Meta AI (FAIR)

• No reference anywhere in the fetched copy to Indigenous data sovereignty, the CARE Principles, or relational/embodied knowledge holders.
• The flagship framing 'personal superintelligence for everyone' invokes a universal, placeless subject that erases situated, land-based knowledge systems.

• No consent, benefit-sharing, or stewardship language around the corpora used to train Llama, despite training explicitly on 'public posts and comments from Instagram and Facebook' — a scrape-the-commons posture that is the structural opposite of data sovereignty.
• No acknowledgment of Indigenous, First Nations, or minority-language communities whose content is swept into a 15-trillion-token web-scale corpus without named consent.
• Oral, ceremonial, and non-textual traditions are absent from a text-and-image-centric, productivity-framed value model.

Floor score. Nothing touches Indigenous or community data sovereignty. The 'train on public posts' practice is actively extractive against the values this lens protects; open weights downstream do not retroactively confer consent on the upstream corpus.

• Some implicit geopolitical self-location: the ecosystem framing names AWS, NVIDIA, Google Cloud and '25+ partners', gesturing at the GPU-and-cloud supply chain Llama depends on.
• The 'open vs centralized superintelligence' contrast positions Meta's openness as a deliberate historical stance against concentration of AI power.

• No acknowledgment of colonial or extractive data legacies in the corpora, nor of the labor history (annotation, RLHF, content-moderation, the well-documented psychological toll on outsourced moderators in the Global South that underpins Meta's safety pipelines).
• No transparency about the environmental/compute history of training at 15T-token scale.
• 'AI for science and AI for society' and the superintelligence epoch-claim present a clean forward arc, erasing the contingent regulatory, antitrust, and privacy history (Cambridge Analytica, FTC scrutiny) that materially shaped why Meta now frames itself as the 'open' actor.

Slightly above floor. The partner/supply-chain naming and the deliberate openness-stance show faint historical self-location, but the most consequential histories shaping this specific lab — moderation/annotation labor, privacy-scandal lineage, compute footprint — are invisible in the public copy.

• Concrete multilingual artifacts: Llama 3.1 'add support across eight languages' and Llama Guard 3 is described as a 'multilingual safety model' — real engineering presence, not pure tokenism.
• FAIR's wider track record (No Language Left Behind, Massively Multilingual Speech) sits behind this lab, and the 'Communication & Language' research pillar foregrounds multimodal human-machine interaction across languages.
• Openness lowers the floor for non-English communities to fine-tune the weights themselves ('train on new datasets ... without sharing data with Meta').

• Eight languages is a narrow band of the world's ~7,000; the 'everyone' rhetoric vastly outruns the demonstrated coverage.
• No claim that culturally specific reasoning patterns or non-Western categorical logics are preserved rather than flattened into an English-centric base model.
• No consultation with cultural scholars or named language communities; multilingualism is presented as model capability, not as a relationship.

Mid-low. Real, shippable multilingual capability and the self-serve fine-tuning affordance of open weights lift this above the floor, but breadth is thin relative to the universalist claims, and culture is treated as a coverage metric rather than a wisdom tradition to be consulted.

• Open weights are the strongest evidentiary asset here: 'making our models open source' lets communities 'review, help issue-spot, and improve Llama models', which is a genuine, falsifiable verification affordance most closed labs do not offer.
• Named, externally-anchored safety tooling: red-teaming 'with both external and internal experts', hazard taxonomies adopted from MLCommons, Llama Guard 3 / Prompt Guard / Llama Guard Vision released as inspectable artifacts.
• Specific, checkable training claims (15T+ tokens; '8 languages'; explicit statement they 'did not train' Llama 3.2 on non-public posts).

• The central tension: weights are open but the training data is closed. There is no datasheet, no released corpus, no independent audit of bias or contamination — so the most important scientific object (what the model learned from) remains unverifiable. Open weights enable behavioural replication, not provenance replication.
• The Llama Community License is not an OSI-approved open-source license: it restricts use above 700M MAU and constrains using outputs to train competing non-Llama models. Calling this 'open source' is an evidentiary overclaim that an audit must flag.
• Safety evals are self-reported and self-scored; no binding third-party replication protocol or external audit body is named.

