OpenAI wants outsiders to grade frontier AI safety—using OpenAI's answer key

Technology analysis with satire // SOURCED REPORTING

On September 22 OpenAI proposed how third parties should assess frontier-AI safety claims. A day later its CEO told the UN humans must stay in control. Real documents. Still vendor homework, not a finished independent exam.

What the slide says: outside experts should pressure-test frontier AI with safety cases, real access, and reports that separate findings from spin.

What it still requires: a pinned model version, assessors who can publish uncomfortable results, and proof that applies to the tool in your tab—not the speech in the press room.

OpenAI spent two late-September days building that vocabulary in public. First a principles document for third-party assessments. Then CEO remarks at the UN Security Council about human control, shared standards, and not sprinting on hope. You asked whether the system is safe. The company answered with a syllabus for how someone else should check. That is not nothing. It is also not a seatbelt already clicked.

On September 22, 2026, OpenAI published proposed priorities and principles for third-party assessments of frontier-AI safety claims, safeguards, evaluations, and incidents. On September 23, CEO Sam Altman told the UN Security Council that powerful AI should remain under human control and that countries need shared standards plus incident reporting. Both are OpenAI positions, not completed independent audits.

ANALTHROPIC’S
DAILY TOP 5:
HOW NOT TO GET
FUCKED BY AI

Fill four nouns before you upgrade trust

Anyone reading 'third-party assessed' marketing after this announcement When a page or sales note claims third-party assessment or a safety case, write four blanks—claim tested, model version, who assessed, what could be published without vendor edit—and stop nodding until each has a noun or an explicit 'unknown.' OpenAI's own principles push scoped claims and reports that separate findings from interpretation; adjectives alone fail that bar. A four-line note sits beside the pitch with fields filled or marked unknown. This is buyer hygiene, not a substitute for a formal audit you are not equipped to run. [S1]

CEO stage lights are not your statement of work

Freelancers whose clients forward UN clips as delivery policy Add one scope sentence to the next client email: final deliverables still need your human review against the brief; public safety speeches and vendor principles are background context, not acceptance tests for this workflow. Human-control rhetoric does not move liability when a bad draft ships with your name on it. The emailed or signed scope contains that human-ownership line. Not legal advice; it only keeps keynote language out of your acceptance criteria. [S2]

Renewal email beats keynote faith

Small-business owners expanding AI seats this quarter Before renewing or adding seats, email the vendor three written questions: which safety claims cover our exact tier, whether any third-party assessment report exists for that version, and whether findings can be shared under NDA. OpenAI lists scoped claims, disclosed conflicts, and actionable findings as assessment features—use the same checklist on whoever invoices you. Replies saved next to the contract; missing answers tagged as open risk in your notes. A vendor can refuse; refusal is data, not a cue to threaten litigation. [S1]

Incident reporting starts as a one-pager you own

Builders shipping features that call frontier models Draft a one-page internal incident template: date, model ID, prompt class, observed failure, human decision, and whether anything crossed your data boundary. Require it for high-impact misses even before any external form exists. OpenAI publicly argued for faster incident reporting and monitorability; waiting for a global template is how nothing gets written. Template lives in your wiki or repo and one dry-run is filed. Internal notes are not public disclosure and do not prove model alignment. [S2]

Human control is a permission map

Everyday users enabling assistants, agents, or connectors List every inbox, drive, calendar, and payment surface your assistant can touch; revoke access you cannot explain in one sentence; keep named human approval for external sends and money movement. UN remarks stressed human control and oversight; account permissions are the slice you actually operate. A dated access list or screenshot from this week; high-risk actions still require you. Revoking local access does not audit the vendor; it only limits your blast radius. [S2]

OKAY, HERE’S THE ACTUAL SHIT

OKAY, HERE'S THE ACTUAL SHIT.

