Cohere

• The Aya initiative covers 101 languages including 50+ described as 'underserved', which incidentally brings some Indigenous and low-resource languages into model scope.
• Cohere Labs frames its mission as 'Changing where, how, and by whom breakthroughs happen,' gesturing at participation by communities historically excluded from AI.

• No acknowledgment of Indigenous data sovereignty or the CARE Principles for Indigenous Data Governance.
• No mention of consent, benefit-sharing, or community control over the language data ingested for multilingual models — 'underserved languages' are framed as coverage gaps to be filled, not as living relational knowledge held by sovereign communities.
• No preservation of oral, ceremonial, or non-textual knowledge; the entire stack is text/retrieval-centric.
• The enterprise 'your data stays yours' commitment protects the paying customer's data, not the communities whose languages train the base models.

Multilingual breadth touches Indigenous and minoritised languages, but framing is extractive-by-default: languages are 'covered' as a benchmark frontier with no sovereignty, consent, or CARE framework. The 'your data stays yours' guarantee is conspicuously a customer-protection clause, throwing the absence of community-data protection into relief. Slightly above floor only because Aya's underserved-language focus is materially better than English-only labs.

• The research framing acknowledges that breakthroughs have historically happened in a narrow 'where/how/by whom', implicitly conceding a geographically concentrated past.
• Catalyst Grants and Scholars Programs implicitly recognise structural exclusion of researchers from underrepresented backgrounds.

• No acknowledgment of colonial data-extraction legacies underlying multilingual corpora.
• No transparency about the geopolitical economy of compute — GPU access, cloud partner dependencies (AWS, Azure, GCP, OCI), or labor in data annotation.
• No historical humility about AI's inheritances; the narrative is forward-tilted and product-clean.
• Regulatory and jurisdictional context (Canadian base, cross-border data flows) is treated only as a security/compliance selling point, not a historical condition.

There is a faint structural awareness that the past was exclusionary (the whole equity-programs framing), but no engagement with colonial extraction, compute geopolitics, or annotation labor. History appears only as a problem the programs will fix going forward, not as an inheritance the lab itself sits inside.

• Genuine multilingual depth: Command across 49 languages, Transcribe across 14, Aya across 70-101 languages with 50+ underserved.
• Global MMLU is a human-verified benchmark explicitly built to test performance 'across diverse linguistic contexts' and support 'more equitable AI evaluation.'
• Open Science Community spans 4,500 members across 150 countries.
• Human verification of the benchmark suggests some real consultation with native speakers.

• Multilingual is framed as language coverage and benchmark performance, not as preservation of culturally specific reasoning patterns or non-Western categorical logic.
• No mention of cultural scholars, anthropologists, or community linguists shaping model behavior beyond data collection and verification.
• Western enterprise-utility framing (RAG, retrieval, productivity) is treated as the universal end-use; cultural knowledge is instrumentalised toward that.
• Languages counted as integers ('101 languages') — a quantification that can flatten the heterogeneity within and across them.

This is Cohere's strongest lens. The multilingual program is substantive and the human-verified Global MMLU plus a 150-country community are real plurality signals. Capped at mid-range because the orientation is performance/coverage rather than preservation of distinct epistemologies, and culturally-specific reasoning is not engaged beyond benchmark accuracy.

• Public commitment to releasing research and models openly via Cohere Labs and the Open Science Community ('all models and findings released publicly').
• Aya and Global MMLU are concrete, publishable research artifacts enabling third-party scrutiny.
• Security posture is evidence-backed: SOC 2 Type II compliance, annual third-party audits, penetration testing, a bug bounty program, and a published 'Secure AI Frontier Model Framework.'
• Mentions adversarial testing and data lineage tracking.

• Core commercial models (Command, Embed, Rerank, North, Compass) are closed/proprietary API — no open weights for independent verification of the production stack.
• Audits cited are security audits (SOC 2, pen tests), not independent audits of training data, bias, or model behavior.
• No third-party replication protocol for the commercial models; openness is concentrated in the research/Labs arm, not the products customers actually deploy.
• Known-limitation disclosures for the flagship models are absent from these pages.

