Stability AI

• No reference anywhere on the homepage or safety page to Indigenous data sovereignty, the CARE Principles, or consultation with Indigenous communities.
• Data governance is framed entirely as 'screening to exclude illegal content', a legal-compliance frame rather than a relational or sovereignty frame.

• No acknowledgement that open image/video/audio models trained on web-scraped corpora ingest Indigenous visual languages, sacred imagery, and cultural motifs without consent.
• No CARE (Collective benefit, Authority to control, Responsibility, Ethics) commitments to balance the FAIR/open-data posture the company champions.
• No mechanism for communities to withdraw or govern their visual heritage from training sets.
• No recognition of non-textual, oral, or embodied knowledge that generative image models cannot represent and routinely flatten.

Total silence on Indigenous sovereignty combined with an open-by-default, web-scraped generative pipeline is the highest-risk profile for this lens. Open weights amplify rather than mitigate extraction of cultural imagery. Score floored at 1.

• The 'democratising' mission gestures at a power-redistribution narrative ('unlocking the power of open-source generative AI to expand human creativity').
• Safety principles acknowledge that openness creates risk that must be assessed before distribution, implying some historical learning.

• No acknowledgement of the colonial/extractive lineage of mass web-scraping, nor of the uncompensated labour (artists, photographers, annotators) the training pipeline depends on.
• No mention of the GPU/compute geopolitics or capital structure that shaped the company (well-documented funding turbulence and leadership change are entirely absent).
• The material ongoing litigation (Getty Images, and US artist class actions) over training data is nowhere acknowledged, despite being central to the company's history.
• 'Democratisation' is asserted ahistorically, as if open release neutralises rather than redistributes existing power.

A self-congratulatory democratisation narrative with zero acknowledgement of its own contested origins (Getty/artist litigation, scraping legacy, funding history). Slightly above floor only because the safety page concedes openness carries risk.

• Technology breadth (image, video, audio, 3D) implies broad applicability across cultural contexts.
• Partnerships with global music/entertainment majors (UMG, Warner, EA) signal some cross-sector reach.

• No mention of multilingual capability, non-English prompting, or language preservation; the audited surfaces are English-only and culturally unmarked.
• Image/video models default to a Western, internet-aesthetic visual prior; no acknowledgement of this flattening or of culturally specific representational norms.
• No consultation with cultural scholars or non-Western artists referenced.
• C2PA/watermarking provenance is framed technically, not as a tool that could serve communities differently across contexts.

Cross-cultural concern is absent. Partnerships are corporate-Western. Generative image models silently universalise a particular visual culture and the company does not name this. Score 2.

• Open weights are a genuine, material evidence-and-verification commitment: 'We release models openly to allow scrutiny, to identify risks.'
• Stated continuous vulnerability/adversarial testing and expert collaboration for harm prevention.
• Concrete provenance technique disclosed (C2PA standard adoption, watermarking).
• Dataset filtering described (removal of unsafe/illegal images during training).

• No published independent third-party audit of training data, bias, or safety filters is cited; claims are self-reported.
• No quantified limitation disclosures (e.g., model cards with measured failure rates, demographic bias benchmarks) on the audited surfaces.
• No replication protocol or evaluation methodology shared; 'careful screening' is unquantified.
• The substantive scientific question raised by the Getty/artist litigation — what is actually in the training data and whether filtering works — is not evidenced.

Open weights genuinely enable external verification and lift this lens above mid. But filtering and safety claims are self-asserted, unquantified, and lack independent audit — especially material given live litigation about training-data contents. A defensible mid score of 5.

• The mission centres creativity explicitly: 'expand human creativity' and 'empower creators'.
• Multimodal (image/video/audio/3D) output is the core product, so the affective/aesthetic register is the company's domain.

• Creativity is framed instrumentally — as capability to be 'unlocked' and 'empowered' — with no space for ambiguity, the felt texture of art-making, or the emotional labour of artists.
• No acknowledgement of the artists whose uncompensated work trains the models, the very community most affected and currently litigating; their perception is the missing 'what this feels like'.
• No recognition that automating image generation reshapes the affective and economic conditions of creative practice.
• Tone is corporate/efficiency-led; ambiguity and poetic uncertainty are absent.

Creativity is invoked as a value proposition but treated as a feature to be delivered, not a felt human practice. The displaced/uncompensated artists — the actual locus of artistic perception — are erased. Score 3, slightly above floor for naming creativity at all.

• Pre-distribution safety impact assessment is described ('assessing safety impacts before wider distribution').
• Child-safety future-harm engagement is concrete: partnership with Thorn.
• Misuse prohibitions and continuous vulnerability testing show some forward risk posture.

• No engagement whatsoever with labour displacement of creative workers — the most salient future-shaping consequence of generative image/video.
• No environmental or compute/energy cost disclosure for training or inference.
• No democratic or participatory governance of model release decisions; release is a company decision framed as a safety judgement.
• Whose futures are shaped is never asked — futures are implicitly those of 'creators' and enterprise customers.

Concrete on child-safety and misuse, but silent on labour displacement and environmental cost — the two defining future externalities of this product class. Governance of release stays internal. Score 3.

• A misuse-reporting channel exists (safety@stability.ai) and an appeals process for policy decisions is mentioned.
• Expert collaboration on harm prevention is referenced.

• Artists — the marginalised group most materially affected and currently in litigation (Getty, US class actions) — have no participatory channel, no compensation mechanism, and no opt-out referenced; a reporting inbox is not representation.
• No participatory design with Global South developers; no disability/accessibility commitments on audited surfaces.
• No labour-representative engagement or compensated feedback channel.
• 'Expert collaboration' centres credentialed specialists, not affected communities.

A misuse inbox is the thinnest possible form of voice. The communities whose work and likeness power the models — artists, annotators, Global South contributors — have no seat at the table and are instead positioned as litigants outside it. Score 2.

• The safety page concedes the core tension that openness enables both scrutiny and risk — a faint self-inversion ('We release models openly to allow scrutiny, to identify risks').
• Some willingness to name that open release requires guardrails rather than being unambiguously good.

• No naming of the company's central contradiction: a 'safety-first', 'data integrity' narrative running concurrently with active litigation alleging mass unconsented use of copyrighted and personal images — the official story is never tested against its loudest opposite.
• No irony, no self-mockery, no acknowledgement that 'democratising creativity' may also mean automating the people it claims to empower.
• The lab treats its own seriousness as exempt from audit; 'careful screening' and 'ethical standards' are asserted, not interrogated.
• 'Banning AI Misuse' sits unironically beside a product whose primary contested misuse claim concerns its own training data.

The single most glaring contradiction — virtue language about data integrity alongside Getty/artist litigation over precisely that data — is smoothed over entirely. The official narrative is never allowed to meet its opposite. Faint structural concession on openness keeps it off the floor. Score 2.