A revision of “A Journey Through the Stages of AI Integration: Introducing the Aiciety Classification”

Why a revision

The AIciety Stages of 2023 had a correct intuition and a wrong frame. The intuition: the integration of AI into a society can be observed on a scale, and a shared vocabulary for its stages is more useful than talking about “AI” as if it were the same thing everywhere. That intuition has held. The frame, however, was a timeline: ten stages following one another and ending in a society in which nobody needs to work any more. In 2023 this was a forgivable reflex – ChatGPT was six months old – and in 2026 it is no longer tenable.

Three assumptions of the first version have not survived the years since:

  • That the stages follow one another. They coexist. In 2026, software development works with autonomous agents while across much of Africa, Asia and Latin America, according to the International AI Safety Report 2026, fewer than one in ten people use AI at all. The same society stands on several stages at once, depending on sector, region and person.
  • That the scale has a destination. The upper stages were written as prophecy. The OECD considers four trajectories equally plausible for 2030 – from stagnation to acceleration. A scale that knows its ending is no longer an observation but a narrative.
  • That work is the measure. Stages 7 to 10 measured progress by how much human labor becomes “redundant”. The evidence of 2026 shows reallocation, not disappearance: no overall effect on employment, but declining demand for early-career workers in exposed occupations. And the real event is a different one: it is not work that disappears but capability that is distributed. Anyone who pays for a subscription commands means that in 2023 required departments.

A fourth assumption was tacit and is the most consequential: that society as a whole climbs the stages – as an acting unit that “develops frameworks” and “adapts”. It does not. Stages are reached by individuals who use a system; by companies that build on it; by capital that finances data centers. Society follows and discovers where it stands. This revision keeps the ten stages as vocabulary and replaces the frame: the timeline becomes a map, the prophecy becomes a measuring instrument, and work as the yardstick gives way to the question of the degree to which AI has become part of a society’s structure.

How to read the scale

The scale measures a degree, not a point in time: how far AI has moved from tool to structure in a given domain. It can be applied to a society, but equally to a sector, a company, a household. The value that comes out is always that of a slice; whoever places “society” on a stage is naming an average across very unequal parts.

Three reading notes:

  1. Map, not timeline. The stages describe states that occur simultaneously. A society does not move as a whole from 6 to 7; its software industry stands at 7, its healthcare at 5, its administration at 4. The interesting quantity is not the stage but the spread – and who causes it.
  2. The measure is structure, not work. The lower stages ask whether AI is conceived, built and used; the upper ones ask whether institutions would still function without it. The four thresholds from the definition of AIciety – delegated agency, machine-to-machine communication, institutional reorganization, recursive self-reinforcement – mark the transitions between stages 6 and 10.
  3. The driver is individuals. No stage is reached by resolution. It is reached when enough people and companies use a system because it serves them, and when someone risks the capital that builds the infrastructure for it. Policy cannot lift a society up. It can slow it down – and even that only locally: what may not be built in one jurisdiction is built elsewhere and imported, along with the rules that apply there.

Each stage is described below with three entries: what characterizes it, how to recognize it, and where it applies in 2026. The dates of the lower stages refer to Western technological history; other world regions have passed through them in a different order and at different speeds, or skipped them.

The ten stages

Level 0 – No concept

Society has no notion of artificial intelligence. The desire for a thinking artifact exists only indirectly – in myths of the animated stone, in golem and automaton as story, not as project. How to recognize it: there is no word for it. Status in 2026: historical.

Level 1 – Imagination

Society can imagine AI but cannot build it. Fiction and mechanical simulations – the Mechanical Turk, automata – anticipate what the technology cannot do. How to recognize it: the debate is conducted in novels. Status in 2026: historical; fiction still supplies the images in which fears and hopes are expressed.

Level 2 – Foundations

The theoretical and technical foundations emerge: Turing’s question of 1950, the Dartmouth conference of 1956, the first programs. AI is considered feasible; first reflections on application and ethics appear long before there is anything to apply. How to recognize it: research funding without a product. Status in 2026: historical – with one lesson: the precautionary debates of this stage almost never matched the later reality.

Level 3 – First applications

Simple systems show what might be possible without delivering it – ELIZA, early translators, game programs. How to recognize it: demonstrations impress, products are missing. Status in 2026: historical.

Level 4 – Specialization

AI works in bounded fields: expert systems, industrial automation, chess. Society discusses benefits and risks concretely for the first time, but as a specialist debate. How to recognize it: AI is a tool for specialists. Status in 2026: where digital infrastructure is lacking, this is still the prevailing form.

Level 5 – Diffusion in the background

AI is present in many companies and households without being noticed: search engines, recommendations, spam filters, route planning, image recognition. Regulation covers only peripheral aspects, above all data protection. How to recognize it: the technology is there, the concept is not. Status in 2026: large parts of the world’s population and of most administrations stand here – the 2010s that never ended.

Level 6 – General use

General-purpose models are used directly by broad sections of the population: writing, learning, research, coding, conversation. Society debates AI publicly and begins to pass laws – mostly before it knows what it is regulating. How to recognize it: at least a third of adults use a model weekly; AI is an election and school topic. Status in 2026: the stage of Western societies as a whole. Worldwide, according to the Safety Report, at least 700 million people use leading systems weekly; in some countries more than half the population – the fastest diffusion of any technology in history, without any program to diffuse it.

