Commitment: 10 Labs, One Inconsistent Promise

Last verified: September 20, 2026. Anthropic’s Responsible Scaling Policy lays out, in one table, two different things: what Anthropic itself plans to do, and what it thinks the whole industry should do. Then it says so, plainly: “we cannot unilaterally and unconditionally commit to staying in line with the industry-wide recommendations.” Google DeepMind’s Frontier Safety Framework does something structurally similar with its highest security tier — DeepMind writes that Security Level 4 is warranted, then adds that reaching it “must be taken on by the frontier AI field as a whole,” not by DeepMind alone. Two of the most detailed safety documents in the field each publish a standard its own author says it does not expect to meet by itself. That gap — between what an organization recommends and what it commits to — is exactly the territory NIKOLAI’s Commitment element was built to map. This is the third standalone deep-dive in CASRAI’s frontier-AI-safety cluster to open a single NIKOLAI element in full, following the same crosswalk format as the capability-threshold and security-level pieces — and by row count, it is the richest element NIKOLAI has mapped so far: ten sources, four different modalities, and one flagged research trap worth its own warning label.

The short version: two labs (Anthropic, OpenAI) publish language that reads as genuinely binding. Two more (Google DeepMind, xAI) publish a mix — firm on one clause, aspirational or conditional on the next, inside the same document. Meta commits to one thing (a publication cadence) while leaving a related one (a whistleblower protocol) explicitly unfinished. The EU’s GPAI Code of Practice runs on a structural opt-in that lets a signatory commit to less than the full code and still be a signatory. The US government’s own commitments are the least visible of the ten — CAISI has announced agreements exist, not what is in them. And two entirely different documents share almost the same working title, which is reason enough for its own callout box below.

What NIKOLAI Means by “Commitment”

NIKOLAI, CASRAI’s own independent, unendorsed dictionary of frontier-AI-safety elements, defines Commitment as element N9 — part of the Commitments and Governance track. NIKOLAI’s own working definition, which it labels an editorial synthesis rather than a quotation from any single source, is: “a public statement by an identified party (a developer, a government body, or a named individual signatory) to do or not do something, carrying an explicit modality (binding / unilateral / conditional / aspirational), a scope, an effective or target date, and a status value that can be tracked over time.” The element is currently tagged proposed in NIKOLAI’s schema (release nikolai-v0.1, inside the broader nikolai-v0.2 dictionary) — NIKOLAI is upfront that this is a working definition, not a settled one.

The modality field is doing most of the work in that definition, and it is also where the ten crosswalk rows below actually disagree. A binding commitment reads as an obligation with no stated exit. A unilateral commitment is something one party does regardless of what others do. A conditional commitment only holds if a stated trigger occurs. An aspirational commitment is a stated goal without a mechanism forcing it to happen. None of the ten sources below name their own modality using NIKOLAI’s four-way split — NIKOLAI is reading each source’s own language and classifying it, which is precisely why every row is a shadow mapping rather than a confirmed one (more on what that means below).

The Crosswalk: Ten Rows, Four Kinds of Promise

Verified directly against NIKOLAI’s live Commitment element page on September 20, 2026, and cross-checked against primary source documents where those documents are public. Confidence and evidence tier are NIKOLAI’s own labels for each row, not CASRAI’s assessment of how “good” a commitment is.