On September 22, 2026, OpenAI released Priorities and principles for effective third party assessments. The company says third-party reviews should examine safety cases, safeguards, capability evaluations, and critical misalignment incidents. It describes a safety claim as a specific assertion that can be checked against evidence, and a safety case as a structured argument that includes evidence, assumptions, uncertainties, and remaining risks. [S1]

In the same document, OpenAI proposes concrete features it presents as making assessments more meaningful: clearly scoped claims, proportionate assessor access, transparent methods and uncertainties, disclosed conflicts, security controls, actionable findings, and responsible publication practices. It says reports should distinguish direct findings from interpretation and identify what the evidence supports. [S1]

Those lines are OpenAI's proposal. The text does not establish that a particular model, safeguard, or assessment has already been independently validated under the framework, and it does not create an industry-wide requirement. [S1]

On September 23, 2026, OpenAI published Sam Altman's remarks to the United Nations Security Council. As delivered in the company transcript, he said powerful AI systems should remain under human control; that companies should not accept too much technological risk merely to keep competitive pace; and that systems should not be built without a strong case for alignment, monitorability, and safety guarantees. [S2]

The remarks call for national and international frontier-AI standards for measuring capabilities, assessing risks, judging safeguards, preserving human oversight, and incident reporting, and they argue such standards should not lock in incumbents or favor one business model. That is a stated policy position and a speech transcript, not documentation of a binding global regime already in force, and not independent proof that OpenAI's systems meet the standard described. [S2]

WHY AN ACTUAL HUMAN SHOULD CARE

WHY AN ACTUAL HUMAN SHOULD CARE.

If you buy tools, run a small team, freelancing pipeline, or school-adjacent workflow, "third-party assessed" is about to become ambient marketing weather. Principles documents travel faster than access agreements. The honest split is simple:

What the deck says: assessments should be scoped, evidence-backed, conflict-aware, and clear about uncertainty. [S1]

What your decision still requires: which claim, which model version, who paid whom, what the assessor could publish, and whether any of that maps to the product tier you actually call.

OpenAI's framing is not automatically empty. Preferring safety cases over vibes, separating findings from interpretation, and pushing for incident reporting and human oversight are coherent asks. Shared measurement standards, if they ever arrive without turning into incumbent armor, could give ordinary buyers a common checklist instead of seventeen incompatible keynote PDFs. [S1][S2]

The lived risk is category error. A UN sentence about human control does not revoke a connector permission on your laptop. A vendor principles post does not replace your acceptance test, your data boundary, or your named human approver for money movement. When the next pitch deck launders this week into "we're independently covered," your job is to downgrade poetry into four nouns and keep working until the nouns exist.

WHAT WE DON’T KNOW

WHAT WE DON'T KNOW.

This packet does not show which third parties will receive meaningful access, on what timeline, under what disclosure limits, or with what right to publish negative findings. It does not show whether other labs will adopt OpenAI's proposed principles, ignore them, or negotiate softer cousins. [S1]

The UN remarks frame standards and incident reporting as goals; they do not document enforcement, membership, or effective oversight already operating at global scale. [S2]

We also lack, in the supplied evidence, specific independent scores, incident tallies, or completed assessment reports produced under the new rubric for named production systems. A call for a "strong case for alignment" is not the same thing as the public case.

SATIRE: HUMANITY SUPPORT

Issue: Third-party safety assessment returned 'needs human' and a PDF of the company's preferred grading curve.

Response: Thank you for contacting Humanity Support. We have logged your request for independent scrutiny of planetary autocomplete. Your case has been routed to Rubric Design, which reports to the same office that wrote the exam. Please remain under human control while the inspector practices stamping. Estimated completion: after the capability lands.

Status: Closed — vocabulary upgraded; evidence still pending

OpenAI did the corporate two-step cleanly: publish how third parties ought to grade frontier safety, then tell the UN the species should keep the steering wheel. Take the vocabulary if it helps you ask sharper questions. Do not confuse a well-written syllabus with a completed exam, and do not outsource the last approval to a speech.

Your replacement would like you to trust the process. Keep the receipt instead.