Level 7 – Delegated action

AI systems carry out tasks rather than merely answering: they use tools, write and review code, execute transactions, negotiate. The first threshold of the definition has been crossed. Institutions begin to organize themselves around the capability: examinations are redesigned, software work becomes oversight of agents, research runs through automated hypothesis generation. This is the greatest decentralization of capability since the printing press – every individual commands a staff that until now only corporations and governments could afford. How to recognize it: the question is no longer “What does the system answer?” but “What has it done, and who is liable?” Status in 2026: reached in software development, parts of customer service and of science; the length of tasks agents complete autonomously doubles roughly every seven months according to the Safety Report; among small and medium-sized enterprises, the OECD’s 2026 survey finds only 3.6 percent using agentic systems so far.

Level 8 – Machine to machine

A relevant share of social communication is authored and read by no human: agents negotiate with agents; contracts, orders and code travel between systems. The second threshold has been crossed; AI is a participant in communication, not merely its medium. Society finds that it behaves as if machines were members – and the law must set points of attribution it did not previously need. How to recognize it: there are markets on which machines trade on both sides. Status in 2026: emerging – agent protocols, multi-agent systems; the OECD describes this as a “socio-technical paradigm”.

Level 9 – Recursion

AI accelerates the development of AI. The pace of the technology is set by the technology, and society can no longer catch up with it through procedures designed in years. The fourth threshold has been crossed. How to recognize it: the most important research results come from systems that design their successors. Status in 2026: not reached and contested; forecasters put the probability of a dramatic acceleration by the end of the decade at 8 to 20 percent. The 2023 version did not know this stage; it is the most important addition.

Level 10 – Permeation

The distinction between AI and society no longer has analytical use: AIciety in the full sense of the word. The 2023 version described this stage as a state in which “most people no longer need to work”. That was a promise nobody can keep, and a picture that makes the human being a recipient. More accurately: what counts as work has shifted, as writing and the printing press shifted it. What this stage looks like nobody can plan; whoever tries will not reach it. How to recognize it: one cannot know before being there. Status in 2026: a possibility, not a forecast.

Where we stand in 2026

To the question of which stage “we” are on, there is no single answer in 2026 – and that is precisely the finding. The spread is wider than at any point in the history of technology, and it does not run along the borders one would expect.

  • The laboratories: Level 8 to 9 in trial. Multi-agent systems are operated there, and AI is already being used to accelerate AI research – with contested results but undeniable intent.
  • The leading sectors: Level 7. Software development, parts of customer service, parts of science work with delegated action and have begun to organize around it.
  • Western societies as a whole: Level 6. Broad use, public debate, first laws – and a healthcare system, an administration, an education system that mostly stand at Level 5 or 4.
  • The EU in particular: Level 6 with the brakes on. The AI Act sought to regulate Level 7 in advance and had to defer its high-risk obligations in 2026 because the standards they rested on did not exist. This has not prevented Level 7 in Europe; it has relocated it – to providers outside and to users who import their systems.
  • The larger part of the world’s population: Level 5 or below. Not because rules are missing, but electricity, networks and access.

Two observations on this map. First: the distance between laboratories and everyday life is two to three stages, and it is growing, because the frontier doubles every seven months and everyday life moves at the pace of procurement procedures. Second: the distance between world regions is not technical. The same models that run in San Francisco run in Lagos – if the power holds and access is not blocked by export controls or licensing requirements. The scale therefore measures not only how far a society has come, but also how much it stands in its own way.

What moves a society up – and what does not

Every ascent on this scale in the past three years had the same cause: someone was allowed to build something, and someone was allowed to use it. No stage was reached through a strategy, no transition triggered by a framework. The 2023 version credited the upper stages to society “developing ethical and legal frameworks” – as if that were the engine. It is the reaction. The engine is three things:

  • Freedom to build. Open research, open-weight models, the exchange of weights and methods. That the capabilities of a frontier model can be transferred to a smaller one for a few thousand dollars is the reason Levels 6 and 7 no longer happen in only three cities.
  • Freedom to use. Access without approval. The 700 million weekly users asked nobody; they found a tool that served them and kept it.
  • Capital and energy. Data centers worth hundreds of billions of dollars, privately financed, at private risk – and power grids that carry them. Where either is missing, a society stays at Level 5, however clever its laws.

Conversely, there is nothing that lifts a society up: no subsidy programs that replace what users do not want; no frameworks that generate capabilities. There is only what slows it down – and even that works differently from how its authors believe. A rule that makes building conditional on proof of harmlessness prevents not the risk but the knowledge of the risk; it relocates Level 7 to another jurisdiction and imports its results. Restriction is not freedom, and it is not even protection. What the upper stages actually require is older than any technology: liability for harm, transparency about provenance, prohibition of fraud and deception, courts that learn from cases. These institutions have survived every earlier stage. They will survive this one too – if they are allowed to work rather than being replaced by permissions.

An instrument, not a prophecy

The AIciety Stages are an instrument of observation. They do not say where a society is going, but where its parts stand and how far apart they are. Whoever uses the scale for prediction makes the same mistake as the 2023 version: confusing a direction with a destination. The upper stages are possibilities whose arrival depends on decisions nobody has yet made – in laboratories, in data centers and in billions of everyday lives in which someone trusts a system or does not.

The first version of this post, of May 9, 2023, remains online; it documents how the concept arose and what seemed plausible six months after ChatGPT. This revision will meet the same fate. That is not a flaw but the method: a scale for a technology that doubles every seven months must grow with it.

Criticism, counter-proposals and other calibrations are welcome – in the comments or via the contact page. Of particular interest: who sees their own sector on a different stage, and why?

Sources: International AI Safety Report 2026 (February 3, 2026) · OECD: The agentic AI landscape and its conceptual foundations (February 13, 2026) · OECD D4SME Survey 2026 · What is AIciety? (definition with the four thresholds).

Leave a Reply

Your email address will not be published. Required fields are marked *