Where Does Legal Authority Live in an AI System?
- Russell E. Willis
- May 4
- 5 min read
Law begins with authority.
Not with harm. Not with liability. With authority.
Before a court asks who is responsible, it must assume that someone—or some institution—stood in a position to decide, to act, to govern what happened. Authority is the quiet precondition beneath doctrine. It is what allows law to draw lines between actor and action, between decision and consequence, and to hold those lines long enough for responsibility to attach.
For most of the history of law, that assumption has held. Authority could be located. A contract was signed. A product was designed. A decision was made. Even when responsibility was contested, the field on which it was contested remained visible.
Artificial intelligence does not displace that field. But it alters its structure.
In a traditional legal frame, authority appears as an event. Someone acts. A choice is made. Responsibility follows. The doctrines we rely on—tort, product liability, administrative law—presume that authority can be tied, however imperfectly, to a moment in which a decision took place.
That presumption begins to strain when we turn to contemporary AI systems.
A hiring outcome, a credit determination, a clinical prioritization—these no longer emerge from a single decision. They arise from a layered process: data selected and inherited, objectives defined and optimized, models trained and tuned, systems embedded within organizational workflows, outputs accepted or overridden, and feedback loops reshaping future behavior. Each element contributes. None alone determines.
The decision, if we still want to call it that, is no longer a point.
It is a pattern.
And so the shift becomes unavoidable: Authority is no longer best understood as an event. It is a structure.
And structures are not simply located. They must be interpreted—read across the relationships, constraints, and conditions that give rise to outcomes.
Law has long depended on the ability to locate authority. AI does not make that impossible. But it renders authority less singular, more distributed, and increasingly shaped by the systems that produce outcomes—straining the assumption that a single decision-maker can be identified as the locus of responsibility.
The idea that authority can be cleanly assigned to a single actor for a single decision begins to lose its presumptive force.
It is the kind of pressure that arises when laws formed under conditions of relative stability encounter systems shaped by a more fluid and interdependent age.
This pressure is not theoretical. It is already visible in doctrine.
In tort, causation becomes harder to isolate. Harm may be real and measurable, yet the pathway from system to outcome resists reduction—and demands interpretation across the system rather than attribution to a single act.
In product liability, the idea of defect becomes unsettled. Is the defect in the model, the training data, the objective function, or the conditions of deployment? Each component may perform as intended. The system, taken as a whole, may not.
In administrative law, agencies are required to provide reasoned explanations for their actions. Yet when those actions are mediated through complex systems, the obligation to explain remains while the object of explanation becomes more diffuse. The explanation risks becoming either too abstract to satisfy or too technical to be meaningful.
None of these doctrines collapse. They continue to operate. But they do so under increasing strain, as if they are being asked to map a terrain that has shifted beneath them.
What emerges from this strain is a condition that is easy to overlook.
Systems can be compliant without being responsible.
An organization may follow established procedures, document its processes, validate its models, and satisfy regulatory requirements. It may, in every formal sense, do what is expected of it. And still, the system it deploys may produce outcomes that are systematically biased, difficult to contest, or misaligned with the purposes for which it was introduced. This is not simply a failure of diligence. It is a misalignment between how authority is structured and exercised, and how responsibility is assigned—and, increasingly, how authority must be interpreted.
In AI and the Crisis of Control, I described this as a widening gap between the power of systems and our ability to take responsibility for what they do. Accountability, as our legal structures typically understand it, looks backward. It asks whether rules were followed, whether duties were breached, whether liability can be assigned.
Responsibility asks something more demanding. It asks whether we remain answerable for the systems we have set in motion—not only when they fail, but in the ordinary course of their operation. That question becomes harder to answer when authority itself becomes harder to isolate—and must instead be interpreted within the system.
Artificial intelligence does not remove authority. It redistributes it. It is present in the selection of data, in the shaping of objectives, in the calibration of thresholds, in the decision to deploy, in the willingness to rely on outputs, and in the presence—or absence—of oversight. It is exercised not once, but repeatedly, often indirectly, and sometimes invisibly.
What changes is not whether authority exists, but how it is structured and exercised. It becomes distributed across the system—embedded in its components, conditioned by its design, and emerging through its operation. Which means it can no longer be fully understood by asking where it resides. It must be interpreted within the system that produces the outcome.
This introduces a shift that law is only beginning to absorb.
If authority is no longer singular, responsibility cannot be limited to the final decision-maker. It must extend across the participants who shape the system: those who design it, those who train it, those who deploy it, and those who structure the conditions under which it operates.
This does not dissolve responsibility. It expands it.
But expansion creates tension. Legal frameworks are more comfortable assigning responsibility than distributing it. They depend on clear lines—of duty, breach, and causation. Participation introduces overlap. It complicates attribution. It resists the clean separations on which doctrine depends.
And yet, to ignore participation is to misinterpret how authority is exercised within the system.
The deeper challenge, then, is not simply to locate authority more precisely. It is to recognize that authority now resides in structures that extend across time, across roles, and across institutional boundaries.
This suggests a different orientation for law.
Less emphasis on isolating the decisive moment. More attention to governing the ongoing system.
Less reliance on identifying a single responsible actor. More effort to align responsibility with the architecture of participation.
Less confidence that explanation alone will suffice. More insistence that institutions remain answerable for what their systems do.
At its core, this is not only a doctrinal problem. It is an institutional one.
Do we understand ourselves as building systems that then operate beyond us, or as remaining participants in systems for which we are continuously responsible? The first posture treats authority as something exercised and then relinquished. The second treats it as something that must be sustained. Artificial intelligence does not force us to choose between these views. But it makes the consequences of that choice more visible. Because while authority may diffuse, consequences do not.
Law has long depended on the ability to locate authority. That capacity remains necessary.
But it no longer captures the full reality in which legal responsibility must operate.
The task is no longer simply to locate authority. It is to interpret how authority is exercised within the systems through which it operates—and to remain responsible within those systems even when control is no longer easy to see.
That is the work now before us.
And it is not a small one.
Comments