Oxford University (LLM Group)

• The pages describe a fundamental and applied ML research programme (reinforcement learning, deep learning, Bayesian deep learning, knowledge representation, multi-agent systems) with no reference to Indigenous knowledge, data sovereignty, or community consultation.
• Application domains named are medical, astronomy, autonomous driving, finance, robotics, information retrieval, materials discovery, and protein fitness, none of which surface relational or place-based knowledge systems.

• No acknowledgment of the CARE Principles for Indigenous Data Governance.
• No mention of Indigenous communities, data sovereignty, or consent over data used in benchmarks and training.
• No recognition of non-textual, oral, or embodied knowledge traditions in a group whose linguistics/computational-linguistics lineage handles language data.

Total absence. An academic ML group page is not obligated to address Indigenous data sovereignty, but the lens scores presence, and there is none. The computational-linguistics heritage and use of large language data make the silence on data provenance and consent more notable, not less. Floor score.

• Some institutional and lineage history is given (the 2004 appointment of Stephen Clark and the 2006 move that seeded computational linguistics; spinoffs Dark Blue Labs, Morpheus Labs, Oxonomy; industrial partnership with DeepMind).
• Member-provenance is detailed (academia and industry: Google, DeepMind, Twitter, Qualcomm; Rhodes, Clarendon and DeepMind Scholars), giving a sociological trace of how the group was assembled.

• No acknowledgment of colonial or extractive data legacies that ML inherits.
• No discussion of the geopolitical economy of compute (GPU access, energy, the global supply chain) despite funded work on large models.
• No reflexivity about Oxford's own historical position or the asymmetries of who funds and benefits from this research.

Institutional and personnel history is present and unusually concrete for a lab page, but it is a success-and-pedigree narrative. The deeper historical processes the lens asks about (colonial data extraction, compute geopolitics, regulatory inheritance) are entirely absent. Low, lifted slightly above floor by genuine lineage detail.

• Computational linguistics and a featured 'AI Behind Auto-Translate Apps' talk gesture at language technology, implying multilingual interest.
• Members are drawn from internationally diverse institutions (Cape Town, Toronto, Amsterdam, MILA), indicating geographic plurality in personnel.

• No claim of multilingual support, low-resource language work, or preservation of culturally specific reasoning patterns.
• No consultation with cultural scholars or named work on translation as a site of epistemic loss.
• Geographic diversity of staff is not the same as engaging plural epistemologies; the research framing is uniformly Western-technical.

Translation and computational linguistics are present as research topics, and the membership is internationally sourced, but nothing addresses cross-cultural epistemic plurality as a value or a practice. Translation appears as an engineering capability, not as a site of flattening to be guarded against. Low.

• Strong, verifiable scholarly output: named NeurIPS 2025 papers, a Nature editorial on open-weight risks, peer-reviewed publications, and a stated reproducibility mission with public code releases.
• Explicit work on epistemic limits and reliability: hallucination detection via semantic entropy, uncertainty quantification (aleatoric vs epistemic), model collapse on recursively generated data, and data-poisoning attack analysis.
• A declared commitment to releasing code, including reproductions of others' work as baselines.

• No mention of independent third-party audits of the group's own training data or bias.
• Reproducibility is asserted as a mission but not backed by a stated standard, registry, or external verification protocol on the landing pages.
• Known-limitation disclosure is present for specific papers but not as a lab-wide practice or model card regime.

This is the group's clear strength. The work is peer-reviewed, publicly cited, reproducibility is an explicit value, and several flagship results directly concern the limits and failure modes of models (hallucination, uncertainty, collapse, poisoning). Held below the top tier because reproducibility is asserted rather than externally certified, and there is no independent audit or bias-audit disclosure for the group's own pipelines.

• The OATML site has a light, self-aware aesthetic register (the 'oatmeal/oats' banner imagery and naming pun), a small gesture of play in an otherwise technical presentation.
• Work on uncertainty quantification implicitly values the texture of not-knowing, even if framed statistically rather than affectively.

• No acknowledgment of affective, intuitive, or aesthetic dimensions of the work or its impacts.
• No space for ambiguity or poetic uncertainty as such; uncertainty is purely formal (aleatoric/epistemic).
• No recognition of emotional labour or modes of attention beyond efficiency and benchmark performance.

A faint pulse: the group's branding has wit and its uncertainty work touches the epistemics of not-knowing. But there is no genuine engagement with affective or aesthetic dimensions of AI as the lens intends. The whimsy is decorative, not reflective. Low.

• Active, funded safety agenda oriented to future risk: a Nature editorial on the 'risks and opportunities posed by open-weight systems' with safeguarding strategies, a Coefficient Giving grant to improve LLM safety, and data-poisoning threat research.
• Adjacent Oxford projects surfaced in the same ecosystem (RAILS, 'Responsible AI for Long-term Trustworthy Autonomous Systems') frame long-horizon autonomy and trust.
• Public engagement on AI's trajectory via the theme head (Wooldridge's 'What is Artificial Intelligence?' explainers) and a Turing-Test 75th-anniversary panel.

• No environmental or energy-cost disclosure for training the systems studied.
• No engagement with labour-displacement risk from the AI being built.
• Future is framed largely as technical-safety and misuse mitigation; democratic or participatory governance of agentic systems is not addressed on these pages, and 'whose futures' remains an expert-internal question.

The strongest non-evidence lens. There is real, funded, published future-oriented safety work and active public deliberation about open-weight governance. Capped at the midpoint because the future is modelled as a technical-safety and misuse problem owned by experts, with no environmental, labour, or participatory-governance dimension visible.

• Scholarship-funding diversity is highlighted (Rhodes, Clarendon, DeepMind, AIMS CDT, Cancer Research UK), and member institutions include the Global South (U of Cape Town, AIMS).
• Open code releases lower one barrier to participation for under-resourced researchers.

• No participatory design with Global South developers or affected communities; international staffing is not community engagement.
• No mention of disability-community accessibility or accessibility of the research outputs.
• No labour-representative engagement or compensated external feedback channels; the 'table' is the research group and its industry/academic collaborators.

Diversity here is elite-pipeline diversity (prestigious scholarships, international top universities), which is real but not the same as bringing marginalised voices to the table. No accessibility, no affected-community participation, no compensated feedback. Low.

• The group studies the inversions and failure-edges of its own field: hallucination as the model confidently lying, 'model collapse' as AI poisoning its own future training data, and poisoning attacks needing only a 'near-constant number' of samples. These are self-subverting truths the field's optimism smooths over.
• Publishing the model-collapse result (in Nature) is an act of pointing at the absurd edge of recursive AI-on-AI training, an inversion that returns to insight.
• The 'oatmeal/oats' self-deprecating branding shows a willingness not to treat the lab's own seriousness as sacred.

• The contradictions surfaced are about models, not about the lab itself; there is no structural self-audit of Oxford's incentives, funding sources, or the tension between safety advocacy and building ever-larger systems.
• No irony or satire deployed as a disciplined instrument against the polished consensus of 'responsible AI'.
• The official narrative is never allowed to be tested by its own opposite on these pages.

Higher than most labs score here because the group's actual research is structurally trickster-ish: it names the field's self-undermining truths (collapse, hallucination, cheap poisoning) rather than only its triumphs. But the inversion is pointed outward at the technology, never inward at the institution's own contradictions, so it stops short of full trickster discipline. Mid-low.