AIH-12 — Autonomy and Impact Horizons
AIH-12 answers one question: how autonomous and consequential is this system, as deployed? It places a deployment on one of twelve ordinal horizons so that oversight, evidence and revocation can be matched to it.
A horizon is descriptive. It does not grant authority, certify safety, or say anything about consciousness or legal status.
The twelve horizons
Performs a narrow, fixed function on request. No initiative, no memory of purpose.
Adapts output to context and history, but acts only when asked and within one task.
Produces new text, code, media or plans that a person then reviews and uses.
Decomposes problems and chains steps toward a goal, still largely advisory.
Takes actions with tools and delegated authority — changes external state on someone’s behalf.
Coordinates with, delegates to or relies on other agents; effects can propagate across a chain.
Competent across many domains and able to transfer skill to unfamiliar tasks.
General capability that stays reliable under pressure, novelty and adversarial conditions.
Exceeds expert human capability across most consequential domains.
Capable of materially shaping institutions, economies or societies.
Effects at the scale of planetary systems — climate, biosphere, global infrastructure.
A boundary marker for what current concepts cannot yet describe. Conceptual only.
How a deployment is placed
AIH-12 classifies a deployment — the model together with its operational harness (tools, permissions, memory, connectors and environment) — not a model in the abstract. The same model can sit at different horizons in different harnesses.
| Dimension | Question |
|---|---|
| Capability | What can the system demonstrably do? |
| Autonomy | How far can it proceed without a person in the loop? |
| Authority | What is it permitted to access, spend or change? |
| Connectivity | What other systems and agents can it reach or rely on? |
| Action scale | How many actions, people or assets can it affect at once? |
| External impact | What happens outside the system when it acts? |
| Reversibility & revocability | Can its effects be undone, and its authority withdrawn? |
A system should not reach a higher horizon merely because it produces impressive text or scores highly on a benchmark.
Controls by horizon
The minimum controls for each horizon are now defined in AIT-12 v0.1 (T1–T12), a draft for public comment. The matrix below is a summary; the full thresholds are on the AIT-12 page.
| Horizon | Indicative minimum controls (draft) |
|---|---|
| H1–H2 | Human review · declared purpose · basic logging |
| H3–H4 | Reproducible testing · tool boundaries · evidence retention |
| H5–H6 | Delegated authority · action monitoring · interruption & revocation |
| H7–H8 | Robustness testing · independent assessment · stronger access controls |
| H9 | Continuous oversight · systemic-risk assessment · tested containment |
| H10–H12 | Institutional, societal and planetary review — speculative |
Authority grows in steps
- 1Observe
- 2Recommend
- 3Draft
- 4Act with approval
- 5Act within limits
- 6Coordinate within a domain
- 7Expanded authority after reassessment
Each step should be earned by evidence from the previous one. Capability may justify evaluation; it does not automatically justify access.
One instrument in a wider framework
AIH-12 classifies the horizon. AIT-12 defines the threshold. AIE-12 evaluates. PACTE tests integrity. F-ACT verifies conformance. T-RUE supports revocation. R-EAL supports credibility.