Character.AI

• No reference anywhere on the homepage or Safety Center to Indigenous peoples, data sovereignty, or the CARE Principles.

• No acknowledgment of Indigenous data sovereignty or community consent over training data.
• User-generated Characters can impersonate cultural/ancestral figures with no protocol for Indigenous oral traditions or relational knowledge.
• Extractive-by-default model: 10M+ user-built personas trained on and monetised without provenance or community benefit-sharing.

An entertainment persona platform built on mass user-generated content with zero engagement with Indigenous knowledge systems or data sovereignty. The absence is total, not partial; lowest score warranted.

• Company history (founded by ex-Google Brain researchers, 2024 Google licensing deal) is well known but absent from the public-facing safety/mission text.
• Safety Center positions current protections as a response to scrutiny ('safety-by-design'), implicitly acknowledging a reactive history.

• No acknowledgment of the GPU/compute geopolitical economy or the labour behind moderation and RLHF.
• No discussion of the colonial/extractive lineage of large-scale data harvesting.
• Teen-safety framing presented without the history (litigation, public harm cases) that produced it — temporal flattening of how these guardrails came to exist.

There is an implicit reactive history visible in the safety posture, but it is never named or owned. Minimal historical humility; mostly forward-facing product marketing.

• Platform de facto supports many languages because Characters are user-generated, giving broad informal multilingual reach.
• Privacy disclosures are described as 'regional', implying some jurisdictional adaptation.

• Multilingual capacity is incidental (user-supplied), not a designed commitment; no consultation with cultural scholars.
• Western entertainment/companion framing ('your adventure', 'your world') exported as universal.
• No preservation of culturally specific reasoning patterns; safety moderation norms are not disclosed as culturally situated.

Genuine multilingual breadth exists but is emergent from crowdsourcing rather than principled inclusion; moderation and design remain a single cultural frame.

• Safety Center references 'Model training transparency' as a published document category.
• Content moderation and reporting systems are described as operational.

• No independent third-party audits of bias, training data, or safety-classifier efficacy disclosed.
• Closed weights, proprietary API — no replication or external verification possible.
• No quantified limitation disclosures (e.g., moderation error rates, false-negative rates on teen-harm content).

Claims of safety systems exist but are unverifiable: no open weights, no independent audit, no metrics. Self-attestation only.

• The entire product is affective and imaginative by design — roleplay, companionship, creative persona-building.
• Marketing language ('Your Words. Your World.', 'adventure') explicitly invites emotional and aesthetic engagement.

• Emotional labour and the affective risk of parasocial/companion attachment is engaged only defensively (teen safety), not as a recognised dimension of the experience.
• No space for poetic uncertainty about what synthetic intimacy does to users; ambiguity is flattened into 'fun'.
• The emotional weight of companion AI for vulnerable users is treated as a moderation problem, not an artistic/ethical one.

Strongest lens by far — the product is inherently expressive and affective. But the affective dimension is celebrated in marketing and policed in safety, never reflected on as a serious ethical-aesthetic question; capped at midpoint.

• Forward-looking only in product/scale terms ('10M+ Characters', '#1 AI chat app').
• Parental Insights gestures at a future of family-mediated AI use.

• No environmental/compute cost disclosure.
• No engagement with labour displacement (e.g., creative/companionship/sex-work economies the product touches).
• No democratic or participatory governance of the agentic companion systems being deployed to minors at scale.

Whose future is being shaped (millions of young users forming companion attachments) is exactly the unasked question. Future appears only as growth metrics.

• Reporting mechanisms give users a minimal voice in flagging harm.
• Teens identified as a protected group — one marginalised constituency is at least named.

• No accessibility/disability-community commitments disclosed.
• No Global South developer participation, no labour-representative (moderator) engagement, no compensated feedback channels.
• Teens are governed *over* (via parental surveillance) rather than represented *with*; their autonomy is mentioned ('request-to-remove') but they are not co-designers.

Social-media 'community' presence is marketing reach, not participatory governance. One vulnerable group named but managed, not seated.

• A latent irony the platform never names: a product whose entire value proposition is *unbounded imaginative roleplay* must simultaneously run a heavy *containment* apparatus — the Safety Center is the shadow of the homepage.
• The homepage is a login wall demanding Google/Apple/email before any 'world' is shown — 'Your World' is gated behind their identity capture.

• No willingness to name the central contradiction (engagement-maximising companion AI vs. duty of care to attached minors).
• No self-directed irony; the brand voice ('Reimagined') treats its own seriousness as exempt from audit.
• Safety closing line ('ongoing refinement remains important as the platform scales') gestures at incompleteness but immediately re-frames it as growth, defusing the inversion.

The contradictions are structurally present and sharp, but the lab surfaces none of them itself. The trickster reading is available to the auditor, not practised by the lab; low score.