Organization	Source	Match	Evidence	What it actually says
Anthropic	Responsible Scaling Policy v3.4	Exact	High	Splits its own planned mitigations from its industry-wide recommendations, and says outright it cannot unilaterally commit to the latter.
OpenAI	Preparedness Framework v2	Exact	High	“We won’t deploy these very capable models until we’ve built safeguards to sufficiently minimize the associated risks of severe harm.”
Google DeepMind	Frontier Safety Framework v3.1	Close	Medium	Aspirational: “aim to share relevant information with appropriate government authorities” (§5.2), not a stated obligation to do so.
xAI	Frontier AI Framework, 30 Jun 2026 draft	Close	Medium	Modality split inside one document: a firm annual risk assessment, versus “may” for trigger-point evaluations and for any shutdown response.
Meta	Advanced AI Scaling Framework v2	Close	Medium	Firm publication commitment per release (§2.2.1); a related whistleblower protocol is explicitly still “developing” (§2.3.1).
EU	GPAI Code of Practice	Exact	High	Full signatories adhere to all three chapters; xAI signed only the Safety and Security chapter and must meet the other two “via alternative adequate means.”
US Government	NIST / CAISI	Close	Low	CAISI “announced new agreements with Google DeepMind, Microsoft and xAI” — but the terms of those agreements are not public. A genuine transparency gap, not a mapping CASRAI can complete.
METR	Common Elements of Frontier AI Safety Policies	Close	Medium	Its own nine-element taxonomy uses two different phrasings side by side — “Commitments to…” for halt conditions, “Intentions to…” for evaluation method, accountability, and policy updates — without treating them as synonyms.
Safety Framework Cards (discovery)	SSRN working paper	None (unverified)	Low	Uses the phrase “mitigation commitments” in a discovery snippet only; the full paper is paywalled and was not read for this pass. Flagged honestly rather than dropped.
Seoul Summit signatories	Frontier AI Safety Commitments (2024) / “Pacing the Frontier”	Exact	Medium	Two org-level and individual-level commitment instruments — see the false-friends note directly below.

Two Labs, Two Kinds of Escape Hatch

Anthropic’s RSP v3.4 is the cleanest case of a document that separates what it will do from what it thinks should happen. Its own text: “The distinction between our plans as a company… and our industry-wide recommendations… reflects the limitations of any single AI developer’s ability to ensure safety across the industry. In particular, we cannot unilaterally and unconditionally commit to staying in line with the industry-wide recommendations.” That is not evasive language — it is Anthropic explaining, in its own document, exactly why a recommendation and a commitment are not the same speech act, and NIKOLAI’s crosswalk records that distinction as the row’s defining feature rather than glossing over it.

OpenAI’s Preparedness Framework v2 reads differently. Its governing sentence — “We won’t deploy these very capable models until we’ve built safeguards to sufficiently minimize the associated risks of severe harm” — has no stated exception and no industry-wide/company-only split. NIKOLAI scores both Anthropic and OpenAI Exact/High, but for different reasons: Anthropic for the precision of its own hedge, OpenAI for the absence of one.

Google DeepMind and xAI: Commitment That Changes Mid-Sentence

DeepMind’s Frontier Safety Framework v3.1 commits to a process (running its own risk assessment protocol) but not to an outcome once that process flags a problem: “if we assess that a model has reached a CCL that poses an unmitigated and material risk to overall public safety, we aim to share relevant information with appropriate government authorities.” “Aim to” is aspirational language attached to the exact moment the framework is supposed to matter most.

xAI’s 30 June 2026 draft Frontier AI Framework splits its modality inside a single paragraph: “xAI will conduct a full systemic risk assessment and mitigation process of our frontier models at least once a year” — a firm, dated commitment — immediately followed by “xAI may also conduct smaller-scale model evaluations at appropriate trigger points,” where “may” governs the release-triggered evaluations, the tiered-availability response, and the shutdown language elsewhere in the same document. One document, one modality for the calendar-driven obligation, a different one for everything that depends on judgment calls. Worth flagging separately: this framework’s own PDF metadata still labels the file a confidential working draft, and no xAI statement disambiguating draft-versus-final status was found in this pass.

Meta: One Commitment Finished, One Still “Developing”

Meta’s Advanced AI Scaling Framework v2 makes an unambiguous publication commitment: “We will publish a preparedness report for each closed or open Frontier AI release.” Two sections later, on a related but distinct governance question, the same document is candid about not being done: “Meta maintains a comprehensive whistleblower and complaint policy, and is developing further protocols.” NIKOLAI records both halves of that sentence under the same element, because both are commitment language — one with a trackable status of “complete,” one honestly marked “in progress.”

EU GPAI Code of Practice: A Commitment Structure With a Built-In Opt-Out

The EU’s Code of Practice is structured so that a full signatory adheres to all three chapters — Transparency, Copyright, and Safety and Security. But the Safety and Security chapter is, by the Commission’s own description, “only relevant to the small number of providers of the most advanced models,” and a provider can sign that chapter alone: “xAI signed up to the Safety and Security Chapter; this means that it will have to demonstrate compliance with the AI Act’s obligations concerning transparency and copyright via alternative adequate means.” The commitment instrument itself has a modular structure that most of the other nine rows do not — you can be a signatory of one-third of it.