A real bifurcation: the Labs/research arm is genuinely open and verifiable (open models, public findings, Global MMLU), which lifts the score above mid; but the commercial products are closed-weight proprietary API, and the prominent third-party audits are security/compliance rather than bias/data audits. Evidence culture is strong on security, thin on model-behavior accountability.

• No meaningful engagement. Language is uniformly enterprise-functional: 'productivity,' 'retrieval,' 'security,' 'compliance.'

• No acknowledgment of affective, intuitive, or aesthetic dimensions of language and AI.
• No space for ambiguity or poetic uncertainty — everything is framed as optimization and control.
• No recognition of emotional labor (annotators, users, communities).
• Modes of attention beyond efficiency are entirely absent; efficiency and 'scalable' impact are explicit guiding principles.

Floor score. The corpus is wholly instrumental and efficiency-oriented. Even the multilingual work — the natural home for affective and aesthetic attention — is framed as coverage and benchmark performance. There is no register for feeling, ambiguity, or the non-instrumental.

• The 'Responsible' guiding principle commits to 'advancing AI safety aligned with societal values.'
• A published 'Secure AI Frontier Model Framework' signals some forward governance posture on frontier risk.
• Catalyst Grants fund 'projects driving real-world change,' gesturing at distributed future benefit.
• Data-sovereignty deployment options shape a future where enterprises retain control.

• No engagement with labor displacement risk — striking for a company selling 'workplace productivity' and agentic systems explicitly aimed at automating knowledge work.
• No environmental or energy/compute cost disclosure anywhere in the audited text.
• No democratic or participatory governance of agentic systems; 'societal values' is undefined and unaccountable.
• Future is shaped around enterprise customers (Oracle, Salesforce, SAP, financial institutions) — whose futures are centered is implicitly the buyer's, not the displaced worker's or the public's.

There is a named responsibility principle and a frontier-safety framework, lifting this off the floor, but the substance is thin: agentic workplace automation is sold with zero engagement with displacement, no environmental disclosure, and 'societal values' left undefined. The future being modelled is the enterprise buyer's, made explicit by the client roster.

• Concrete, structured programs aimed at inclusion: Research Scholars Program ('mentorship and infrastructure access for rising researchers from underrepresented backgrounds'), Catalyst Grants (free API access for academic and civic organizations), and the 150-country Open Science Community.
• Underserved-language focus (50+ via Aya) brings Global South linguistic communities into scope.
• Free participation in educational programs lowers a real access barrier.

• No disability-community accessibility commitments anywhere in the audited text.
• No labor representative engagement — data annotators and the workers behind multilingual corpora are invisible.
• Inclusion is researcher-and-developer-centric (who builds), not end-user or affected-community-centric (who is acted upon).
• No compensated feedback channels for affected communities; participation is framed as opportunity/access, not power or governance seat.

Among the better lenses: real, named, free programs that broaden who participates in research, plus genuine Global South reach. Held to mid-range because inclusion is about access for builders, not governance power for affected communities; disability and labor voices are entirely absent, and there are no compensated community feedback channels.

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

• No naming of the central contradiction: 'your data stays yours' is offered to enterprise customers while base models are trained on others' languages and web data with no equivalent sovereignty guarantee.
• No reckoning with selling 'workplace productivity' agents (which displace labor) under a 'Responsible / societal values' banner.
• The polished consensus ('industry-leading,' 'trustworthy AI,' 'every corner of the globe') is treated as exempt from its own audit.
• No space where the company's seriousness about safety is allowed to be questioned by, e.g., the closed-weight nature of the very products it sells.

Floor score. Corporate AI communications are structurally trickster-free, and Cohere is no exception. The material smooths over at least two sharp contradictions (data-sovereignty-for-buyers vs. extraction-for-training; responsibility-framing vs. labor-automating products) without a flicker of self-irony or willingness to be inverted.