A delivery drone swerves around an obstacle and damages a parked car. Who is liable? The answer turns on simple questions. Which device was it, which software version was running, did the evasive command come from a human or from the onboard logic? In principle each of these can be answered. In practice no one was obliged to record them, and where records exist they sit with the operator and can be changed after the fact. The machine acted, and yet it has dropped out of the chain of responsibility.
This is where a speech comes in that Prof. Tom Fuerstner, founder of the Vienna machine-identity company Riddle&Code, gave in the German Bundestag on 10 June 2026. His proposal is sober: machines that act need an address. Such an address is neither personhood nor a right. A verifiable entry is enough, one that makes a system addressable and distinguishable and connects it to liability, oversight and correction. “Whoever has no address,” Fuerstner says, “drops out of the chain of responsibility.”
What is striking is what Fuerstner does not ask. Whether machines think or become conscious he calls interesting for philosophers and not the most pressing political matter. The pressing question is a different one: is an acting system accountable? This shift, away from the question of consciousness and toward the question of how the action is structured, is also the move Path Ethics makes.
Path Ethics does not judge a system by its essence. It asks how the system is built: how independently it acts, how it is coupled to other systems, and in what way it persists over time. Three questions that apply to organisms as much as to software.
Why Oversight Grasps at Nothing
The third of these questions explains a growing problem. Classical device and product oversight was historically designed for systems whose action-relevant state remains relatively stable after approval. Learning and continuously updated systems put that assumption under pressure. A toaster certified in 2024 still toasts the same way in 2026. A learning system does not. A purchasing agent whose model has been updated twice since a questionable order is, at the time of inspection, a different system than at the time of the act. Oversight may then be examining a different system state from the one that was relevant to the act in question. Path Ethics calls this a persistence illusion: an order appears stable because its form endures, while its actual grip has long since eroded.
Fuerstner's second proposal answers this. He calls it notarisation: a machine that acts should leave traces that cannot later be rewritten at will, at registration, at the release of a mission, at the change of a software version. The model is old. Aviation knows it as the flight recorder, a deliberately rigid device that holds unchanged what happened while everything around it is in motion. The more mobile the system, the more rigid the layer must be that keeps it reconstructable.
The Second Question
So far the demand holds. But an address answers only one question: who acted. It leaves a second one open: what that action does to the people and systems it is connected to. Imagine a surveillance fleet that fulfils everything the address debate asks for. Every device identified, every mission signed, every chain of responsibility complete. This fleet is fully accountable. And it can at the same time hollow out the ease with which the people under surveillance move and assemble in public space. Complete addressing, and harmful all the same.
Path Ethics distinguishes here between couplings that strengthen both sides and those in which one side gains at the other's expense. Accountability and harmlessness are two different things. An address order can itself become a highly precise instrument of surveillance. A machine order that asks only for addresses and leaves the quality of the couplings unexamined remains half-finished.
The Window
Such an order is taking shape. Europe's identity infrastructure is being built out by the end of 2026, the duties of the AI Act take effect in stages. Where the legislator leaves the address question open, platforms and device makers answer it daily, through product decisions and de-facto standards. Such an architecture is easy to shape early and hard to change later. Those who are included only after it has hardened have significantly less room to shape it.
Fuerstner is right about the address. The open question is whether the coming machine order also asks the second one, while its architecture is still fluid.
→ The full short analysis “Addresses for Machines” follows Fuerstner's speech step by step, with sources, case examples and three checking questions for the legislator. Download the analysis as PDF.