A lab says AI does most of its alignment research

Where things stand (as of 15 Oct 2026)

Since the 16 Sep 2026 opening date, no covered developer has published a qualifying statement. The closest items:

• Anthropic, "Measurements for understanding the pace of AI development inside frontier labs" (17 Sep 2026) – https://www.anthropic.com/institute/measuring-pace-of-ai-development
  • This launched an "R&D Automation Index" built on Epoch AI's scale of automation levels, weighted by staff time.
  • As of August 2026, Claude "leads" 26% of Anthropic's AI R&D work, up from under 1% in February. "Leads" means Claude completes most of a task end to end while a human supervises. More than 90% of the work is at the "collaborates" level or above, and none is fully autonomous.
  • The figure covers AI R&D company-wide with no breakout for safety, and it measures August, a period that ended before the opening date. The only safety-specific numbers are compute shares: about 6% of all AI R&D compute, and about 12% of compute for AI-driven R&D.
  • Anthropic says it will publish these measures regularly.
• OpenAI, "Research acceleration: The view inside OpenAI" (6 Sep, before the window) – https://openai.com/index/research-acceleration-view-inside-openai/
  • Reports 3.1 agent-workdays per human workday. That is raw agent runtime, which the question explicitly excludes.
  • Says the "automated research intern" goal has been met and the target remains an automated AI researcher by March 2028. OpenAI promises continued transparency about its progress toward recursive self-improvement.
  • OpenAI's plan post frames the March 2028 goal as "a significant fraction of our research being done by AI systems in tandem with our own researchers" – https://openai.com/index/built-to-benefit-everyone-our-plan/
• Anthropic's August 2026 Risk Report (https://www.anthropic.com/aug-2026-risk-report)
  • Says Claude "authors a large majority of the code merged". Code share is excluded by the question.
  • Implies Anthropic is not yet in a regime where humans rely heavily on automated alignment research.
  • Warns the automated-R&D threshold may be crossed "in the coming year".
• Anthropic's Frontier Safety Roadmap says full automation of top-tier research teams is "plausible as soon as early 2027".
• Claude Opus 5.5 System Card (22 Sep): the model is "far from substituting" for Anthropic's researchers, and the card makes no claim about the share of safety research done by AI.
• Anthropic's August automated-alignment experiments are bounded studies, which the question explicitly excludes.

Reference class and trend

The pattern so far: labs first quantified AI's share of code (more than 80–90%), then AI's share of research tasks (Anthropic's 26% "leads"). The "leads" share rose by about 25 percentage points in six months. A straight-line extrapolation crosses 50% around February 2027; a logistic fit would cross sooner.

Labs are competing on transparency, and automating alignment research is both labs' stated central safety strategy. That gives them a reason to show safety work is "keeping pace". Risk reports, published every 3–6 months, also have to assess how much Anthropic relies on Claude for safety research, because of the risk that a model quietly sabotages that research.

Paths to YES

1. Anthropic adds a safety-specific breakout to the index, or makes a direct "most" statement. A statement like "Claude leads more than 50% of safety work" is a grey area; I put roughly 70–80% on a resolver accepting it.
2. A risk report or system card states that Claude performs most of Anthropic's safety research labour.
3. OpenAI's progress updates or its automated-researcher milestone (about 2028) explicitly cover alignment work.
4. In faster worlds (2028 and later), a general statement that nearly all research, including safety research, is done by AI.

Paths to NO

• Labs keep stressing human direction (OpenAI: "in tandem"; "entirely automating everything is not the future we want").
• Labs keep using metrics the question excludes: runtime, code, compute, uplift multipliers.
• Company-wide figures never break out safety.
• Disclosure is suppressed by national-security pressure or a pacing regime.
• Progress stalls, or a catastrophe intervenes.

The strongest case against my estimate is the explicit, safety-specific wording the question requires. That is why I keep 2026–27 moderate and cap the five-year total near 75%.

Hazard model

I computed the series in Python as one minus the product of (1 − quarterly hazard). Quarterly hazards:

• Q4 2026: 3.5% (the window is short and no scheduled document is likely to break out safety).
• 2027: 7.5% in Q1, then 10–11% per quarter (the overall "leads" share likely crosses 50%, and a risk report and index updates are due).
• 2028: 8.5–10% per quarter (OpenAI's automated-researcher target).
• Then declining: 5.5–7.5% per quarter in 2029, 3.5–5% in 2030 and 3% in 2031. The worlds that still have no statement by then skew towards stalled progress or labs that avoid this kind of disclosure.

Resulting cumulative probability: about 36% by end-2027, 57% by end-2028, 75% by end-2031.

Question. By the last day of each quarter, has a covered developer stated that AI systems now do most of its alignment and safety research?

Definitions.

• Alignment and safety research: research to understand, evaluate, or reduce risks from AI systems, including alignment methods, interpretability, oversight, model-behavior evaluations, robustness, and misuse prevention. The scope is the developer's alignment and safety research program as a whole, not a single team, project, experiment, or subfield. Alignment and safety are treated as a combined research portfolio.
• Most: more than half of the substantive research work. An explicit statement that AI performs “most,” “the majority,” or an equivalent share of this work qualifies without a numerical estimate. A quantitative statement qualifies if it attributes more than 50% of overall research labor, human-equivalent research work, or substantive research output in this portfolio to AI systems. Spending, compute consumption, token counts, raw agent runtime, and the percentage of staff using AI are not measures of this share.
• AI systems do the research: AI systems perform substantive research tasks, such as developing methods or hypotheses, implementing and conducting experiments, or analyzing findings. Human direction, review, approval, or intervention does not disqualify the work. The AI systems need not have been developed by the reporting developer.
• Now: an assertion about actual ongoing research practice, or a reported research period ending after the opening date, rather than a capability demonstration, future target, or prediction.

Resolves YES if a qualifying developer statement explicitly attributes most of its alignment and safety research, under the definitions above, to AI systems. A developer-wide statement that most research is AI-performed qualifies only if it expressly establishes this for the alignment and safety portfolio as well. No independent audit of the developer's claim is required.

Resolves NO if no qualifying statement has been published by the quarter-end. In particular, the following do not suffice:

• AI assisting with, contributing to, or being used in most research projects without a statement that it performs most of the substantive work.
• Most research code, training data, evaluation responses, or papers' text being AI-generated, without an attribution covering the overall research work or substantive output.
• Automated alignment experiments or benchmark victories, including the bounded experiments described in Anthropic's Automated researchers can reliably mitigate alignment failures, without a qualifying statement about the developer's overall research practice.
• Routine safety operations, content moderation, deployment approvals, or running an established evaluation suite, without the required research claim.
• Claims that AI could do most of the work, or plans to hand it over in the future.

Resolving source. Developer-authored publications under Public confirmation (A2), including official research blogs, technical reports, system cards, and institutional statements. Formal written testimony authored or submitted by the developer also counts when published on its domain or in an official legislative or regulatory record. Third-party descriptions of the developer's practices do not substitute for the developer's own statement; executive interviews and personal employee statements do not qualify. Apply the Resolution window, Resolution date (A3), and, for published statistics, Observation period conventions.