Inflection AI

• No reference to Indigenous knowledge, data sovereignty, or community consultation anywhere in the public-facing or policy materials reviewed.

• No acknowledgment of CARE Principles or Indigenous data sovereignty.
• No consultation processes with Indigenous communities for a product ('Pi') built to occupy intimate, relational conversational space.
• No protection for oral, relational, or non-textual knowledge against extractive training-data ingestion.
• The 'royalty-free, perpetual, irrevocable' training license is the precise mechanism by which relational knowledge becomes extractable without provenance or consent at the community level.

The product's whole premise is relational intimacy, which makes the total absence of any relational-knowledge sovereignty framing more conspicuous, not less. A perpetual irrevocable training license over the most personal category of conversational data, with zero Indigenous or community-sovereignty consideration, scores a floor 1.

• Acceptable Use Policy language mirrors EU AI Act prohibited-practice categories (social scoring, biometric categorisation, untargeted facial-recognition scraping, workplace emotion inference), implying awareness of a specific regulatory lineage.

• No acknowledgment of colonial or extractive data legacies underpinning large-model training.
• No transparency about GPU access, compute supply chains, or the data-labour economy.
• No historical context for the company's own turbulent history (the 2024 Microsoft talent-and-licensing arrangement that hollowed out the original founding team) and what that means for stewardship continuity.
• Regulatory mirroring is presented as a static rulebook, not as a contested, historically produced settlement.

Some implicit historical inheritance is visible in the regulatory-shaped prohibitions, but it is borrowed compliance scaffolding rather than reflexive historical humility. No engagement with the material or institutional history that shaped the lab. Scores 2.

• No multilingual commitments, no cultural-scholar consultation, no recognition of culturally specific reasoning patterns in the reviewed materials.

• 'Pi' is marketed as a universal personal/empathetic AI, yet there is no statement of language coverage or cultural adaptation.
• 'Emotional intelligence' is treated as a universal, culture-neutral category, when affect norms are deeply culturally variable.
• No acknowledgment that empathy, comfort, and 'relational' interaction are translated through a default Western, English-language affective register.
• No consultation with cultural or linguistic scholars disclosed.

A product whose core claim is emotional/empathetic intelligence universalises one affective register while disclosing nothing about cross-cultural validity or language plurality. The flattening is structural to the value proposition. Scores 1.

• Outputs are explicitly flagged as potentially unreliable and not a substitute for professional advice, a genuine if minimal limitation disclosure.
• Content moderation described as 'a combination of automated technologies and human review.'

• Closed weights ('proprietary API') with no third-party replication or external audit pathway.
• No independent audit of training data or bias disclosed.
• No empirical evidence offered for the central 'emotionally intelligent' / 'empathetic' claim, which is asserted rather than measured.
• No model cards, evaluation benchmarks, or known-failure taxonomies in the reviewed materials.

A boilerplate accuracy disclaimer is the only evidentiary commitment. The headline empathy claim carries no verification, and the closed-weight posture forecloses external replication. Scores 2.

• Unusually for an AI lab, affect is the explicit core of the offering: 'emotionally intelligent,' 'relational,' 'human-centered.'
• The relational/transactional contrast gestures at modes of attention beyond pure efficiency.

• Emotion is instrumentalised toward 'enduring value for individuals and enterprises' rather than honoured as ambiguous or non-extractable.
• No space for poetic uncertainty, grief, or unresolvable ambiguity; affect is rendered as a smooth, monetisable surface.
• No acknowledgment of the emotional labour involved on the user side, nor of the asymmetry of confiding in a system that holds a perpetual training license.
• The aesthetic is reassurance, not honest affective complexity.

This lens scores comparatively higher because affect is genuinely centred rather than absent, the highest score in this audit. But the affective register is managed, smoothed, and routed to commercial value, so it stops short of honest artistic perception. Scores 4.

• AUP engages near-term harm vectors (surveillance, manipulation, discrimination, social scoring) with concrete prohibitions.

• No environmental or compute-cost disclosure.
• No engagement with labour displacement, despite an explicit enterprise/API go-to-market that automates interaction work.
• No democratic or participatory governance of the agentic/empathetic system; futures are shaped unilaterally by the firm.
• Whose future the 'enduring value for individuals and enterprises' serves is asserted, not deliberated.

Harm-prevention prohibitions are present but defensive and compliance-shaped. There is no forward, inclusive, or deliberative modelling of the futures an empathetic-AI-at-enterprise-scale produces. Scores 2.

• AUP protects vulnerable groups from manipulation and bans exploitation of vulnerabilities and discriminatory uses, a protective gesture toward marginalised users.

• Protection is framed as restrictions on misuse, not as participation: no participatory design, no Global South developer engagement, no disability-community accessibility commitments disclosed.
• No compensated feedback channels or labour-representative engagement.
• Users of an intimate empathetic AI in vulnerable states are objects of policy, never co-authors of governance.
• No accessibility statement for Pi despite a multi-device, conversational product.

Vulnerability appears only as a class to be shielded from above, never as a constituency at the table. The one concrete user-agency mechanism ('!delete') is exit, not voice. Scores 2.

• No irony, self-questioning, or willingness to name internal contradiction in any reviewed material; register is uniformly earnest and reassuring.

• The central contradiction goes unnamed: a product selling intimate, 'relational,' empathetic trust simultaneously claims a 'royalty-free, perpetual, irrevocable' license over the contents of that intimacy.
• No acknowledgment that 'transforms interactions from transactional to relational' is asserted by an entity whose own license terms are maximally transactional.
• The lab's seriousness is treated as exempt from audit; no space where the official empathy narrative is tested against its own commercial opposite.
• Inversion that would clarify (an empathetic AI that profits from confided emotion) is precisely what the polished surface conceals.

The materials are wholly solemn and self-consistent on the surface while sitting on a sharp unacknowledged paradox between marketed relationality and extractive licensing. Zero structural inversion is permitted. Scores 1.