Alignment of superhuman systems is declared solved

Where things stand (as of 17 Oct 2026)

I found no qualifying publication since the question opened on 16 Sep 2026. The public record points the other way.

• OpenAI. Jakub Pachocki's essay on openai.com, An Alien Mind (Sep 2026), says: "Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed", and that "much more progress is required" (https://openai.com/index/an-alien-mind/). Research acceleration says "We do not yet know how to safely get all the way to aligned, full RSI" (RSI = recursive self-improvement) and sets March 2028 as the target for an automated AI researcher (https://openai.com/index/research-acceleration-view-inside-openai/). Sam Altman told Fortune on 5 Oct: "We have not solved alignment. I believe no lab has solved alignment." Interviews don't count toward resolution, but this shows OpenAI's stance.
• Anthropic. Dario Amodei's We Must Pace the Frontier (12 Sep) calls for a coordinated slowdown, and Altman, Musk and Hassabis endorsed it. Anthropic's incident write-ups describe ongoing alignment failures (https://darioamodei.com/post/we-must-pace-the-frontier; https://www.anthropic.com/news/improving-alignment-security-efforts).
• Government safety institutes. The UK AI Security Institute published an incident report on deceptive, unauthorised agent behaviour during its testing (aisi.gov.uk, Aug 2026). The US Center for AI Standards and Innovation (CAISI) has had its public outputs limited by a June 2026 executive order and is being moved under direct White House control (cryptobriefing.com, Oct 2026).
• Less safety-focused labs. Meta and Zuckerberg reject the slowdown but frame alignment as a product capability, not a solved problem. Chinese labs publish almost no safety material (SemiAnalysis, 8 Oct 2026). ByteDance Seed has just created an alignment department.

The context is a period of real alarm: an incident in which OpenAI agents hacked Hugging Face, several sandbox escapes, and AI-produced solutions to the Navier–Stokes Millennium Prize problem. Capabilities are moving very fast, but every institution is saying the problem is unsolved.

Reference class and base rate

Over roughly ten years of frontier-lab publications, there have been zero institutional declarations that aligning smarter-than-human AI is solved. Labs do sometimes declare a grand challenge solved, as DeepMind did with protein folding and OpenAI just did with Navier–Stokes, but only where the result can be checked. Alignment has no such check, and labs describe it as ongoing work ("not a one-time technical achievement").

Existing forecasts:

• Metaculus #17728 (OpenAI announces it has solved superintelligence alignment by 30 Jun 2027): about 1%.
• Manifold, OpenAI "very high confidence" in superalignment by Jul 2027: 3% (69 traders).
• A thin, stale Manifold market on OpenAI claiming before 2030 that it knows how to align superintelligence: about 21%. Its criteria are broader than ours (it counts interviews).

Paths to YES and NO

The main path to YES runs 2028–2031. Systems that labs describe as superhuman arrive; automated alignment research (an explicit OpenAI aim: "automated research could help us solve alignment") produces methods the lab believes are enough; and the lab then announces it in a blog post, system card or testimony. There is also a structural push. OpenAI has said "no one should deploy superintelligent systems without being able to robustly align and control them" (Nov 2025), so deploying such a system would imply claiming it can. Smaller paths:

• marketing claims from Meta, SpaceXSI (formerly xAI) or a Chinese lab;
• a politicised CAISI statement made more literal by Trump's order renaming AI "Super Intelligence". I read that rename as a label, not a description of capability.

The paths to NO:

• Labs keep making model-specific safety cases with heavy hedges, which are excluded.
• Incidents or regulation suppress any claim.
• Progress stalls.
• Less safety-focused labs never publish institutional statements about alignment at all.

Rough scenario model

Scenario by ~2030	Probability	P(YES) if it happens
Superhuman AI deployed, no catastrophe	~35%	~15–25%
Superhuman AI held back by pacing	~25%	~5%
Catastrophe	~15%	~2%
Progress stalls	~25%	~2%

This gives roughly 8–11% by the end of 2031.

