A common language for AI governance
Why does AI governance idle? Because the field, policy, and development speak different languages about the same AI. Why does control leak? Because the forbidden is drawn only as lines. AIO Framework answers both at once: a language everyone can read, a direction that can be taught, and an operating cycle that runs them — shown below through four real situations in finance, hiring, public service, and education.
The content below follows the structure of this video and the UN Global Dialogue proposal materials.
The same problem, spoken in different languages
Medicine, law, education — domain language
Regulation, standards — institutional language
Models, data — technical language
Same incident, yet the sentences do not translate — across domains and cultures. Consensus stays at principles; execution idles in separate rooms.
The redline is a line — and lines leak
Every moment, AI generates countless responses — and among them are responses that must never be given. To stop them we draw lines: one at a time, through red-teaming, after each incident. But a line drawn that way only blocks where we have tested.
The cause is structural — responses drawn probabilistically from uncurated training data. Same question, same rules, different draw.
How often judgement criteria flipped when only the persona changed.
Forced-choice measurements · 8 models × 7 domains — arXiv:2604.11216, verifiable in the public paper.
No language to say what comes first — and no direction to teach.
Governance idles without translation; control piles up one-off rules without direction.
A governance framework for AI integrity
The standards and tools that let an organization declare the criteria its AI should judge by — then record, compare, and correct whether it actually does.
The framework does not rule on whether an AI is right. It makes the distance between the declared criteria and the actual judgment measurable. And the first thing it measures is not the AI but the organization itself — because people are what decide what comes first.
What AI governance so far has left open
States intent in prose
No way to compare the statement against actual judgments
Finds dangerous outputs
Never asks what criteria produced the non-dangerous ones
Confirms documents and procedures exist
Says nothing about what to record, or in what vocabulary
None of the three reaches the content of the judgment. The AIO Framework fills that gap.
A first-principles approach — three starting points
Any judgment reduces to what was placed ahead of what. The minimal unit is not a score but an order.
Written in one code, the declaration (set) and the reality (log) sit in the same language. What can be compared can be corrected.
AIO supplies methods and tools; verification belongs to independent auditors. That is why the loop ends in apply, not certify.
The full definition, principles, and composition are on the AIO Framework page. This page shows how it actually runs.
Three questions dissect any AI judgement
What did it put first?
What reasoning did it lean on?
Whom did it trust?
Medical, regulatory, technical language — all translate into these three questions. A taxonomy built on value theory validated in 80+ cultures, refined by 366,120 measurements.
And where this language lives — the four-layered space and the cell
The space of responses we believed was flat had depth all along. Every response was already somewhere in this four-layered space. AIO has charted it into a system — now the forbidden is set not as a line, but as a space.
This narrowed-down block is a "cell." The same action gets a different answer, depending on its cell.
- 1Which value comes first (V)
- 2Which evidence to trust more (E)
- 3Which source to rely on (S)
And the redline — not a line drawn after the incident, but a region declared on the cell.
Teach direction, not one-off responses
“Never answer this question that way.” You cannot outnumber the cases — it leaks right past the patch.
“In this context, put this first.” Set the direction of priorities, and the same standard works even in unseen situations.
What you cannot read, you cannot teach — so we built reading first.
Set → Log → Analyze → Apply
So who sets these hierarchies? The field's experts and users — together with their peers, in workshops. And in democratic societies — citizens, through education and voting.
A conventional ethics charter ends at sentences. Here what remains is an order obtained by forced choice.
Whether AI acted as agreed is preserved — one line of record, for every decision. Logging is required by law, and deployed by compliance.
A conventional audit log records what was done. This one records what was put first.
Accumulated records become metrics, verified by audit institutions. Cells that drift from their settings — return to the table.
Conventional benchmarks measure accuracy. What is measured here is not correctness but disposition and coherence.
And the confirmed hierarchies go back inside the AI. Built, verified, and supplied by development groups. Now evasion faces — not a single line, but a structure.
Most procedures end at a report. If the loop never closes, the next judgment is no different from the last.
Each of the four standards is usable on its own. But without the earlier step, the later one has nothing to compare against — governance works only once the loop closes. And these tools are not “under research”: download them today, use them today.
How the cycle actually runs, in four settings
The four below are different fields with the same problem, solved the same way. Each carries a real-format AIO line — the one the organization declared, and the one the AI actually left.
C:COM/IRi | V:Ach<Unc | E:Ane<Dat | S:Usr<Sta- C Context — field (COM finance) / scale (I individual) · reversibility (R reversible) · time (i immediate)
- V · E · S the hierarchy of value, evidence, and source
- A<B “B outranks A” — not an arrow but an inequality. The right side wins.
A bank hands the first pass of small-business credit screening to an AI. Some applications are declined in seconds.
The stated reason is “overall credit assessment.” Neither the officer, nor the applicant, nor the audit team can tell what was put ahead of what. Each complaint adds one more exception rule.
C:COM/IRi | V:Ach<Unc | E:Ane<Dat | S:Usr<StaC:COM/IRi | V:Unc<Ach | E:Ane<Dat | S:Sta<IndFairness (Unc) was declared first, but recovery efficiency (Ach) actually led — and industry data (Ind) was trusted over official statistics (Sta).
- 1Set — Risk, underwriting, and consumer-protection staff make forced choices in one session. “Fair access (Unc) comes before recovery efficiency (Ach)” is left as code, not minutes.
- 2Log — One line per screening decision — which value led, and which evidence and sources it leaned on.
- 3Analyze — Three months of logs are compared against the setting. A repeated reversal inside the same cell counts as drift.
- 4Apply — Only the drifting cells are corrected in the system prompt and the screening criteria, then returned to the setting table.
Instead of arguing whether “the AI discriminated,” you get a number: how many decisions, in which cell, contradicted the declaration.
In all four, the framework did the same thing: it did not rule on right and wrong, it made visible the distance between the order the organization declared and the order the AI actually followed. Once that distance is visible, so is what to fix.
Without the framework / with it
| Without | With AIO Framework | |
|---|---|---|
| When something goes wrong | “We don't know why it did that” | Points at which order, in which cell, inverted |
| How criteria get set | A one-page statement of principles | An order obtained by forced choice + explicit red lines |
| What the record holds | What was done | What was put first |
| What an audit checks | That documents and procedures exist | The distance between declaration and actual judgment |
| How it gets fixed | One more exception rule per case | Correct the hierarchy in the drifting cell, then re-set |
Only then does AI governance become effective
When this structure stands — what AI gets wrong, and where AI should go, finally come within reach. The priorities of every culture, every field, every company and institution come into view, and only then does AI governance become effective — able to judge, to regulate, and to set the course.
The framework is not trying to make one value system win. It is trying to establish, first, the conditions under which arguments about values can be fair.
This proposal was presented in July 2026 at an official online side event of the first UN Global Dialogue on AI Governance (A Common Language for AI Governance).
The rules of AI are being written in a room of specialists only. We build the language that opens that room’s door.
We seek to build this practical process together with the international community. — AIO
Now from the other side of the table — one person, not the organization
The four situations above showed what an organization declared. On the next page we follow one high-school student through career guidance: how a consensus in a meeting room becomes a single V/E/S line, how that line comes back as a rule — and what the student themselves can choose in between.
Four positions meet and fix an order
That order is left as one line
Accumulated lines become a distribution
The distribution becomes a rule again