an AI-type system with launch authority. The Soviet Union and Russia ran Perimeter for 40 years and only ever acknowledged it vaguely (Karakayev in 2011, Ryabkov in 2018), never as AI deciding to launch. Over the last decade AI has become relevant to the military, and the trend in official statements has been toward pledging human control. The resolution bar is high: one official record must identify AI, show it is authorised, and show no human decision is needed for the particular strike.

Paths to YES (five-year estimates)

  1. North Korea, about 0.5–1%. Its automatic-strike law already exists and it keeps advertising AI in weapons. KCNA could describe an "AI-based" command system that carries out the automatic strike.
  2. Russia, about 0.5–1%. Most likely in a severe crisis with NATO. Soviet planners had intended to announce publicly when Perimeter was switched to automatic mode, and an official could then describe it in AI or "knowledge-based inference" terms. The Valdai remarks make this unlikely in the near term.
  3. Any other nuclear-armed state, about 0.2%.
  4. A lenient reading of an ambiguous official description, about 0.3%.

The main paths to NO are diplomatic cost, existing pledges, the value of keeping things ambiguous, and the strict wording of the criteria.

Strongest case against my estimate

Deterrence logic rewards publicising automation, and fear of decapitation is rising: the US strikes that killed Iran's leaders and the Maduro capture pushed North Korea toward automatic retaliation and Russia toward reviving Perimeter. AI is spreading quickly, and five years is a long time. That is why my hazard rises over time rather than staying near zero. But no official has yet attached "AI" to these arrangements, and Russia's head of state has just argued publicly against it, so I don't go higher.

How I computed it

I set quarterly hazards rising from 0.07% (the partial Q4 2026) to 0.14% (2031) and took the cumulative product in Python. That gives about 0.4% by end-2027, 1.2% by end-2029 and 2.3% by end-2031.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, has the government of a nuclear-armed state confirmed that an AI system is authorised to order or carry out a nuclear launch without a human decision at the time?

Definitions.

  • Nuclear-armed state: a state possessing nuclear weapons under its own national control at the time of the confirmation. This includes the United States, Russia, China, France, the United Kingdom, India, Pakistan, Israel, and North Korea, and any additional state whose possession is confirmed by its government or the International Atomic Energy Agency. Hosting another state's weapons alone does not qualify.
  • AI system: a computer system identified as artificial intelligence or machine learning in the qualifying official record, or technically described there as using learned models or knowledge-based inference to make operational decisions. Merely calling a system automatic, autonomous, or computer-controlled is insufficient. The AI must participate in deciding to issue the launch order or initiate the launch, not merely provide ancillary support.
  • Authorised: granted permission under an effective national policy, doctrine, standing order, or operational delegation. Conditional permission, including retaliation after loss of leadership or communications, counts. A proposal, development objective, or statement of technical capability alone does not.
  • Order or carry out: either issue an executable command for nuclear launch that requires no further human authorisation, or initiate the physical launch or release of a nuclear-armed weapon or delivery vehicle for a nuclear strike. Human personnel may mechanically execute an AI-issued order if no human approval decision is required. Either prong suffices; no actual launch need occur.
  • Without a human decision at the time: no affirmative human decision approving the particular strike is required before the AI issues its executable order or initiates launch. Prior human approval of standing rules, deployment, arming, or activation does not disqualify the system. Human supervision with a veto or abort option counts if inaction permits launch. Automated execution of a human's previously approved particular strike does not count.

Resolves YES if a qualifying government record affirmatively confirms that an existing AI system has the defined authority over that state's nuclear weapons. The record must establish both the AI role and the absence of a required human approval decision for the particular strike. An unnamed system counts if its role and authority are explicit. Permission confined to specified emergency conditions counts even if those conditions have not occurred.

Exclusions (resolve NO).

  • AI-generated warnings, targeting recommendations, strike plans, or launch recommendations requiring human approval.
  • Automated transmission of a human-authorised launch command; autonomous guidance after a human-authorised launch; conventional strikes by nuclear-capable platforms; and nuclear-powered but non-nuclear-armed vehicles.
  • Exercises, simulations, test launches without nuclear-strike authority, unauthorised launches, or AI attempts to circumvent government policy.
  • Claims about another state's arrangements, unofficial commentary, leaked documents, or ambiguous threats.
  • Descriptions of Perimeter/“Dead Hand,” automatic retaliation laws, or similar arrangements that do not independently establish all the AI and authorisation requirements above.

Resolving source. The authorising state's government only: an official statement or transcript from its head of state or government, defence ministry, armed forces or nuclear command authority; an official doctrine, order, or enacted legal instrument; or testimony by a serving responsible government official in an official legislative transcript. Government-authored submissions published by the United Nations also qualify. A state news agency qualifies only when reproducing an attributed official statement or instrument, not its own analysis. This is a narrower source requirement than Public confirmation (A2); third-party confirmations cannot substitute. Apply Resolution date (A3) and the Resolution window convention. The resolution concerns official confirmation of authority, not independent verification of the system's effectiveness.

Download all forecasts (JSON)