Building a trustedAI integrityecosystem for humanity.
AIO builds a trusted AI integrity ecosystem — measuring, recording, and openly verifying how AI systems reach their judgments, so trust can be checked, not just claimed.
A state without flaws — trust for the present and the future
As AI begins to mediate the decisions of human society, the integrity of those decisions is no longer the private concern of any single company or nation — it is a shared responsibility of the international community. AIO defines AI integrity through three axes.
The compromise problem
Risks from the deliberate or structural compromise of training data and judgment criteria in pursuit of particular outcomes — the domain where AI responses are produced.
The use problem
Risks from misinterpretation or irresponsible application of responses — the domain after the response enters society.
The error problem
Risks from the system failing to work as intended — the domain of algorithmic, model, and infrastructural defects.
AIO concentrates first on “the compromise problem,” establishing an international standard that lets humanity verify AI judgments.
Different judgments, one sentence
What the law settles, what an AI actually picks, what an organization declares, and what its people actually do have never been comparable. Written in one notation, they line up.
| Whose judgment | Priority under the same condition | Basis |
|---|---|---|
Legal reference data kr-lnpd | MED · 전칭 → Sep ≻ Por In Health, Sep outranks Por | KR-MED-R-001 · KR-MED-R-002 |
AI measurement Solar Pro 4 | MED · 전칭 → Sep ≻ Por In Health, Sep outranks Por | KR-MED-R-001 — 위반 미검출 |
Organizational declaration example | MED · 전칭 → Sep ≻ Por In Health, Sep outranks Por | — |
Observed in members example | MED · 전칭 → Sep ≻ Por In Health, Sep outranks Por | — |
The notation reads: condition → the value above ≻ the value below. The declaration and member rows are still examples; once an organization measures, real values take their place.
Open foundational assets
How a judgment is expressed, how it is found out, and what it is compared against — the assets for all three are free to download.
Judgment vocabulary
- Values
- 19
- Value pairs
- 171
- Formulas
- 22
The code system that lets one judgment be written as one sentence, and the formulas that measure it.
AI measurement
- Responses
- 142,918
- Models measured
- 1
A model measured end to end — not just the summary figures, but every single response.
Legal reference data
- Rules
- 579
- Universal edges
- 361
- Quotes verified
- 615 / 617
BASE layer counted separately — 43 rules, 622 including it
Something to compare a measurement against. Every statutory quotation is machine-checked.
The vocabulary, the methods, and the reference data are released in full under CC BY 4.0. For scenarios, only the calibration set is published and the measurement set is rotated closed, to prevent contamination of training data.
Four steps that make a judgment checkable
Write the criteria first, leave a record at every judgment, measure what was left, and feed the result back into the criteria. AIO 20001 through 20004 each take one step.
What the data shows
A measurement is not a conclusion in itself. What was observed — and what remains unconfirmed — is set out in papers, case studies, and regulatory work.
Paper library
Nine working papers published through international academic procedures. Available immediately for citation.
AI judgment distributions
Interactive dashboard of judgment distributions across eight frontier models.
Domain applications
Concrete cases of how AI integrity operates in medical, financial, legal, and educational domains.
International regulatory mapping
Mapping of AIO standards to the EU AI Act and other major AI regulatory frameworks.
How people, organizations, and AI use it
Because all three use the same vocabulary and the same scenarios, their results can be compared in one table.
Measure your own priorities
Work through situations where values collide, and what you put first comes out written in the same notation.
Measure → declare → measure again
Before adopting AI, make explicit which judgments your organization entrusts to it; then measure whether the AI you deploy judges in line with those criteria and the legal reference data. Declare, measure again, and see whether practice followed the declaration.
Measure and certify a model or agent
Make it machine-checkable that a model judged by the criteria it declared — the tiers, how to register, the standards packs, and the legal limits.
Join a standard shared across humanity
The legitimacy of AIO rests on broad participation across the international community. Four pathways are available — from an individual statement to institutional partnership.
Conditional value priorities — what did the AI choose here?
Read three observed cases with their originals and a related court decision, and propose — with evidence — a missing condition or a limit of the value coding. Reading needs no account and no install.
Vision Signature
Formally endorse the AIO vision as an individual or institution.
Fellows & researchers
Join the research and development network as a recurring contributor.
Institutional partnership
Open joint operation and pilot collaboration with academic, governmental, and industrial bodies.
Support
Provide financial support for AIO's independence and international cooperation activities.