SuperintelligenceMONITOR
Published by Center for Artificial WisdomLatest newsletter: Issue #1 · draft v1.11

Daily evidence review · 2026-11-17

What the latest evidence changes

We examine new results, the work behind them, and what they could change for the future we hope for with superintelligence. Each judgment needs evidence about the conditions that connect capability to consequences.

Provisional rationale · October 20, 2026

Why we monitor the trajectory

An enduring better future with less suffering requires wisdom; wisdom requires intelligence without reification. Future SI may reinforce or loosen reification — the current trajectory is deeply uncertain. CAW investigates and influences the development and deployment of AI without reification, or artificial wisdom; SIM monitors the trajectory and evidence.

Reification means treating a provisional model, category or measure as a fixed truth, detached from the context and dependencies that give it meaning.

This is our provisional guiding rationale. The necessity claims guide our inquiry; they are not findings established by the cases below. Neither a capability gain nor a favorable test result establishes wisdom or an enduring reduction of suffering.

The reasoning and its open questions

Daily notes can be published by automated source review before human review. No per-note prepublication human sign-off is recorded here. Publication and review status.

New evidence concerns how institutions use and oversee AI. Apollo will pilot assessments inside OpenAI, while making clear that access and disclosure remain limited. Anthropic announced shorter vulnerability disclosure windows, effective January 1 rather than immediately. OpenAI’s study of scientists reports perceived time savings, but does not establish faster discovery experimentally. These developments make oversight, maintainers’ capacity and verification costs central to interpreting capability progress; none establishes an overall trajectory toward wise superintelligence.

Sources checked and limits of this review

Review since the November 16 briefing. Opened METR, ARC Prize, Epoch AI, Google DeepMind, OpenAI and Anthropic primary watchlists and searched current reporting and independent-research leads. TechCrunch’s November 16 coverage led to OpenAI’s announcement and Apollo Research’s own November 16 account; the latter supplies material access and disclosure qualifications. OpenAI’s current article and news index show November 16 despite a search snippet displaying November 13. Checked Anthropic’s original November 16 disclosure-policy update, including its January 1 effective date and exclusions. Read OpenAI’s November 16 science study and the full report’s methods and limitations: messages cover August 20–September 22, and the survey September 30–October 7. Publication dates are distinct from study and implementation dates. FrontierML’s original post is dated October 13; RSI Control’s arXiv record dates to October 20, revised October 21. Their renewed coverage was not treated as new research. No additional material new finding was verified on the other named watchlists in this bounded review. Coverage depends on accessible sources and indexing. Planned outside assessment is distinguished from published independent results, lab claims from measurements, and participant estimates from causal evidence. SI relevance is editorial interpretation. No human editorial approval, overall wisdom verdict, stop/proceed recommendation or pivotal designation is claimed.

Evidence note 1 · Assessor and lab announcements of planned independent evaluations; no results yet; SI relevance is editorial interpretation

OpenAI’s outside-assessor pilot comes with access and disclosure limits

Apollo Research’s November 16 announcement says six evaluators will work within OpenAI for four months, examining possible subversion during training, testing and internal use. OpenAI separately names Apollo, METR, Transluce and Goodfire as assessors in its first phase. These are agreements for assessments, not published safety findings.

Our interpretation. Editorial interpretation: access during development could reveal failures that a short test before release misses. Public confidence depends on which evidence assessors can inspect and disclose.

Main limitation. Apollo says the pilot does not yet meet its full access and disclosure criteria. OpenAI selects interview participants and can withhold information or redact supporting evidence for stated reasons, though not merely because findings are unfavorable; Apollo’s claim verdicts cannot be redacted. Some access details remain unsettled. Apollo is paid for the work, independently of its results.

Next evidence. The promised reports, including withheld evidence and unassessed questions, disclosed constraints, and whether findings lead to verifiable changes.

Evidence note 2 · Lab policy announcement and capability rationale; SI relevance is editorial interpretation

Anthropic shortens disclosure windows, starting January 1

Anthropic announced that vulnerabilities reported from January 1, 2027 will generally have 30-day disclosure windows for high- and critical-severity issues and 45 days for others, replacing 90 days. Existing reports keep their deadlines. The lab attributes the change to faster AI-enabled discovery and rediscovery.

Our interpretation. Editorial interpretation: faster discovery changes the time defenders have to repair systems. Capability gains can increase pressure on maintainers as well as help them; the outcome depends on deployment speed and resources.

Main limitation. This is a lab policy and rationale, not an independent measurement of net security benefit. Exceptions permit negotiated timelines for difficult fixes. The seven-day window for active exploitation is unchanged. AI-discovered findings without human review are excluded from this update.

Next evidence. Time to usable fixes and deployment, rediscovery and exploitation rates, and evidence about burdens on maintainers under the new windows.

Evidence note 3 · Lab-authored observational study; participant estimates and forecasts; SI relevance is editorial interpretation

Scientists report time savings; faster discovery remains unmeasured

OpenAI’s November 16 study analyzes over 580,000 scientific messages from 1,819 pilot participants and surveys 251 scientists. Respondents estimated that equivalent work at equivalent quality would have taken 1.9 times as long without AI. Usage records cover August–September; the survey ran September 30–October 7.

Our interpretation. Editorial interpretation: assistance across research tasks could change scientific capacity, including work relevant to future AI. Whether it improves verified discoveries depends on checking outputs and on experimental, staffing and funding constraints.

Main limitation. This is a lab-authored observational study of selected applicants and survey respondents. The time comparison is self-reported, not experimentally measured acceleration. Usage categories do not establish scientific correctness, and reported expectations of faster discovery are forecasts. The sample need not represent scientists generally.

Next evidence. Independent prospective comparisons of verified research outputs, including correction time and bottlenecks outside model-assisted work.