Mila (Quebec AI Institute)

• Mila runs a dedicated, recurring Indigenous Pathfinders in AI program (three cohorts to 2026) for First Nations, Inuit and Metis talent, explicitly framed to 'bridge Indigenous perspectives with artificial intelligence technologies'.
• Pedagogy is grounded in a specific Indigenous epistemology rather than a generic gesture: 'rooted in the Nehinuw concept of Teaching Each Other (kiskinaumatowin), introduced by Keith and Linda Goulet, which views learning as a reciprocal process.'
• Community-driven, language-revitalisation projects are surfaced and credited (G(AI)M for Mohawk language, Buffalo in Motion, SAIGE scholarship matching), indicating data and problem framings owned by communities.
• Material barriers are addressed concretely: a $5,800 stipend plus travel and accommodation support, reducing extractive 'free labour' dynamics.

• No explicit reference to Indigenous data sovereignty frameworks (OCAP, CARE Principles) or who holds rights to data/models produced by cohort projects.
• Knowledge flows are described as reciprocal but there is no governance statement on benefit-sharing, IP, or how community knowledge feeding AI systems is protected from downstream extraction.
• Program is talent-pipeline shaped (career pathway into the AI ecosystem); less evidence that Mila's own core research data practices are restructured by Indigenous governance.

Substantially above sector norm: a sustained, named, community-rooted program with a specific Indigenous epistemology and paid participation. Held below 9-10 because data sovereignty (CARE/OCAP), IP and benefit-sharing governance are never named, and the framing centres pipeline access more than restructuring of Mila's own data practices.

• Some institutional history is given (founded 1993 by Yoshua Bengio; university affiliations), situating the lab in a Quebec academic lineage.
• Governance work acknowledges power asymmetries in the present: policymakers have 'limited access to technical AI expertise' and the information space is 'full of competing claims and vested interests'.

• No acknowledgement of colonial data-extraction legacies, despite running an Indigenous program on territory whose history is unaddressed (no land acknowledgement in audited text).
• No engagement with the geopolitical economy of AI: GPU/compute access, data-labour supply chains, or environmental inheritance of the field.
• Linear 'transformative technology' framing presents AI as an arriving force to be managed rather than a contingent product of specific historical and economic choices.

Institutional history is present but historical humility about AI's colonial and material inheritances is essentially absent. The Indigenous program makes the silence on colonial data legacy and the lack of a land acknowledgement more conspicuous, not less.

• Genuine institutional bilingualism: every audited page exists in parallel French and English, reflecting a Quebec francophone context rather than English-default tokenism.
• Indigenous-language revitalisation is treated as a first-class AI application (Mohawk via G(AI)M), preserving a culturally specific knowledge system.
• International, multi-jurisdiction governance partnerships (UNESCO, UN-Habitat, GPAI, G7) bring non-single-culture framings into the policy work.
• A dedicated GPAI project on 'Real Diversity and Gender Equality in AI' signals attention to perspective plurality across the AI life cycle.

• Multilingualism is structurally Franco-English; languages of the Global South or non-Western reasoning traditions appear only via individual projects, not as institutional commitment.
• Responsible-AI framing leans on Western categorical vocabulary (fairness, bias, trustworthiness, alignment) presented as universal, without naming culturally specific reasoning patterns it may flatten.
• No evidence of cultural scholars consulted on the epistemology of the AI methods themselves (vs. on application domains).

Real, structurally embedded bilingualism plus Indigenous-language work lifts this above token presence, but the universalising fairness/alignment vocabulary and Franco-English centre of gravity cap it mid-range.

• Strong evidentiary culture: dense, dated, peer-reviewed publication stream with DOIs and arXiv links (Science, European Journal of Radiology, AI Chemistry), open to verification.
• Explicit naming of model limitations as a research target: 'tackling the lack of epistemic humility within large language models, which can confidently state false things.'
• The flagship risk paper ('Managing extreme AI risks amid rapid progress', Science 2023) openly states 'there is a lack of consensus' and that 'AI safety research is lagging' rather than projecting certainty.
• Methodological transparency in showcased work (cross-institution train/test splits, named statistical tests, reported precision/recall deltas).

• Openness is publication-level, not artefact-level: no statement on open weights, open training data, or independent third-party audits of Mila's own released models.
• Bias/fairness work is described programmatically but without independent external audit protocols or replication commitments for Mila outputs.
• Stated 'open' openness posture is not evidenced by any weights/data release policy in the audited text.

