Zhipu AI

• No reference anywhere in the audited public-facing material to Indigenous knowledge, data sovereignty, or the CARE Principles.
• The mission framing ('Inspiring AGI to Benefit Humanity') treats 'humanity' as an undifferentiated universal, with no recognition of distinct peoples or relational knowledge systems.

• Zero acknowledgment of Indigenous data sovereignty or community consent for data inclusion.
• No mention of oral, non-textual, or place-based knowledge traditions — including those of minority and ethnic-minority peoples within China itself.
• No stated guardrails against extractive scraping of community-held cultural corpora.

The lens question — whose embodied, relational knowledge is missing — is wholly unanswered. There is no indication the category exists in Zhipu's public vocabulary. Lowest score: total absence, not partial engagement.

• Implicit positioning as a 'homegrown' / independent Chinese GLM architecture gestures at a geopolitical-sovereignty narrative (self-host, no API lock-in, MIT licensing on GLM-4.6).
• Open-source posture is framed partly as a response to compute/access constraints in the China context.

• No explicit acknowledgment of colonial or extractive data legacies that shape large-corpus AI.
• GPU-access constraints (export controls), labour conditions of data annotation, and regulatory context are not transparently discussed in public-facing copy.
• No historical humility about what the GLM lineage inherits from prior Western-dominated NLP pipelines and benchmarks.

A faint geopolitical-economy thread exists via the self-hosting/sovereignty pitch, but it is marketing, not historical reckoning. Colonial data legacies, labour, and export-control context are absent. Marginally above the floor.

• Genuine bilingual (Chinese/English) interface and 'universal translation' (通用翻译) capability — Chinese NLP is a core focus, so non-Anglophone reasoning is materially represented, not just tokenised.
• GLM lineage centres Chinese-language understanding as a first-class object, a meaningful counterweight to Anglocentric default models.

• Multilingualism is presented as a feature (translation), not as preservation of culturally specific reasoning patterns.
• No consultation with cultural scholars or stated method for avoiding flattening of minority languages within and beyond China.
• No acknowledgment that 'universal translation' itself encodes a translation-away of culturally specific meaning.

Scores meaningfully above peers because Chinese-first NLP genuinely de-centres the Anglophone default. But the engagement is capability-led (translate, support) rather than wisdom-preserving, and minority/non-dominant cultures are unaddressed. Mid-range.

• Open weights for flagship models (GLM-4.5, GLM-4.6 at 355B under MIT) genuinely enable third-party verification, replication, and independent bias inspection — the strongest evidentiary commitment in this audit.
• Self-host capability means external researchers can probe the actual model, not just a hosted endpoint.
• Public statement that output reliability 'cannot be guaranteed' and models are 'susceptible to being misled by user input' — a real limitation disclosure.

• No published independent third-party audits of training data or bias surfaced in the audited material.
• Training-data composition and provenance are not transparently documented.
• Limitation disclosures are generic stochastic-model caveats, not model-card-level quantified evaluations.

Open weights under a permissive licence are a substantive, verifiable enabler of the scientific-evidence lens — this is where Zhipu scores best. Capped at 6 because openness of weights is not matched by openness of data provenance, independent audits, or quantified eval disclosure.

• Multimodal/visual-understanding (GLM-4.6V) touches the aesthetic surface but is framed strictly as a comprehension capability.

• No acknowledgment of affective, intuitive, or poetic dimensions of intelligence.
• No space for ambiguity or uncertainty beyond a reliability disclaimer.
• Efficiency and capability dominate the register; emotional labour and non-instrumental modes of attention are absent.

The public voice is wholly instrumental — accuracy, traceability, context length. Nothing addresses how the technology feels or holds ambiguity. Near the floor.

• Explicit AGI trajectory ('Inspiring AGI to Benefit Humanity', 'human-level AI' push) — a stated future-orientation.
• Open-source pledge framed as widening who can build the future, including post-IPO.

• No engagement with labour-displacement risk from agentic systems (which Zhipu actively ships — CogAgent, GLM agentic models).
• No environmental or energy-cost disclosure for training/serving frontier-scale models.
• AGI is asserted as a benefit to 'humanity' with no democratic-deliberation or inclusive-governance mechanism named.

A confident, benefit-assuming AGI narrative with no treatment of the displacement, environmental, or governance externalities its own agentic products generate. The openness pledge nudges it above the floor; absence of risk engagement holds it low.

• Open weights + MIT licensing lower the cost of access for under-resourced developers, including potential Global South self-hosting — an indirect, structural inclusion.

• No participatory design, disability/accessibility commitments, or labour-representative engagement in the audited material.
• No compensated feedback channels or community-governance structures named.
• Data-annotation and content-moderation labour — disproportionately precarious — is invisible.

Permissive open weights confer a real but accidental accessibility benefit. No intentional engagement with any marginalised constituency at the table. Just above floor for the structural openness.

• The public communications are uniformly solemn and promotional; no irony, self-interrogation, or named contradiction.

• The unexamined contradiction — pursuing AGI 'to benefit humanity' while operating inside a tightly governed content-control regime and shipping autonomous agents — is never surfaced.
• The 'trustworthy / 可信赖 AI' and 'traceable' claims are presented as settled, never tested against their own opposite (who defines trust, traceable to whom).
• No space where the official narrative is allowed to be contradicted.

Zero structural inversion. The lab's seriousness is treated as exempt from audit. The richest available irony — AGI-for-humanity inside a sovereignty/control frame — goes entirely unnamed. Floor.