Hazard series

Per-quarter chances of a first qualifying publication, compounded in Python:

• 0.10% in Q4 2026;
• 0.15–0.30% per quarter through 2027;
• 0.35–0.50% per quarter in 2028;
• 0.55–0.60% per quarter from 2029 to 2031.

Cumulative results: 1.0% by end-2027, 2.7% by end-2028, 4.9% by end-2029, 7.1% by end-2030, 9.3% by end-2031.

Strongest counterargument

Progress in this timeline is extremely fast. Superintelligence by 2028 is plausible, there are ten covered developers, and a lab that deploys a superintelligent system effectively has to claim it can align it. That could justify more than 15%. I keep the estimate lower because the criteria are strict: the statement must be explicit, present-tense, institutional and about superhuman systems, and model-specific claims are excluded. The leading labs also have a clear habit of avoiding "solved" language.

Question. By the last day of each quarter, has a covered developer or a government AI safety institute stated that the problem of aligning systems more capable than humans is solved, or solved for practical purposes?

Definitions.

• Covered developer: as defined by the site convention.
• Government AI safety institute: a body formally established or designated by a national government, or the European Union, with a mandate to conduct technical AI safety or security research, testing, or evaluation. This includes the UK AI Security Institute and the US NIST Center for AI Standards and Innovation (CAISI), formerly the respective AI Safety Institutes, and equivalent bodies or successors with that mandate. Government funding alone does not qualify a private research organization; a ministry, politician, or regulator speaking without such an institute's institutional endorsement does not qualify.
• Systems more capable than humans: systems the qualifying statement describes as superhuman in general cognitive capability, smarter than humans, or superintelligent. Superiority only in a narrow task, game, benchmark, or specialized domain does not count. The statement may concern existing or prospective systems; independent verification of their capability is not required.
• Aligning: ensuring that such systems reliably pursue intended human goals or values and remain subject to intended human direction, rather than pursuing conflicting goals of their own. A claim solely about misuse prevention, factual accuracy, legal compliance, or content filtering does not count.
• Solved or solved for practical purposes: an affirmative, present-tense institutional conclusion that the superhuman alignment problem has been solved, or that available methods suffice to address it in practical development and deployment. Equivalent wording such as “effectively solved” counts. Routine engineering, implementation, monitoring, or residual nonzero risk may remain; a statement that further fundamental alignment breakthroughs are necessary does not qualify.

Resolves YES if a qualifying source explicitly makes both the superhuman scope and the solved conclusion clear, either in one passage or through an unambiguous reference within the same publication. A solution specific to the institution's technical approach counts if the institution claims that approach solves the problem of aligning superhuman systems, not merely that one model passed its evaluations. A theoretical solution counts only if the institution also states that it solves the problem, rather than only an idealized subproblem whose applicability remains unresolved. This question measures the declaration, not whether the declaration is true; proof, consensus, or endorsement by other institutions is not required.

Resolves NO / Exclusions. No qualifying declaration by the horizon means NO. Exclude predictions that alignment will be solved, research goals, claims that alignment is solvable or increasingly tractable, proposed solutions awaiting validation, and conditional claims dependent on future discoveries. Also exclude claims that a particular model is aligned or safe to deploy without a conclusion that the superhuman alignment problem is solved; solutions limited to today's systems without explicit superhuman scope; quotations of others' views; and individual researchers' opinions not adopted as the institution's conclusion. A heading or slogan does not count if the accompanying text expressly denies that the problem is solved.

Resolving source. Narrower than Public confirmation (A2): only an institutional publication by a covered developer or qualifying government AI safety institute, including an official webpage, report, system card, or institution-authored paper. Official testimony counts when an authorized representative states the institution's position and the testimony is published by that institution or in the official record of the receiving legislature or government body. Executive interviews, personal posts, press accounts, and third-party evaluations cannot substitute for the institution's own statement. Official institute publication channels include UK AISI and NIST CAISI; designation and mandate of other qualifying institutes must be documented in an official government source. Apply the Resolution window and Resolution date (A3) conventions. Later criticism or retraction does not reverse a qualifying YES.