AI21 Labs

• No reference of any kind to Indigenous knowledge, data sovereignty, or the CARE principles across the homepage or about page.

• No acknowledgment of Indigenous data sovereignty or CARE principles
• No consultation with Indigenous communities mentioned
• No recognition of oral, relational, or non-textual knowledge traditions
• Training-data provenance is silent, so extractive practices cannot be ruled out

An enterprise foundation-model vendor whose public-facing surface is organised entirely around regulated-industry deployment (Defense, Finance, Healthcare). Indigenous knowledge is wholly absent; nothing here even gestures at data sovereignty. Floor score.

• Academic lineage is named explicitly — Stanford professors (Yoav Shoham, Dan Jurafsky, Sebastian Thrun), Hebrew University and Technion researchers — which acknowledges an institutional history.
• Israeli national/academic context is implicit in the founding pedigree.

• No acknowledgment of colonial or extractive data legacies
• No discussion of the geopolitical economy of compute (GPU access, supply chains, labour)
• No transparency about regulatory constraints or jurisdictional positioning
• No historical humility about what foundation models inherit

Names an academic genealogy, which is more than zero, but treats history purely as a credential. No reckoning with the material or colonial histories that condition foundation-model development. Score 2.

• Long-context processing (Jamba) is foregrounded as a capability, which has latent multilingual relevance, but no language coverage is stated.
• Founder Dan Jurafsky is a computational-linguistics figure; cross-lingual depth is plausible but unclaimed on these pages.

• No stated multilingual support, even token presence
• No preservation of culturally specific reasoning patterns
• No consultation with cultural or linguistic scholars surfaced
• Western enterprise/business logic is presented as the universal frame

The product surface is monolingual-by-implication and organised around Global-North enterprise verticals. Latent linguistic competence exists in the team but is neither claimed nor operationalised publicly. Score 2.

• Claims of traceability and auditability: 'Each decision the system makes is traceable and auditable' — a verifiability-adjacent commitment, if narrow.
• Jamba is a published, technically distinctive architecture (hybrid SSM-Transformer), giving the lab some real scientific track record beyond marketing.
• Targets regulated, evidence-sensitive sectors (compliance, precision).

• No independent audits of training data or bias disclosed
• No third-party replication protocols
• No systematic known-limitation disclosures (hallucination, failure modes)
• 'Closed' openness level — weights are not open for external verification

Some genuine scientific substance (Jamba architecture, auditability language) lifts this above the floor, but as a closed-weights vendor with no published audits, replication protocols, or limitation disclosures on these pages, evidentiary openness is thin. Score 3.

• The register is uniformly instrumental — reliability, control, productivity. No affective or aesthetic dimension is present.

• No acknowledgment of affective or intuitive dimensions
• No space for ambiguity or poetic uncertainty
• No recognition of emotional labour
• Efficiency is the only mode of attention on offer

A purely efficiency-and-control vocabulary with zero affective surface. 'Superproductivity' is the antithesis of poetic attention. Floor score.

• Agentic systems ('AI agents', 'Maestro: reliable knowledge agents') are explicitly the product, so the lab is shaping agentic futures.
• Auditability/traceability gestures faintly at governance of those agents.

• No engagement with labour-displacement risk — 'superproductivity' implies it without naming it
• No environmental or compute-cost disclosure
• No democratic or participatory governance of agentic systems
• No inclusive deliberation about whose future 'superproductivity' serves

Actively building agentic futures while leaving their distributional, labour, and environmental consequences entirely unexamined in public. The futures shaped are enterprise-owners'; no one else is at the deliberation table. Score 2.

• No marginalised constituency appears anywhere on the audited surface.

• No participatory design with Global South developers
• No disability-community accessibility commitments
• No labour-representative engagement
• No compensated feedback channels

The table is set for enterprise buyers and no one else. No accessibility, no Global South participation, no labour voice, no compensated feedback. Floor score.

• The communications are uniformly solemn and self-serious; no irony, self-interrogation, or named contradiction.
• The coinage 'superproductivity' is, read against the grain, an unintentional self-parody — productivity inflated past the point of meaning.

• No willingness to name contradictions (a 'trust' vendor selling to 'Defense' is left unexamined)
• No disciplined use of irony or paradox
• No space where the official narrative is tested by its opposite
• The lab's own seriousness is treated as exempt from audit

Zero trickster capacity. The most generative inversion (trust-marketed AI sold into Defense; 'humanity' served via 'superproductivity') is precisely what the copy is built to not notice. Floor score.