We have inherited a vocabulary for describing human-machine relationships that was built for cockpits and factory floors that is no longer adequate. As artificial intelligence moves from tool to agent, and from agent to sovereign, we need a more precise account of what it means to delegate decision-making to a non-human system. This matters for both business and politics. The AI autonomy proposed in this article provides the frame to understand the evolution towards a non-human government, where humans gradually delegate their decision-making to machines for performance reasons. The question then is how to keep these systems legitimate.
The foundational taxonomy comes from human factors engineering. In 1978, Thomas Sheridan and William Verplanke proposed a ten-level scale of automation, from “the computer offers no assistance” to “the computer acts autonomously and ignores the human.” Forty years later, the same logic was adapted by the Society of Automotive Engineers to classify self-driving vehicles, producing the now-familiar SAE Levels 0 through 5 from “no automation” to “full self-driving.” This type of scale is a useful tool to discuss AI autonomy. However, it doesn’t yet exist in the realm of political science, probably because it long belonged more to science-fiction than present day reality. But things have changed now.
Most frameworks were designed for discrete, bounded, engineered systems operating in defined physical environments. A car, an aircraft autopilot, an industrial robot. Their unit of analysis is the task: at what point does a machine take over the steering, the braking, etc? What they cannot describe (and were not designed to describe) is the gradual migration of cognitive and political agency from human actors to non-human systems operating not on roads or in factories, but inside the epistemic infrastructure of governance itself.
Here I propose a different scale. designed to understand non-human authority. Not for engineered systems, but for the institutional and governmental deployment of AI. The question it asks is not “how much of the driving does the machine do?” but “at what point has sovereign decision-making migrated to a non-human actor?” and, crucially, whether anyone noticed when it did. Because this shift is happening right now.
The difference matters because the political stakes are entirely absent from engineering taxonomies. An autopilot or autonomous car are not illegitimate; they are certified, regulated, and subject to liability frameworks. But an autonomous welfare determination system raises questions that no airworthiness certificate can answer: who is responsible when it errs? What procedural rights does the claimant hold? By what criteria is performance defined, and who defines those criteria?
“At Level 4, the human decision-maker remains formally in place, signing, voting, announcing, while the substantive content of the decision has been shaped upstream by non-human systems.”
A scale to name AI autonomy levels in governments
I propose six levels, defined by the substantive authority. It is meant to understand the degrees of human sovereignty over political decision-making. The different levels go from fully sovereign human, not assisted by AI tools, to having more and more AI tools assisting decision-making, shaping human thinking, suggesting solutions being merely ratified by humans, to the level of AI autonomy where it is taking decisions without human input or veto. We are more advanced than Level 0, and obviously not at Level 6. You, and any observer, are able to see at which level we are today, which will differ depending on the AI system we’re considering.
Because this is happening more and more, and we aren’t likely to go backward, it is of the greatest importance that we create solutions to keep this new political decision-making legitimate.
This scale maps the gradual evolution of AI from a political tool to a political actor.
The invisibility of the transition to AI autonomy in politics
The most consequential feature of this scale is not the distance between Level 0 and Level 5. It is the invisibility of the movement from Level 1 to Level 2. A recent study from University College Dublin illustrates this with uncomfortable precision. Researchers found that AI-generated summaries of European Parliament debates systematically distorted political representation: one political group’s share of the summarised record doubled, while another’s was cut by more than half, and a third disappeared entirely from the algorithmic account. Nobody decided this. No policy was changed. The distortion is a structural consequence of how language models weight and select content. This is not science-fiction, it is already operating at the heart of EU institutions and has consequences in today’s world.
This is a Level 2 failure (filtered cognition producing invisible epistemic distortion) but it illustrates the mechanism that drives the entire scale. The transition between levels is not debated or even announced. Without a scale reference, the move from unmediated deliberation to filtered cognition was likely not even conceptualised in that sense. The shift happens incrementally, invisibly, and often under the banner of efficiency. By the time it becomes visible, it is typically irreversible. Not because reversal is technically impossible, but because the institutional infrastructure, the decision-making culture, and the political economy of the new arrangement have been rebuilt around it. Who would come back from automatic translation? Who would go back to hand-written minutes of meetings?
Nobody returned from statistical governance to pre-census administration. Nobody dismantled the bureaucratic state to restore charismatic oral sovereignty. The sequence of transformations in political cognition throughout human history (oral, scriptural, statistical, computational) is not reversible. The question is therefore not whether non-human systems will govern, but whether the degree of AI governance will be legitimate.
What legitimacy requires at each level of AI autonomy
A theory of legitimate non-human government cannot simply demand a return to Level 0. That demand is historically naive and politically inert. What it can require (and what democratic theory, properly updated, does require) is that three conditions be met at whatever level a system operates.
First, alignment: the objectives optimised by non-human systems must genuinely correspond to human welfare comprehensively defined, not to proxy metrics. A system that optimises parliamentary summary for salience by the standards of its training data is not aligned with democratic representation, even if it is technically optimising for something that correlates with engagement. The technical alignment problem is, at its core, a political philosophy problem.
Second, demonstrability: the performance of non-human systems must be independently verifiable and contestable by the governed. A system that operates at Level 3 or Level 4 and cannot be audited is opaque and therefore ungovernable. In worst-case scenarii, it could even lead to a new type of performance authoritarianism.
Third, revisability: the governed must retain collective authority to redefine what counts as good performance. This is the democratic remainder in any legitimate account of non-human governance. Not consent to each decision, which is neither possible nor necessary, but sovereign control over the criteria by which decisions are judged. This condition is what distinguishes legitimate non-human government from technocratic capture or from the performance legitimacy model in which the governing party defines its own success criteria.
Where we are
Most advanced democracies currently operate between Level 2 and Level 3 in the domains that matter most: information environments, public administration, welfare, and the labour market. The drift toward Level 4 is visible and accelerating. Argentina’s proposed legislation for autonomous AI corporations (entities operated by AI agents with limited liability and no required human shareholders) marks the first serious legislative acknowledgment that Level 5 is not a thought experiment.
The Sheridan scale was useful because it gave engineers a shared language for thinking about where human oversight ends and machine authority begins. We need this scale to provide a shared language to discuss AI autonomy in the political domain. Not to resist the transition, which is underway and will not be reversed, but to distinguish, at each level, between legitimate and illegitimate forms of non-human governance. Without that language, we will not notice when the transition happens. We will simply find, one day, that we are already there.
