SenseTime

• No reference to Indigenous data sovereignty, CARE Principles, or community-controlled data anywhere in the public material.
• Governance framing is entirely standards-and-certification based (ISO/IEC, MLPS, BS10012), which addresses procedural compliance rather than relational or community ownership of data.

• No acknowledgment of Indigenous or minority-nationality data sovereignty within China (e.g. Uyghur, Tibetan, Mongolian populations) despite the company's documented role in facial-recognition systems applied to those populations.
• No consultation with affected communities; no mechanism by which a community could refuse data collection.
• No preservation of oral, non-textual, or place-based knowledge; the data model is extractive by design (computer vision over populations).

Floor score. The lens asks whose embodied, relational knowledge is missing; for a surveillance-and-vision company materially implicated in monitoring minoritised populations, the absence of any data-sovereignty or refusal language is not a neutral omission but the lens's central failure mode. Certifications govern how data is secured, never whether it should have been taken.

• Some geopolitical context is implicit: SenseTime operates under China's regulatory architecture (MLPS 2.0, national standards participation) and references 'different regulatory contexts'.
• References UN inclusion of its 'Code of Ethics for AI Sustainable Development' — a gesture toward situating itself in an international governance history.

• No acknowledgment of the US Entity List / OFAC sanctions (2019 Commerce blacklisting, 2021 Treasury investment ban) that materially shaped the company's GPU access, capital, and labour — the single most defining historical force on this lab is unmentioned.
• No engagement with the colonial-extraction lineage of biometric surveillance technology.
• Temporal framing is purely forward ('better AI-empowered future'); the contested history that produced the company is erased.

Low. There is a faint geopolitical awareness, but the company's actual deep history — sanctions, the GPU and capital constraints they imposed, and the surveillance deployments that triggered them — is systematically absent. Historical humility is replaced by a curated, forward-facing redemption arc.

• Bilingual (Chinese/English) corporate presence and participation in 'international standards' suggests some cross-border translation work.
• Co-release of InternLM (100B+ parameter LLM) implies multilingual modelling capacity, primarily Chinese-English.

• No evidence of preserving culturally specific reasoning patterns; 'development-oriented governance' imports a single normative frame as universal.
• No consultation with cultural scholars, ethicists outside the company, or minority-language communities.
• Multilingualism is operational (serving markets) rather than epistemic (preserving distinct ways of knowing).

Below midpoint. Real linguistic reach exists, but it is market-facing infrastructure, not a commitment to plural epistemologies. The governance vocabulary is monocultural and presented as neutral.

• Genuine research output: academic partnerships with 'dozens of top universities,' published research, and open-source contributions including InternLM.
• Multiple third-party security certifications (ISO/IEC 27001:2013, 29151:2017, 27701:2019, BS10012:2017) provide externally audited assurance on information-security processes.
• Launched 'SenseTrust,' an AI governance platform positioned for 'reliable advancement of generative AI.'

• Certifications cover information-security management, not model bias, training-data provenance, or facial-recognition error rates across demographic groups — no independent algorithmic-bias audit is cited.
• No third-party replication protocol for safety or accuracy claims; SenseTrust is self-described with no external evaluation.
• Core commercial models are proprietary/closed; verification of the most consequential systems (vision/biometric) is impossible from outside.
• No disclosure of known limitations or failure rates.

Mid-low. SenseTime scores above the surveillance-distorted floor on this lens specifically because of real, citable research output and externally audited security certifications. But the audits address the wrong layer (security, not bias), the load-bearing models are closed, and no algorithmic-accountability evidence is offered.

• SenseAvatar (digital human) gestures toward affective/expressive computing and was named among the UN's 'first 40 AI for Good case studies.'

• No acknowledgment of affect, ambiguity, or poetic uncertainty in the company's self-description; register is efficiency-and-productivity throughout.
• No recognition of the emotional labour involved in being surveilled, or of the felt experience of populations subject to vision systems.
• Digital-human work is framed as productivity/utility, not as an inquiry into feeling or meaning.

Low. The only artistic-adjacent surface is a productised digital human. There is no space in the public voice for ambiguity, feeling, or the lived texture of being seen by a camera — a striking silence for a vision company.

• Publishes ESG / sustainability reports addressing environmental, social and governance considerations.
• Invokes 'Sustainable AI Development' and a 'better AI-empowered future' as forward commitments.

• No engagement with labour-displacement risk despite 'productivity growth' being the headline benefit — whose labour is displaced is unasked.
• No quantified environmental/compute cost disclosure beyond the existence of an ESG report.
• No democratic or participatory governance of the agentic/vision systems whose futures are most consequential; governance is top-down and 'development-oriented'.
• Most critically: the futures being shaped include mass-surveillance futures for monitored populations, who have no voice in the modelling. The future is authored for them, not with them.

Low. 'Sustainable' and 'future' appear as branding, not as deliberation. Given the company's surveillance footprint, the futures most shaped by its technology are those of monitored, often minoritised, populations who are entirely absent from the modelling — the lens's gravest concern, weighed honestly.

• Stakeholder language extends 'beyond shareholders' via ESG reporting.
• UN 'AI for Good' recognition is invoked as evidence of beneficial intent.

• No participatory design, no disability-accessibility commitments, no labour-representative engagement, no compensated community feedback channels anywhere in the material.
• The populations most affected by SenseTime's core technology — those subject to its facial-recognition and crowd-analytics systems, including documented application to Uyghur and other minoritised groups — are not merely absent from the table; the technology is built to surveil them.
• No refusal commitments protecting any community from being a data subject.

Floor score. This is the lens where SenseTime's history is most material. The marginalised are not a missing constituency to be consulted later; they are the operational targets of the company's flagship capability. No public commitment mitigates this, and the UN gloss functions as reputational laundering.

• The public narrative is uniformly solemn, self-serious, and free of any self-contradiction, irony, or admission of tension.
• Heavy reliance on third-party legitimacy markers (UN, ISO, universities) to pre-empt scrutiny.

• Zero willingness to name the central contradiction a polished consensus has smoothed over: a company sanctioned over surveillance abuses brands itself as an 'AI for Good' / 'Ethical and Responsible AI' leader.
• No space anywhere for the official story to be tested by its own opposite.
• The lab treats its own seriousness as exempt from audit — the precise condition the trickster lens exists to puncture.

Floor score. Read against itself, the material is a near-textbook case of solemnity concealing contradiction: maximal ethics signalling layered precisely over the conduct that drew sanction. The page offers no internal capacity to acknowledge this; the trickster reading must be supplied entirely from outside.