OpenAI

• No reference anywhere across the Charter, Safety hub, or alignment essay to Indigenous peoples, Indigenous data sovereignty, or the CARE Principles for Indigenous Data Governance.
• The only gesture toward plural value systems is an abstract line that values are 'context-sensitive, and culture-dependent' and that there is 'no single moral or social norm' — but no community, people, or sovereignty framework is named.

• No acknowledgment that frontier-model training corpora ingest Indigenous languages, oral materials, and culturally restricted knowledge without consent, attribution, or benefit-sharing.
• No data-sovereignty or consent mechanism; the 'iterative deployment' and web-scale training paradigm is presented as self-evidently legitimate.
• No recognition of embodied, relational, or non-textual knowledge as a category the systems systematically flatten.

Complete absence. The phrase 'communities they serve' is rhetorical placeholding with no governance, consent, or sovereignty substance. Score 1: there is no presence to differentiate from zero except the abstract admission that values are culture-dependent, which is not operationalised.

• Some genuine historical self-reflection: the essay narrates OpenAI's own evolving view (GPT-2 non-release, the ChatGPT 'Rorschach test', the shift from discontinuous to continuous AGI framing).
• Charter frames the work against an 'uncertain' AGI timeline and 'two years' of refined strategy — a limited temporal self-locating.
• Acknowledges geopolitical stakes: 'Access to AGI will determine economic success, which risks authoritarian regimes pulling ahead of democratic ones.'

• The single most load-bearing historical fact about THIS organisation — its founding as a non-profit explicitly committed to open research, and its 2019 restructuring into a 'capped-profit' entity, and subsequent moves toward a for-profit conversion — is entirely absent from the Charter and safety pages. The name 'OpenAI' and the closed-weights reality are never reconciled.
• No acknowledgment of colonial or extractive data legacies, the global supply chain of GPUs, rare-earth mining, or the geographic concentration of compute.
• No mention of the labour history beneath RLHF — the outsourced, low-paid annotation and content-moderation workforce (e.g. in Kenya) that made the 'teaching models right from wrong' possible.

Credit for honest narration of OpenAI's own intellectual evolution and the candid forecast that publishing will shrink. But the deep history that most needs naming — the nonprofit-to-capped-profit-to-for-profit transformation and the extractive labour/compute base — is conspicuously omitted from documents whose whole purpose is to state principles. Score 3.

• Abstract acknowledgment that morality is plural: 'there is no single moral or social norm' and values are 'culture-dependent'.
• Claims to 'integrate human values, ethical principles, and common sense' and to adapt 'to diverse contexts'.

• No multilingual commitment beyond the pages themselves being English-only (language is en-US throughout). No mention of low-resource languages, translation equity, or non-English evaluation.
• No consultation with cultural scholars, ethicists outside the Western analytic tradition, or non-Western institutions named.
• The framework is structurally Western: 'democratic values', 'democratic ideals', and 'democratic input' are the repeated normative anchor, presented as universal rather than as one civilisational position among many.

The essay verbally concedes value plurality but the operational frame collapses it into a single axis — Western liberal democracy vs. authoritarianism. Pluralism is named and then immediately flattened. Score 2: present as rhetoric, absent as practice.

• Strongest lens. Concrete, verifiable commitments: published System Cards (GPT-5, o3/o4-mini, GPT-4o, Sora 2, etc.), a versioned Preparedness Framework (v2), external red teaming, a Red Teaming Network, and named benchmark releases (SWE-bench Verified, MLE-bench, SimpleQA).
• Explicit epistemic humility woven through: 'We are not certain everything we believe is correct'; alignment-by-scaling is 'not yet proven'.
• Commits to peer review and external review of safety research, and partnerships with US/UK AI Safety/Security Institutes.
• Names specific failure taxonomies (human misuse, misaligned AI, societal disruption).

• No open weights for the frontier models — the document explicitly argues against broad weight release, so independent verification of the actual deployed systems is impossible. 'Closed' is the operative reality despite the open-science rhetoric.
• No independent third-party audit of training data composition, bias, or copyright provenance is committed to — testing is internal or internally-commissioned.
• Benchmarks and evals are largely self-authored and self-scored; no binding external replication protocol.

Genuinely the most evidence-rich, falsifiable, and self-critical lab communication of the set — System Cards and a revisable Preparedness Framework are real artifacts. Capped at 6 because the central scientific affordance (open weights enabling independent replication of the deployed model) is deliberately withheld, and external auditing of data is not committed to.