An academic institute with a verifiable, limitation-aware publication record scores well. Capped below 8-9 because verification is via the peer-reviewed literature rather than open weights, open data, or independent audits of Mila's own artefacts.

• Affective, relational framing in the Indigenous program ('holistic and relational learning approach'; reciprocal knowledge flow) admits non-instrumental modes of attention.
• Indigenous art is foregrounded with attribution (Lakota artist Kim Soo Goodtrack's 'Tatanka Leads the Way with Honour and Respect'), treating aesthetic work as meaningful rather than decorative.
• An applied chair on 'ethical use of AI in game development' (Ubisoft-Mila) gestures toward creative-industry sensibility.

• Dominant register is technical-managerial (mitigation, governance, safeguards, frameworks); little space for ambiguity, poetic uncertainty, or the felt texture of living with AI.
• Emotional labour (e.g. of content-moderation, trauma-adjacent work on trafficking/misogyny projects) is named as a problem domain but not as something the workers feel.
• Affect appears mainly in the Indigenous-program subsite, not in Mila's core self-description.

Pockets of affective and aesthetic recognition exist, almost entirely within the Indigenous program. The institutional voice elsewhere is efficiency- and governance-coded, so the score sits below the midpoint.

• Direct engagement with large-scale and existential future risk: the Science consensus paper warns of 'large-scale social harms, malicious uses, and an irreversible loss of human control over autonomous AI systems.'
• Active shaping of democratic AI governance futures (UN Scientific Panel design recommendations 'to ensure the scientific independence, legitimacy and policy relevance'; AI Insights for Policymakers; G7 SME adoption report).
• Climate is a named strategic focus area with applied research (e.g. green-material discovery; David Rolnick's applied climate work), tying futures to environmental stakes.
• Disinformation futures explicitly programmed (June 2026 'Disinformation 2.0: When AI Blurs the Lines' public event).

• No disclosure of Mila's own environmental/compute footprint, even while climate is a stated priority and 'extreme risk' is invoked.
• Labour displacement from AI is largely absent as a risk theme; framing favours adoption ('AI adoption among SMEs') over displacement consequences.
• Governance of agentic systems is discussed at policy level but inclusive deliberation with affected publics (vs. policymakers and experts) is thin.

Serious, materially-backed engagement with catastrophic risk and governance futures lifts this above average, but the missing own-footprint disclosure and near-silence on labour displacement keep it from the upper band.

• Compensated, low-barrier participation for an under-represented group (Indigenous Pathfinders: stipend + travel/accommodation), a concrete anti-extractive channel.
• Applied projects directly serve marginalised stakeholders: countering online human trafficking, modern slavery, and 'sexist bias in text' / misogyny (NeurIPS 2023 project).
• Explicit gender-equality and diversity policy work (GPAI DGE project) and a Human Rights conference convening 'civil society and policy-making communities'.
• Faculty roster shows substantial international and gender diversity among named researchers.

• No mention of Global South developers as participatory design partners, nor of disability-community accessibility in Mila's tools or processes.
• No labour-representative or worker-voice engagement (e.g. data-labour, content-moderation workers) despite trafficking/abuse problem domains.
• Marginalised groups largely appear as beneficiaries of, or subjects of, projects rather than as governance decision-makers over Mila.

Concrete, compensated inclusion and rights-oriented applied work are real strengths; held mid-range by the absence of disability access, Global South co-design, labour voice, and any structural role for marginalised groups in governing Mila itself.

• Willingness to name an uncomfortable contradiction in its own field: that LLMs 'confidently state false things' and that 'AI safety research is lagging' even as the field accelerates.
• A scheduled public event ('Disinformation 2.0: When AI Blurs the Lines') invites scrutiny of AI's own distorting effects, a mild reflexive gesture.
• The institute platforms critics (Kate Crawford on 'the urgent need to regulate AI'; Yuval Noah Harari on AI and the future of civilization), letting outside voices test the narrative.

• No structural inversion of Mila's own seriousness: the institute never turns its critique on its own incentives (industry partners, venture arm, compute dependence).
• Tension between 'AI for the benefit of all' / 'AI for humanity' branding and a venture arm + corporate-partner network ('Mila Ventures', Toboggan/Ubisoft/ServiceNow) is left entirely unexamined.
• No irony, satire or self-mockery as disciplined instrument; the register is uniformly earnest and solemn.

Mila will name contradictions in the field and host external critics, which earns more than a floor score. But it never applies that inversion to its own commercial entanglements (venture arm, industry network) or its 'benefit of all' framing, so the trickster function stops at the edge of self-audit.