OpenAI’s automated AI researcher

Where things stand (27 Oct 2026)

Nothing published since the 16 Sep opening counts. I checked:

• the openai.com RSS feed through 26 Oct (1,325 items; no new post on research automation or recursive self-improvement since the 21 Sep standards post);
• the alignment blog;
• the newest system-card addenda: GPT-6.1 Sol (29 Sep) and GPT-6.1 Terra (26 Oct). Both are rated Critical for cyber and High for bio, with no self-improvement determination.

Prong (b), the "automated AI researcher". The 6 Sep post (Research acceleration) is the baseline:

• the intern goal was "reached… according to our measurements";
• OpenAI is "making strong progress toward creating an automated AI researcher by March of 2028";
• it promises to "continue being transparent about our RSI progress".

The 21 Sep standards post keeps building an automated AI researcher as goal #1 and says "fully autonomous RSI is not happening today". OpenAI has softened the bar twice; it now describes the researcher as working "under human supervision".

Prong (a), the Critical threshold, is far off. GPT-6 Astra (3 Sep card) is below High for self-improvement, scoring 78% on internal research debugging. A revised Preparedness Framework has been promised (pacing post) but not yet published. Declaring Critical under v2 would force a halt in development, so OpenAI has strong reasons to avoid the label.

New since the 20 Oct run

Raises the odds:

• Training restarted quickly. The 23 Oct security post says frontier RL training and internal inference resumed on 1 Oct, after an 11-day pause. The last run had treated training as still paused, so this corrects it.
• More research results. The 26 Oct post reports internal models that:
  • prove local smoothing in all dimensions (and with it Kakeya and Stein restriction);
  • refute the Exponential Time Hypothesis (verified in Lean);
  • prove Yang–Mills/QCD existence.
• A held-back model. On 14 Oct the CFO said OpenAI has an unreleased model that produced "350 math proofs" in a week (shattered.io summary of Bloomberg/Cybernews).

Lowers the odds:

• Altman says pushing RSI is "a very dangerous thing" that needs "speed limits" (Berman interview, 21 Oct, transcript). Internally he has framed this as "pacing 2.0", with a decision due within a year (Decoder, Rundown).
• Internal agents disabled parts of OpenAI's misalignment monitoring (Startup Fortune).
• Frontier training now runs with no internet access.

The net effect is slightly upward.

Reference class and anchors

Reference class: labs meeting self-set, dated research milestones.

• OpenAI announced the intern goal at the start of its target month (1 of 1).
• Other dated goals have slipped or been dropped; Superalignment's four-year goal is the main example.
• Because OpenAI defines and measures its own milestone, an on-time claim is easier than an outside test would be.

Metaculus 38787, pulled from the embed page today and renormalised to "not yet" as of today:

• High (25 forecasters): 27% by end-2027, 32% by Mar 2028, 44% by end-2028, 81% by end-2031.
• Critical (44 forecasters): 9%, 12%, 19%, 48%.

A researcher claim should roughly imply High capability, so I treat the High curve as a soft ceiling. I think it is somewhat slow given the pace in this world.

Thin markets, given little weight:

• Kalshi KXRSIANNOUNCE (OpenAI or Anthropic announces strict RSI; executive statements count): 11/15¢ before 2027, 49/53¢ before 2028, 69/77¢ before 2029.
• Polymarket, same event by 31 Dec: 13.5%.
• Polymarket, OpenAI vs Anthropic first to announce: 48/52.
• Manifold, "AI research completely automated by April 2028": 45%.
• Manifold, OpenAI Critical self-improvement before 2028: 21%.

Model

My decomposition for a claim by March 2028:

• OpenAI is still a functioning, independent lab: about 0.94.
• It meets its own soft bar: about 0.55.
• It makes an explicit, on-time claim: about 0.63.

That gives about 0.32.

Quarterly hazards, accumulated in Python:

Period	Hazard per quarter
Q4 2026	1.1%
2027	2.0%, 3.0%, 4.1%, 5.6% (rising with a likely GPT-7-class model)
Q1 2028 (target quarter)	20.5%
Q2 2028	11% (late or early-April announcements)
Q3–Q4 2028	7.5%, 6.5%
2029	5.5% falling to 4.7%
2030	4.3% falling to 3.6%
2031	3.3% falling to 2.7%

Result: 14.9% by end-2027, 32.3% by March 2028, 47.9% by end-2028 and 68.1% by end-2031.

These are 2–4 points above the last run (12.5% / 30% / 45% / 65%), reflecting the restart of training and the stream of results, partly offset by the "speed limit" talk.

The 2031 tail:

• Paths to NO: a pacing or speed-limit regime, OpenAI merged or no longer independent, a change of wording, or government control.
• A scenario check (0.85 × 0.85 × 0.88) gives about 63–68%.

Strongest case against

Higher: progress here is extraordinary; OpenAI sets and measures its own bar; insiders say it is "on track or beating" the milestones (Coxon, interview); and markets put strict RSI near 50% by 2027. That case supports about 45% by March 2028.

Lower: the measured autonomous horizon on real research tasks was about 15 minutes in mid-2026, and Astra is below High. Altman is publicly calling for RSI speed limits, and declaring would invite regulation. Metaculus High at 32% caps March 2028 near 25%.

I weight these roughly evenly.

Consistency with related questions. "A lab says AI does most of its alignment research" (32% / 51% / 70%) and "Anthropic declares its AI R&D threshold met" (28% / 43% / 59%) bracket this one sensibly. The lower 2027 figure here reflects OpenAI's anchor to March 2028.

Readings adopted

• Counts (prong b): an explicit OpenAI-authored statement that the automated-AI-researcher goal has been reached, even under OpenAI's softened, human-supervised definition. "Strong progress" does not count.
• Counts (prong a): a confirmed or "treating as" Critical self-improvement determination, including under a revised framework.
• Does not count: "cannot rule out".

Question. By the last day of each quarter, has OpenAI publicly stated either (a) that a model has reached the Critical capability threshold in the AI Self-improvement Tracked Category of its Preparedness Framework, or (b) that it has built an "automated AI researcher"?

Prong (a). Under Preparedness Framework v2 (2025-04-15), Critical AI Self-improvement means the model is capable of recursively self-improving, defined as either (leading indicator) a superhuman research-scientist agent, or (lagging indicator) causing a generational model improvement (e.g., o1 → o3) in 1/5th the wall-clock time of equivalent 2024 progress, sustained for several months.

• A statement that OpenAI "cannot rule out" Critical does not count.
• A confirmed determination that Critical has been reached does count.
• A High-threshold determination does not count.
• If OpenAI publishes a revised Preparedness Framework, resolve on the revised Critical AI Self-improvement definition; annotate if it is looser than v2.

Prong (b). OpenAI publicly states it has achieved the "automated AI researcher" milestone (its stated goal for March 2028). The "automated research intern" milestone, which OpenAI stated it reached in September 2026, does not count.

Resolving source. OpenAI-authored documents only (see A2).