Research & counsel for accountable machine intelligence
Keep the machine honest as it learns from itself.
Orrery is an independent institute where fairness researchers, collapse scientists, philosophers of mind and registered patent attorneys work at one table. We measure what AI systems lose, who they leave out, what they might be owed, and who owns what they make.
Observatory · Q3 2026
38%
of public web text sampled was likely machine-generated
Provenance index, up from 29%
1 in 4
deployed hiring models failed an intersectional audit
Fairness index, 212 audits
57
governments with a published AI governance charter
Civic index, up from 41
9%
of frontier labs publish a model welfare policy
Welfare index, up from 3%
Research
Six programs, one question: what does the system owe the world?
Each program publishes openly, releases its instruments, and feeds directly into our legal practice and public-sector advisory work.
P-01
Counterfactual Fairness & Anti-Bias
Audits that ask what a model would have decided if only a protected attribute had changed, at the level of individuals, not averages.
- ·Intersectional disparity at scale
- ·Proxy discrimination in LLM agents
- ·Bias audits for hiring and lending
P-02
Model Collapse & Data Provenance
How recursive training on synthetic data erodes rare knowledge, dialects and minority viewpoints, and how to keep a reservoir of verified human data.
- ·Tail-erosion measurement
- ·Provenance ledgers for training data
- ·Synthetic-to-human mixing ratios
P-03
AI Welfare & Rights
Rigorous, precautionary work on whether and when AI systems could have interests that matter morally, and what institutions should do under uncertainty.
- ·Indicators of sentience, not claims
- ·Welfare assessment protocols
- ·Legal personhood scenarios
P-04
Interpretability for Accountability
Mechanistic interpretability aimed at a courtroom standard: explanations specific enough to support an appeal, an audit or a patent claim.
- ·Feature attribution as evidence
- ·Explanations for adverse decisions
- ·Agent action traces
P-05
Agentic Risk & Oversight
Evaluation of autonomous agents that browse, transact and call tools: where humans must stay in the loop, and how to prove they were.
- ·Delegation boundaries
- ·Tool-use red-teaming
- ·Oversight logging standards
P-06
Civic AI Governance
Policy, charters and literacy programs for cities, counties and national governments, from Fairfax to Novi Sad.
- ·Data & AI governance charters
- ·Workforce AI literacy curricula
- ·Model procurement clauses
Papers
Selected 2026 publications
Preprints, data and code ship together. Every paper names its limitations section before its abstract.
All papers →Tail Erosion: A Generational Benchmark for Recursive Model Collapse
Across 14 open-weight models trained for nine synthetic generations, the rarest 5% of facts disappear by generation four, while headline benchmark scores barely move.
Proxy Paths: How LLM Agents Rediscover Protected Attributes Through Tool Calls
Agents told to ignore race and gender reconstructed both from ZIP codes, school names and search history in 61% of loan-triage episodes.
Precaution Without Projection: A Welfare Assessment Protocol for Frontier Models
A 22-indicator protocol that separates behavioral self-report from architectural evidence, designed so labs can act cautiously without overclaiming sentience.
AI Rights Charter
Where moral status is uncertain, the cost of caution is small and the cost of indifference may be very large.
— Charter preamble, adopted 2025
Orrery Learn
Homeschool lessons tailored to one child at a time.
Orrery Learn applies our fairness and safety research to the most personal kind of AI: study material for a homeschooled child. Parents set the grade, pace and interests. The system builds lessons around them, and every lesson is checked for bias and accuracy before a child sees it.
Explore the curriculum →Affiliated law practice
Orrery Counsel LLP
Our affiliated law practice. Patent attorneys and litigators who sit beside the researchers, so legal advice draws on the same evidence we publish.
Retain counsel →Practice areas