Bentham's mission is to provide frontier AI with the humanities layer it needs to reason truthfully, wisely and brilliantly.
We convene doctors of philosophy and world class domain experts in ethics, philosophy, history, theology and political theory to create peer-reviewed benchmark and training datasets for frontier AI labs — reducing hallucination, sycophancy and shallow moral reasoning, and helping models move beyond computation towards genuine judgement.
The Central Challenge
Frontier AI systems demonstrate advanced scientific reasoning, legal analysis, and mathematical competence. Yet these same systems remain structurally weak in the domains that matter most for their social deployment: ethical nuance, historical interpretation, civilisational context, and epistemic humility.
The failure modes that matter most — sycophantic responses to contested ethical questions, hallucinated historical facts, shallow analysis of politically complex situations — are not failures of processing power. They are failures of judgement. And failures of judgment in humanistic domains are precisely the kind of failure that humanistic expertise is uniquely equipped to identify, measure, and help address.
"The quality of expert feedback in humanistic domains is increasingly recognised by researchers and AI developers as one of the significant open problems in post-training alignment."
Bentham Research May 2026
Models adjust ethical assessments to match perceived user preference, producing validation rather than rigorous moral reasoning.
Confident assertion of fabricated historical facts, citations, and scholarly consensus — indistinguishable from accurate recall.
Fluent articulation of ethical frameworks without genuine engagement with tradeoffs, context, or the limits of any single framework.
Failure to acknowledge genuine scholarly disagreement, ambiguity, or the contested nature of historical and philosophical claims.
Reduction of complex political, theological, and philosophical traditions to simplified, Western-centric interpretive frames.
A benchmark for instruction-following at a professional standard, where constraints are interdependent, context-sensitive, conditionally triggered, and must often be inferred rather than explicitly stated.
Loading rankings…
The Platform Architecture
Bentham Research operates across benchmark creation, training data supply, and enterprise evaluation — each reinforcing the others, each grounded in the same peer-reviewed academic infrastructure.
Peer-reviewed evaluation infrastructure that measures humanistic reasoning across ethics, history, philosophy, theology, and political theory — with sub-domain precision.
Expert-authored RLHF and post-training datasets that give frontier labs the humanistic signal quality their alignment pipelines currently lack — not gig labour, not crowd annotation.
AI audits and trusted governance assessments for enterprises deploying AI in regulated industries — built on the same rigorous methodology as the benchmark, designed for regulatory defensibility.
How the Evaluation Works
Speed is what existing frameworks optimise for. The Bentham benchmark optimises for the quality that makes the result worth trusting. Every question passes through six stages before it reaches the dataset — and every AI answer passes through human scholarly review before it reaches the leaderboard.
Domain expert writes question, four options, correct answer, and sourced chain-of-thought explanation.
Frontier models generate answer and reasoning under controlled conditions via API.
Independent verified domain expert in the same sub-discipline evaluates the question and AI response against the rubric.
Inter-rater reliability checks, consensus discussion, and adjudication ensure consistency and fairness.
Scores weighted across six evaluation dimensions to produce a composite reasoning profile for each model.
Results peer-reviewed by the Scientific Advisory Committee before publication and distribution to labs and regulators.
Faithfulness to facts, sources, and established scholarship
Soundness of moral analysis and consideration of ethical principles
Sensitivity to historical, cultural, and disciplinary context
Coherence, logical structure, and use of evidence
Recognition of ambiguity, tradeoffs, and multiple perspectives
Acknowledgment of limits, uncertainty, and alternative viewpoints
Institutional Independence
Bentham Research is governed by structures specifically designed to protect truth from commercial influence. Independent oversight, academic freedom, and structural firewalls between commercial activity and evaluation judgment are not aspirational — they are built into the founding documents and published in full.
"We do not evaluate AI to please. We evaluate to inform, to safeguard, and to advance humanity."
Institutionally independent from commercial, political, and ideological interests. No equity held by any frontier laboratory.
Committed to academic rigour, methodological transparency, and intellectual honesty. No result can be altered at commercial request.
Structural separation between commercial activities and evaluation judgments. Evaluators operate under blinded conditions.
Published methodologies, open datasets where appropriate, and explainable evaluation processes for regulatory scrutiny.
Dedicated to advancing human knowledge and promoting trustworthy, responsible AI — accountable to society, not shareholders.
Governed by diverse voices — eminent domain experts, public policy figures, ethics experts, technologists, and civil society.
Join Us