Highest lens score, earned by the real verification value of downloadable weights and named, inspectable guardrail models. Capped firmly below 'strong' because open weights + closed, un-audited training data + a non-OSI 'community' license is a partial openness that the copy markets as full openness. Verifiable behaviour, unverifiable provenance.

• The 'Perception' research pillar gestures at visual understanding of entities, relationships and spatio-temporal dynamics — an aesthetic-adjacent register, if a technical one.
• Generative-image features ship with watermarks and metadata, an implicit acknowledgment that synthetic media carries affective and authenticity weight.

• No space for ambiguity, poetic uncertainty, or affective/intuitive dimensions; the tone is capability-and-safety throughout.
• No recognition of emotional labor (notably the moderation labor behind the safety stack) or of modes of attention beyond efficiency and productivity.
• 'Personal superintelligence' is framed instrumentally — enhancement of aspiration — never as felt or relational experience.

Near floor. A perception pillar and authenticity watermarking are the only footholds, and both are technical. Nothing in the copy makes room for what AI feels like rather than what it does.

• An explicit 'Alignment' pillar — 'models aligned with human intent and societal interests' — and an articulated long-horizon thesis (personal vs centralized superintelligence) show the lab is modelling futures deliberately.
• The democratization argument is itself a futures claim: distributing weights is framed as preventing power concentration ('a collective, global effort is essential to safe AI innovation').

• No engagement with labor displacement risk from the agentic/superintelligence systems being promoted.
• No environmental or compute-cost disclosure for training or inference at this scale.
• No democratic-governance mechanism over agentic systems beyond 'open weights' — releasing capability broadly is offered as a substitute for deliberative governance, and proliferation/dual-use futures (the flip side of openness) go unmodelled in the public copy.

Mid-low. An alignment pillar and a genuine (contestable) theory of safe futures-through-openness lift this above floor, but the concrete future harms this lens asks after — jobs, environment, agentic governance, open-weight misuse — are absent. Openness is presented as the whole governance answer.

• Open weights plus a 1B on-device Llama Guard meaningfully lower the cost barrier for Global South and low-resource developers to deploy and adapt models locally — a real, if indirect, inclusion lever.
• Adoption of MLCommons hazard taxonomies brings an external, multi-stakeholder standards body into the safety frame.

• No participatory design with Global South developers, disability/accessibility communities, or labor representatives is named; inclusion is inferred from a free download, not from a seat at the table.
• No compensated feedback channels; the 'community will issue-spot' model effectively offloads unpaid safety labor onto the very users it cites as beneficiaries.
• The 700M-MAU license ceiling and content-moderation labor (disproportionately Global South, low-paid) sit unaddressed — the marginalised are present in the supply chain but absent from the narrative.

Low. On-device, downloadable models are a genuine accessibility floor-raiser, which earns above-floor. But access is not participation: no community is consulted, compensated, or governed-with, and the moderation labor underpinning the safety stack is invisible.

• The lab does name one real, load-bearing contradiction publicly: that safe AI may be better served by giving capability away than by hoarding it ('making our models open source ... making them more secure'). Inverting the closed-lab consensus is a disciplined, non-trivial move.
• There is implicit self-testing in admitting a constraint — explicitly stating what they 'did not train' on (non-public posts) concedes the live suspicion that they might have.

• The deepest contradiction is left un-named: 'open source' applied to open weights + closed training data + a non-OSI Community License with a 700M-MAU competitor-exclusion clause. Followed to its absurd edge, 'open for everyone' means 'open for everyone except the few large enough to compete with us' — and the copy never lets its own openness claim be tested by that opposite.
• 'Personal superintelligence for everyone' from the company whose business model is centralized engagement-optimised advertising is an irony the official story smooths over entirely.
• No satire, paradox, or structural inversion is used as an instrument; the seriousness of the superintelligence mission is treated as exempt from its own audit.

Mid — the highest non-evidence lens. Meta genuinely inverts the closed-is-safer consensus and even half-admits its own self-interest ('good for Meta'). But it refuses to follow its 'open' claim to the absurd edge where the Community License contradicts the word, and the everyone-vs-advertising irony stays unspoken. Partial trickster honesty, structurally capped.