The 15 levels of artificial intelligence maturity. From "a machine that answers questions" all the way to "what kind of society do we want to build?": each level changes the question you ask, and each level has its blind spot.
The goal is not only to measure what a person knows technically. You have to measure a full spectrum of understanding: what the system does, how it acts, by what right, and in what world.
This site is a journey of understanding, not a normative framework: the staircase orders questions, not people. The axes remain independent; the progression tells how the gaze widens.
A two-position switch. It is the most widespread representation, and the most wrong.
Six independent axes. You can be an expert on one and blind on the other five.
Each era does not replace the previous one: it changes the question you ask the system.
"What can AI do?"
"How can it act autonomously?"
"Does it have the right to do it, and how do you prove it?"
"What kind of society do we want to build?"
"What can AI do?"
"A machine that answers questions."
The person uses AI as a chatbot: asking questions, generating text, getting ideas.
"A model I can steer with good instructions."
Context, constraints, examples, roles, output formats, iterations: the person learns to drive the model's output instead of enduring it.
"A probabilistic system trained on large amounts of data."
Tokens, context window, inference, embeddings, temperature, hallucinations: the person understands the mechanics that produce the answer, and why it can fail.
"A component of a work process."
Humans, models, databases, code, automations, validations: the person thinks in integrated processes rather than isolated prompts.
"How can it act autonomously?"
"A system able to reason while consulting sources and using tools."
Search engines, vector databases, APIs, software: the model steps out of its internal memory.
The fundamental shift: the difference between the model's internal knowledge and information retrieved dynamically from the environment.
"A system able to pursue a goal and chain actions."
Goal, plan, tools, action, observation, iteration, memory: the loop replaces the question-answer exchange.
The fundamental shift: moving from "answering" to "acting".
"Several specialized intelligences can collaborate, critique each other, delegate work or compete."
Orchestration, delegation, critique, verification, arbitration, competition: thinking in terms of a society of agents sharing a memory and coordinating indirectly.
"Does it have the right to do it, and how do you prove it?"
"A distributed cognitive infrastructure made of several distinct systems."
Data, model, runtime, memory, identity, policies, orchestration, authorization, proof: distinct layers, replaceable one by one.
A robust AI system must be able to survive the replacement of a model or a provider.
"An intelligence that acts within a space of rights, permissions and responsibilities."
Who is this agent, what do we trust it to do, what is it able to do, and what does it have the right to do: four distinct questions, four distinct answers.
The fundamental equation: Identity ≠ Trust ≠ Competence ≠ Authority.
"An intelligence whose actions and claims must be verifiable."
Declaration, artifact, event, trace, result, verdict, proof: an independent verification chain, link by link.
We no longer ask "does it work?" but "what allows us to prove it?" (provenance, causality, reproducibility).
"An intelligence must not only produce correctly executed actions; it must produce a useful effect in the real world."
Two independent axes: the quality of technical execution, and the real usefulness of the result.
A perfectly executed system can perfectly produce something useless.
"What kind of society do we want to build?"
"A machine that produces propositions under uncertainty."
The epistemic spectrum: absolute ignorance, uncertainty and contradiction, hypothesis and estimation, inference, observation and fact.
Intelligence is not only about giving answers. A mature intelligence must be able to represent what it does not know.
"An intelligence must be subject to an authority above it."
An authority layer above the system: constitution, permissions, enforceable refusals.
The fundamental principle: capability does not grant authority. The legitimate conditions of AI action must be built.
"A population of intelligences evolving in a human, economic, political, technical and institutional environment."
Agents, infrastructures, humans, markets, institutions, protocols, power relations: everything interacts with everything.
The unit of analysis becomes the whole system. AI is a complex-systems problem where behaviors emerge from interactions.
"A new scientific and civilizational object we only partly understand."
Interpretability, emergence, artificial cognition, alignment: living research.
Our current words (model, agent, reasoning, autonomy) are only provisional approximations for barely understood systems.
Four eras, four units of analysis, four questions, four success metrics.
| Era 1 (The Tool) | Era 2 (The Agent) | Era 3 (Governance) | Era 4 (The System) | |
|---|---|---|---|---|
| Levels | 1 to 4 | 5 to 7 | 8 to 11 | 12 to 15 |
| Unit of analysis | The Model | The Action | The Architecture | The Ecosystem |
| Main question | "What does it do?" | "How does it act?" | "Does it have the right?" | "What society do we want?" |
| Success metric | Precision and usefulness | Autonomy and goals | Safety, proof and value | Legitimacy and emergence |
Check every statement that is honestly true for you (or for your organization). The result is an estimate of the highest step of your journey, with the blind spot to watch.
Level 15 · Frontier: this stage is not self-attributable. It names the questions whose answers remain open.
"Artificial intelligence has ceased to be a model problem. It has become a problem of architecture, coordination, authority, proof, value, governance, and finally... of institution."
The evolution has only just begun.