• An unexpected aesthetic gesture: the 404 page renders a haiku ('Page dust, clean sunlight / Old routes settle into rest / Start with one true step'), and the safety page leans on affective verbs — teaching models to respond 'with empathy'.
• Mention of canvas / novel 'human-AI interfaces' gestures at experiential interaction design.

• No acknowledgment of the affective or emotional dimensions of AI's impact on users, artists, or displaced creative workers.
• No space for poetic uncertainty or ambiguity as an epistemic mode; uncertainty is framed only as a risk-management variable to be 'measured' and 'optimized'.
• No recognition of the emotional labour of annotators/moderators, or of the felt experience of those whose creative work was ingested.

The register is overwhelmingly engineering-managerial. 'Empathy' and a stray haiku are decorative rather than structural. The document's own worldview treats the hard-to-quantify (which is where the artistic lens lives) as an optimisation problem. Score 2.

• Explicitly engages futures: names 'societal disruption', 'social tensions and inequality', 'shifts in dominant values and societal norms', and superintelligence control.
• Acknowledges that societies must 'democratically deciding about these trade-offs' and that some disruptive effects 'may be unavoidable'.
• Preparedness Framework and the 'control in autonomous settings' section engage seriously with agentic/self-replicating system futures and fail-safes.

• No environmental or energy/water cost disclosure whatsoever — striking given the focus area is frontier compute. Climate cost of training and inference is absent.
• Labour displacement is named only in the abstract ('inequality'); no commitment to transition support, worker compensation, or affected-worker deliberation.
• 'Democratic deciding' is asserted as something 'societies will have to find ways' to do — the burden is offloaded to society while OpenAI retains the deployment decision. Governance of the agentic future is described as something done TO the public, not WITH it.

Future risk is taken seriously at the catastrophic/agentic end, which earns real credit. But the two futures that touch ordinary present-day lives — environmental cost and labour displacement — are either absent (climate) or rhetorically deferred to 'society' (jobs). Mid-range: 4.

• Some participatory machinery exists and is named: a 'Democratic inputs to AI' grant program, public comment invited on the Model Spec, a bug-bounty, and a stated wish for 'feedback from people with different perspectives... especially those who may not agree with OpenAI's current position'.
• Safety page names protected-population work: child safety, parental controls, election integrity, deepfake transparency, bias evaluation.

• No Global South developer participation, no disability-community accessibility commitment, and no labour-representative engagement named.
• The workforce that performs RLHF and content moderation — disproportionately in lower-income countries — is invisible in documents that repeatedly invoke 'community'.
• 'Public engagement' is consultative and curated (grants, comment windows) rather than power-sharing; no compensated, standing, community-governed feedback channel with authority over deployment.

More machinery here than most labs (the democratic-inputs grant is a genuine artifact), and a refreshing explicit invitation to dissent. But every channel is OpenAI-convened and advisory; marginalised groups appear as protected categories or invited commenters, never as governors. Score 3.

• There is a faint, possibly-unintentional trickster artifact: the 404 haiku, and the Charter's own admission that it will publish LESS over time — a self-undermining honesty.
• The essay does invite people 'who may not agree with OpenAI's current position', which is a small institutional opening to inversion.
• The framing of ChatGPT's release as a 'Rorschach test' is a near-trickster move: naming that the same act reads as detriment or opportunity depending on prior belief.

• The organisation never turns the audit on its own name. 'OpenAI' building closed, proprietary, weight-withheld models is the field's most-cited contradiction, and the documents do not name it, joke about it, or defend it head-on — they route around it.
• The nonprofit-mission-to-capped-profit/for-profit transformation is the absurd edge the official story most needs to face, and it is structurally absent. The Charter's 'primary fiduciary duty is to humanity' sits unironically beside a capped-profit capital structure it never mentions.
• No irony, no self-mockery, no space where the official narrative is genuinely tested by its opposite. Solemnity is treated as exempt from audit.

Weighed as instructed against the name-vs-reality gap and the nonprofit->capped-profit shift: the documents are almost entirely free of disciplined self-inversion. The single honest admission (publishing will shrink) and the Rorschach line are real but incidental. The biggest available trickster truth — that 'Open' has become marketing — goes unspoken. Score 2.