NIST/CAISI: The Row NIKOLAI Can’t Finish Mapping

CAISI, the US government’s AI safety institute, has announced that agreements exist with three labs: “new agreements with Google DeepMind, Microsoft and xAI.” What obligations those agreements actually create — access terms, notice periods, publication rights — has not been made public, and the announcement page CASRAI attempted to verify this against returned a 404 during this research pass. NIKOLAI scores this row Close/Low, and that low score is not a data-entry choice; it is the most accurate label available for a commitment whose content nobody outside government can currently read.

METR: Two Verbs, Same Document

METR’s “Common Elements of Frontier AI Safety Policies” (December 2025) is not itself a lab’s commitment — METR is an evaluator, describing the common structure across published lab frameworks. But its own nine-element taxonomy is worth reading closely on exactly this question: elements 4 and 5 (halting deployment, halting development) are phrased as “Commitments to stop…”; elements 6, 8, and 9 (evaluation elicitation, accountability, policy updates) are phrased as “Intentions to…” The same document, describing the same category of document, reaches for two different verbs and does not treat them as interchangeable — which is itself evidence for why NIKOLAI’s modality field exists as a separate, trackable property rather than a single yes/no “has a commitment” flag.

Safety Framework Cards: An Honestly Unfinished Row

The tenth source in NIKOLAI’s crosswalk is a discovery-stage citation to a paywalled SSRN working paper (Safety Framework Cards), recorded because a public abstract or snippet used the phrase “mitigation commitments” — but the full text was never read for this element, because it sits behind a paywall CASRAI did not access in this pass. NIKOLAI marks this row’s match type “none” and its evidence “low” rather than silently dropping it or inflating its confidence to match the other nine. That is the same discipline the capability-threshold and security-level guides applied to their own weakest rows: an unverifiable source stays in the record, labeled as unverifiable, rather than disappearing.

Seoul Summit and “Pacing the Frontier”: Two Commitment Instruments, Different Signatories

NIKOLAI’s final row actually covers two related but distinct commitment instruments. The Frontier AI Safety Commitments from the 2024 AI Seoul Summit are signed by organizations — major labs commit, as institutions, to risk assessment, intolerable-risk thresholds, described mitigations, a process for halting deployment or development, and continued investment in the underlying capability. “Pacing the Frontier,” a separate open letter published in July 2026 and hosted by the nonprofit Guidelight AI Standards (with Encode AI), is signed by 1,386 individual employees of frontier AI companies, asking “that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.” One is an institutional commitment; the other is a personal-signatory request that a government act. NIKOLAI scores the combined row Exact/Medium because the commitment structure — named party, explicit ask, trackable signatory list — is present in both, even though the parties making the commitment are not the same kind of party.

A Genuine False-Friends Trap: {PACE} vs. {PTF}

Two different documents share almost the same working title, and NIKOLAI’s own research register flags the collision explicitly rather than letting it pass. {PACE} is Dario Amodei’s personal essay, “We Must Pace the Frontier,” which Anthropic’s own RSP v3.4 cites for its unilateral-commitment framing — it is a single author’s argument, not a signed instrument. {PTF}, “Pacing the Frontier,” is the entirely separate July 2026 open letter described above, hosted by Guidelight AI Standards, carrying 1,386 individual signatures. The titles are close enough — “Pace the Frontier” versus “Pacing the Frontier” — that conflating them is an easy mistake to make and an easy one to repeat once made. They are not the same document, they were not written by the same party, and they do not carry the same kind of commitment: one is argument, the other is attestation. If you see either title cited in frontier-AI-safety writing, check which one is actually meant before repeating the citation.

What the Ten Rows Add Up To

Line the ten up and the pattern is not “some organizations commit, some don’t.” Every organization here has published commitment language. The real split is over what a commitment is allowed to leave unresolved. Anthropic and OpenAI leave the least unresolved — Anthropic by naming its own limits explicitly, OpenAI by stating a flat rule with no carve-out. DeepMind and xAI leave the modality itself unresolved, shifting between firm and aspirational language inside the same document depending on which clause is under discussion. Meta leaves one specific, named piece unfinished and says so. The EU leaves the scope of who is bound negotiable, by chapter. The US government leaves the content of its own commitments unpublished. METR, describing all of this from outside, ends up needing two different verbs to cover what it sees. And two commitment instruments eighteen months apart, aimed at the same underlying worry, wound up with names close enough to trade places by accident.

