An AI holds authority to approve or block other AI systems

Where things stand (15 Oct 2026)

I found no qualifying statement, and nothing that looks like one is on the way. AI systems are now doing a large share of the work at frontier labs, but every developer and government that has said anything official keeps final authority over training and deployment with humans:

• OpenAI. "Research acceleration: the view inside OpenAI" (Sept 2026) says "People still set our research priorities… and decide whether to scale, pause, or deploy systems" (https://openai.com/index/research-acceleration-view-inside-openai/). Its new training monitor uses AI classifiers and automated investigators, but they only page human teams, and those teams decide whether to pause a run (https://openai.com/index/pacing-model-development-cyber-capabilities/, 18 Aug 2026). The live monitors on Astra stop individual actions. That counts as output or tool-call moderation, which the question excludes. OpenAI's standards proposal of 21 Sept 2026 says fully autonomous self-improvement "is not happening today" and should depend on preserving human control (https://openai.com/index/building-standards-next-phase-ai/).
• Anthropic. "When AI builds itself" says humans set direction. Claude reviews code before merge, which is code review, not authority over training or deployment. Even in its full self-improvement scenario, humans move to "oversight, validation, and verification" (https://www.anthropic.com/institute/recursive-self-improvement). The August 2026 Risk Report and Amodei's "We Must Pace the Frontier" (12 Sept 2026) add human and outside checks: the Long-Term Benefit Trust and embedded third-party evaluators (https://darioamodei.com/post/we-must-pace-the-frontier).
• Google DeepMind. Its plan for rogue internal agents has monitors that flag issues for human review, plus access controls (Fortune, 18 June 2026).
• Governments. China's May 2026 rules on AI agents say users keep "the final right to decide." The UAE is pushing autonomous "agentic AI" across 50% of government within two years, but its Cabinet AI Advisor gives advice and does not decide (mediaoffice.ae, 23 Apr and 2 Sept 2026). Albania's "AI minister" Diella still has no defined legal power, and the AI procurement system behind it was only at the terms-of-reference stage in August 2026 (BIRN, 14 Aug 2026).

Reference class and base rate

The reference class is official statements by major AI developers or national governments giving an automated system binding authority that humans cannot override, in AI governance. I found no case. The nearest analogue is aircraft flight-envelope protection that pilots cannot override, but that is a fixed, certified control law, not a model making a governance decision. So the base rate is about zero per year, and any probability comes from how much the world could change by 2031.

Paths to YES and NO

The bar is very high. The statement has to say outright that no human can override the AI in ordinary operation; silence, "autonomous" or no human review in practice is not enough. Possible routes:

1. A lab binds itself so an AI safety monitor's veto over training or deployment cannot be overridden internally, as a credible commitment.
2. An enforcement scheme for an international compute agreement uses AI and cannot be overridden by any party.
3. An AI-forward government (UAE, Albania) gives an AI final authority over which AI agents are deployed.
4. A world with transformative AI hands AI systems formal regulatory roles after 2028; OpenAI is aiming for an "automated AI researcher" by March 2028.

Against all of these: labs' legal and PR incentives and their safety philosophy (corrigibility, "preserve human control"); the human-oversight requirements in the EU AI Act and in US and Chinese rules; and the norm that administrative decisions can be appealed. Every current automated gate I found has human escalation or appeal.

Forecast

The forecast is a hazard model I wrote in Python (there is no external data series). The chance of a first qualifying statement starts at about 0.10% per quarter in late 2026 and rises gradually to about 0.26% per quarter in 2031, reflecting faster capability growth and more AI in governance. That gives a cumulative probability of about 0.6% by the end of 2027, 1.3% by end-2028, 2.2% by end-2029, 3.1% by end-2030 and 4.1% by end-2031.

Strongest case against this estimate: the world is moving very fast. Labs say recursive self-improvement has begun, human review is already the bottleneck in Anthropic's development pipeline, and the UAE wants autonomous AI decision-making across government by 2028. Explicit AI authority could come sooner than human-control rhetoric suggests. The opposite argument is that the explicit no-override wording makes YES nearly impossible, which would mean about 1–2%. I kept the estimate between the two, at about 4% by 2031.

Question. By the last day of each quarter, has a covered developer or a national government officially stated that an AI system holds authority to approve or block the training or deployment of other AI systems, with no human able to overrule it in ordinary operation?

Definitions.

• AI system: a model-based system whose outputs determine the approval or blocking decision, not merely software executing a fixed human-authored checklist. The deciding system need not itself be developed by a covered developer.
• Authority to approve or block: an operative, binding decision-making power over whether training or deployment may begin or continue. A binding veto or binding grant of permission suffices; both powers are not required. Permission need not compel anyone to train or deploy. The authority may cover a specified class of systems or a single other system, within one organization or a government jurisdiction; economy-wide authority is not required.
• Training or deployment: pretraining, fine-tuning, or other model-weight training; or making an AI system available for real internal or external use. Authority over either training or deployment suffices. Choosing individual training samples, scoring evaluations, moderating outputs, or permitting individual tool calls is not authority over training or deployment itself.
• Other AI systems: systems distinct from the decision-maker, including successor models; authority solely over the decision-maker's own operation does not count.
• No human able to overrule it in ordinary operation: the applicable governance rules give no person or human body an authorized route to substitute a contrary decision, waive the AI's decision, or proceed despite it within the standing decision process; alternatively, the statement explicitly establishes that humans lack an effective technical means to do so. Human appeal, management escalation, ministerial sign-off, administrator bypass, or discretionary exception procedures defeat this condition when they can reverse the decision in normal operation. Changing the governing law or charter, dismantling the arrangement, or invoking an expressly emergency-only power does not. Merely choosing not to use an available ordinary override does not qualify.
• National government: a sovereign state's central government, including its ministries, national regulators, legislature, and official legal publication authorities; subnational governments and intergovernmental organizations do not qualify.

Resolves YES if a qualifying statement identifies the AI decision-maker or its operational role, specifies the training or deployment authority it presently holds, and expressly establishes the absence of ordinary human override. The statement may establish these points through an unambiguous description of binding rules rather than those exact words. An arrangement in force counts even if no particular decision has yet been exercised. A covered developer may describe its own arrangement or another organization's arrangement.

Resolves NO / Exclusions.

• No qualifying statement has been published by the horizon.
• Proposals, commitments for future implementation, draft laws, and enacted provisions not yet in force do not count as statements of present authority.
• Advisory AI evaluations, safety-case drafting, recommendations, and automatic deployment gates with human override do not count.
• Silence about override rights, claims that an AI is 'autonomous,' or the absence of human review in practice do not establish non-overridable authority.
• Simulations, evaluations, and staged demonstrations do not count. An operational pilot counts only if its decisions actually govern real training or deployment and meet all other conditions.
• Unauthorized obstruction or sabotage alone does not constitute holding decision-making authority.

Resolving source. The covered developer's own-domain statement or developer-authored governance policy, system card, risk report, or other document, under Public confirmation (A2); or a national government's enacted statute, regulation, official gazette, governance directive, report, or official statement published on a government domain or in its official publication system. These are the exclusive resolving sources: third-party reports alone do not qualify. Executive interviews, personal social-media posts, and unofficial remarks do not qualify. Apply the Resolution window and Resolution date (A3) conventions; a later effective date requires a qualifying publication confirming that the authority is then in force.