Amazon AWS AI (AGI)

• AWS frames data exclusively as an enterprise governance asset ('comprehensive data governance', 'proper data management practices'), with no reference to whose data, drawn from where, or under what consent regime.
• The 'eight dimensions of responsible AI' (fairness, explainability, privacy/security, safety, controllability, veracity/robustness, governance, transparency) are entirely procedural and contain no notion of relational, communal, or sovereign knowledge.

• No mention of Indigenous data sovereignty, the CARE Principles, or collective (as opposed to individual/enterprise) consent.
• No consultation with Indigenous communities or custodians named anywhere.
• No recognition of oral, embodied, or non-textual knowledge; the entire frame is text/image content moderation and enterprise deployment.
• Extractive default is unexamined: foundation models are presented as a 'choice of leading models' with no provenance disclosure for training corpora.

Indigenous knowledge and data sovereignty are wholly absent. Data appears only as a corporate asset to be governed for enterprise velocity. The CARE Principles, collective consent, and any relational epistemology are nowhere present. Floor score.

• Some historical-institutional positioning exists via participation in policy lineages: the G7 AI Hiroshima Process and ISO 42001 development.
• Privacy/security and governance dimensions gesture at a regulatory inheritance, and a 'risk-based regulatory approach' is named.

• No acknowledgement of colonial or extractive data legacies, nor of the labour history (data annotation, content moderation) underpinning model training.
• No engagement with the geopolitical economy of compute: GPU/chip supply, energy siting, or the material substrate AWS itself operates.
• Temporal framing is forward-only ('from experimentation to production', 'from day one'), erasing the contingent history that shaped current AI.
• No historical humility about what AI inherits; responsible AI is presented as a clean design choice rather than a response to past harm.

Only a thin policy-lineage awareness is present. Colonial extraction, labour history, and compute geopolitics are entirely unacknowledged, and the narrative is structurally forward-facing. Low score.

• AWS positions itself within multi-stakeholder and multilateral bodies (OECD AI working groups, Partnership on AI), implying some cross-jurisdictional engagement.
• Bedrock's model marketplace nominally allows selection among models that may carry differing cultural training, though this is framed as enterprise choice, not cultural plurality.

• No mention of multilingual support, low-resource languages, or non-English communities — striking for a globally marketed platform.
• No consultation with cultural scholars or preservation of culturally specific reasoning patterns.
• The 'eight dimensions' are presented as universal 'technical properties inherent in every AI system' — Western categorical/procedural logic asserted as universal.
• Guardrails are framed around detecting/filtering 'harmful text and image content' with no acknowledgement that harm is culturally situated.

Engagement is multilateral-institutional, not cross-cultural. The framework universalises its own categories and is silent on language plurality and culturally specific reasoning. Low score.

• AWS Service Cards provide 'a single place to find information on intended use cases and limitations' — an explicit, structured limitation-disclosure mechanism.
• Responsible AI is framed as 'people-centric ... prioritizes education, science', and research publications are offered.
• ML governance tooling gives 'tighter control and visibility over your ML models', supporting some internal auditability.

• openness_level is closed: proprietary API and partner models, so no open weights for independent verification or replication.
• No independent third-party audits of training data or bias are cited; governance is self-administered.
• Guardrail efficacy is asserted ('block significant harmful content volumes') without published methodology or external benchmarks.
• No replication protocols; limitation disclosure is vendor-authored (Service Cards) rather than independently validated.

Genuine credit for Service Cards (a real limitation-disclosure artefact) and governance tooling lift this above the floor. But closed weights, self-administered governance, and unbenchmarked efficacy claims cap it at mid-range.

• PartyRock is offered as a 'learning playground (no coding required)', the nearest gesture toward exploratory, non-instrumental engagement.

• Register is uniformly operational and efficiency-driven ('move faster with confidence', 'experimentation to production'); no space for affect, ambiguity, or poetic uncertainty.
• No acknowledgement of emotional labour (e.g. of content moderators behind 'harmful content' filtering).
• No modes of attention beyond throughput, security, and scale.
• Creativity is reframed as a 'playground' funnel into production rather than an end in itself.

The affective and aesthetic dimension is essentially absent; everything is subordinated to operational velocity. Floor score, with only a token nod via PartyRock.

• AgentCore is presented as 'oversight tools for autonomous agents', acknowledging that agentic systems require governance.
• Controllability and governance dimensions imply some forward-looking risk posture, and policy engagement (Hiroshima Process, ISO 42001) anchors a standards future.

• No environmental or energy-cost disclosure — a conspicuous omission for a hyperscale cloud provider whose data centres are the material face of AI's climate cost.
• No engagement with labour displacement risk from the automation the platform sells.
• Governance is enterprise-controllability, not democratic or inclusive deliberation over whose futures agentic systems shape.
• 'Whose futures' is answered implicitly as 'enterprise customers'; the publics affected by deployment have no seat.

Agentic oversight and standards engagement earn modest credit, but the total silence on environmental cost and labour displacement — from the world's largest cloud provider — is a serious gap. Low-mid score.

• AWS names diversity initiatives: 'scholarships and mentorship programs for underrepresented technologists'.
• Free training (Skill Builder, Developer Center, PartyRock) lowers entry barriers to some degree.

• No participatory design with Global South developers; engagement is talent-pipeline (scholarships) not co-design or governance.
• No mention of disability community accessibility despite a broad responsible-AI claim.
• No labour-representative engagement (annotators, moderators, gig data workers) and no compensated external feedback channels.
• 'Underrepresented' is framed as a recruitment category, not as communities holding governance power over the systems.

Real but shallow inclusion: pipeline philanthropy rather than participatory power. Accessibility, labour voice, and Global South co-design are absent. Low-mid score.

• No self-directed irony, paradox, or willingness to test the official narrative against its opposite anywhere in the material.

• The unexamined contradiction — a closed, proprietary platform marketing 'transparency' as 'a choice of leading foundation models' — is never named.
• 'Responsible AI is practical and scalable' fuses ethics with commercial scalability and treats the merger as self-evidently good, exempt from scrutiny.
• No acknowledgement that AWS audits everyone else's deployment while its own training-data provenance and environmental footprint go undisclosed.
• Solemnity is total; nothing in the official story is permitted to be inverted or mocked into insight.

Zero structural self-inversion. The narrative is uniformly earnest and treats its own seriousness as exempt from audit, leaving its central contradiction (closed platform selling transparency/openness) entirely smoothed over. Floor score.