None of that means any one organization is acting in bad faith. A conditional or aspirational commitment, honestly labeled as such, is more useful than an unconditional one nobody actually expects to be kept. The problem NIKOLAI’s Commitment element is built to solve isn’t that commitments vary — it’s that the word “commitment” alone doesn’t tell a reader which of the four modalities they’re looking at, and right now, checking requires reading each source’s own language directly, the way this page just did.

Where NIKOLAI Fits In

This page is CASRAI’s own deep-dive on NIKOLAI’s Commitment element, N9 in the Commitments and Governance track of CASRAI’s frontier-AI-safety dictionary. NIKOLAI is not affiliated with, run by, or endorsed by any of the ten organizations above, and none of them has been consulted on how NIKOLAI classifies their language. Every row in the table — all ten — is what NIKOLAI calls a shadow mapping: CASRAI’s own independent reading of a published document, carrying its own confidence label (Exact/Close/None) and evidence tier (High/Medium/Low), and nothing more than that unless the organization in question files an explicit Mapping Declaration confirming how it actually uses the term. As of this writing, none of the ten has done so.

Commitment sits next to two other N-track elements published as their own guides in this same content wave. Incident Type (N7, in the Incidents track) is the controlled vocabulary for classifying what actually happened once a commitment like a halt condition or a reporting duty gets triggered. Risk Domain (N1, in the Actors, Models and Scope track) is the classification layer underneath both — the CBRN/cyber/loss-of-control/manipulation categories that a commitment’s scope field and an incident’s type field both draw from. Read together with this page, and with the two guides that established this element-deep-dive format — Capability Thresholds: 14 Labs and Regulators, One Undefined Term and Security Level: Nine Labs and Regulators, One Undefined Standard — the four pages trace one thread: a threshold is crossed, a security posture is supposed to hold, a commitment is supposed to activate, and an incident is what gets recorded if it doesn’t.

Frequently Asked Questions

What does NIKOLAI mean by “Commitment”?

NIKOLAI’s Commitment element (N9) is a public statement by an identified party — a developer, a government body, or a named individual signatory — to do or not do something, carrying an explicit modality (binding, unilateral, conditional, or aspirational), a scope, a target date, and a status that can be tracked over time. NIKOLAI itself labels this an editorial synthesis, not a quotation from any single source.

Which organization has the most binding commitment language?

Of the ten sources in this crosswalk, OpenAI’s Preparedness Framework v2 reads as the most unconditional: “We won’t deploy these very capable models until we’ve built safeguards to sufficiently minimize the associated risks of severe harm,” with no stated exception. Anthropic’s RSP v3.4 is scored at the same confidence and evidence level, but for the opposite reason — it explicitly separates what Anthropic commits to from what it only recommends industry-wide.

What is the difference between {PACE} and {PTF}?

{PACE} is Dario Amodei’s personal essay “We Must Pace the Frontier,” cited by Anthropic’s own RSP v3.4. {PTF}, “Pacing the Frontier,” is a separate open letter published in July 2026, hosted by the nonprofit Guidelight AI Standards, carrying 1,386 individual signatures from frontier-AI-company employees. They are not the same document.

Is NIKOLAI’s crosswalk an official or endorsed mapping?

No. Every row is a shadow mapping — CASRAI’s own independent reading of what each organization has published — unless that organization has filed an explicit Mapping Declaration confirming it. As of this writing, none of the ten organizations and signatory groups in this crosswalk has done so.

Why does NIKOLAI include a row it says is unverified?

The Safety Framework Cards row is included, and honestly labeled match type “none” and evidence “low,” because the underlying paper is paywalled and was not read in full for this pass. NIKOLAI’s own practice is to record a source and flag its limitations rather than omit it or overstate confidence in it.

Commitment: Ten Labs and Regulators, One Inconsistent Promise