The standard argument about artificial intelligence in space is that law has failed to keep pace with autonomy. That is true, but also imprecise. The most important omission in today’s legal architecture is not that the 1967 Outer Space Treaty lacks a theory of machine agency. It is that the treaty system was built for objects, operators and states that could narrate their conduct after the fact. Autonomous systems complicate that assumption. A spacecraft that allocates power, changes trajectory, avoids conjunctions, prioritises scientific tasks or manages surface operations through adaptive software can alter events in ways no existing treaty expressly requires a state to document in a form others can inspect.
That makes the central governance question less metaphysical than evidentiary. International law does not need to decide whether an autonomous agent is a legal actor in its own right. It needs a practical answer to a narrower problem: when an AI-enabled space system causes harm, creates interference, approaches another object, or enters a disputed operational context, what record must exist to explain why the system acted as it did, who authorised the architecture, and which state bears responsibility for the resulting conduct.
This is where the silence matters. The Outer Space Treaty assigns responsibility to states, including for national activities carried on by non-governmental entities, and requires authorisation and continuing supervision. Yet it says little about the evidentiary infrastructure needed to make supervision meaningful when software learns, adapts or acts under constrained communications delay. In mid-2026, a workable standard looks less like a new constitutional settlement for robots in space than like a mandatory logbook for machine conduct.
The treaty system assumes explainable operators
The Outer Space Treaty remains the constitutional core of space law. Article VI places international responsibility on states for national activities in outer space, whether undertaken by governmental agencies or non-governmental entities. Article VII links launching states to liability, while Article IX requires due regard and consultation where harmful interference may arise. The Liability Convention and Registration Convention extend that structure. None of these instruments contemplates software agents directly. That is often described as an historical curiosity. In practice it is a structural problem.
The treaties assume that a responsible state can identify the relevant object, operator and chain of decision. Registration helps answer what was launched and by whom. Liability rules help answer who may pay when damage occurs. But autonomous operations increasingly obscure how a decision was made in the moments before a consequential manoeuvre or allocation of resources. The legal problem is not that machines become persons, but that states can no longer explain what happened.
Why evidence matters more than ontology
Much commentary treats autonomous spacecraft as if the pressing issue were legal personality. That is a distraction. International law routinely governs complex systems without granting them personhood. The harder question is whether a state can still discharge its obligations of authorisation, continuing supervision and due regard when the operational decision is generated by software that was not individually commanded in real time.
Earthbound AI governance has already converged on a useful insight. The OECD principles, NIST’s AI Risk Management Framework and UNESCO’s recommendation all place unusual emphasis on traceability, documentation, accountability and human oversight. Those concepts are often presented as ethical desiderata. In space they become jurisdictional necessities. If two spacecraft interfere with one another, or if an autonomous lander contaminates a sensitive site, the immediate diplomatic question is not whether the AI was intelligent. It is whether the responsible state can furnish a credible account of system design, mission constraints, sensor inputs, command envelopes and decision outputs.
The legal problem is not that machines become persons, but that states can no longer explain what happened.
The governance gap is acute where latency is normal
Autonomy in orbit and beyond is not merely a convenience. In cislunar space, on the lunar surface or farther afield, communications delay and intermittent connectivity make local decision-making attractive and sometimes unavoidable. Collision avoidance, terrain response, power management, thermal balancing and prioritisation of mission tasks may all be delegated partially to software. The farther a mission travels, the thinner real-time human supervision becomes.
The legal problem is not that machines become persons, but that states can no longer explain what happened.
That creates a mismatch with legal language drafted for an era in which meaningful control was implicitly proximate. “Continuing supervision” under Article VI was never defined with machine autonomy in mind. A state may supervise well at the point of licensing, testing and mission design, yet poorly at the point where adaptive software makes an unanticipated trade-off. The answer is not to demand impossible joystick control across millions of kilometres. It is to redefine supervision as a duty to impose ex ante operational boundaries and ex post auditability.
Registration today records hardware, not decision architectures
The Registration Convention was designed to improve transparency through basic data about launched objects. That remains useful, but it is no longer sufficient. For AI-enabled missions, the significant public fact may not only be mass, orbit and function. It may also be the level of autonomous authority granted to onboard software; the mission phases in which autonomous manoeuvring is permitted; the existence of kill-switches, safe modes or geofenced no-go areas; and the identity of the state authority responsible for oversight.
There is no realistic prospect of states publishing source code or commercially sensitive model details as part of routine registration. Nor is that necessary. What is missing is a tier of standardised operational metadata: enough to permit other states to understand the autonomy profile of a mission and enough to support later reconstruction if events go wrong. Space governance has long relied on registries to make objects legible. Autonomous operations require registries that also make delegated discretion legible.
A black box for space AI should be treated as legal infrastructure
Aviation offers an imperfect but useful analogy. Flight recorders do not prevent all accidents; they make accidents investigable. In orbit, a black box is not merely a safety device; it is a constitutional device for accountability. The point is not a sealed orange container, which may be physically impractical for some missions, but a mission-appropriate evidentiary architecture: tamper-evident logs of sensor inputs, system state, model versioning, command constraints, human overrides, significant autonomous decisions and communication interruptions.
Some of this information already exists in engineering practice. The governance failure is that it is not generally framed as a legal minimum. A workable standard would require states, through domestic licensing and supervision, to ensure that AI-enabled missions preserve retrievable records sufficient for accident investigation, interference review and dispute resolution. Where bandwidth is constrained, summary event logs could be downlinked periodically, with fuller records retained onboard and protected against alteration. Where missions are deep-space or long-duration, protocols for delayed but authenticated retrieval matter more than immediate disclosure.
In orbit, a black box is not merely a safety device; it is a constitutional device for accountability.
What a workable standard would actually contain
The sensible route is a technical standard anchored in existing treaty duties rather than a dramatic new treaty on robot rights in space. Such a standard could be adopted first through national licensing rules, soft-law guidelines in multilateral forums, or model clauses in bilateral and plurilateral cooperation arrangements. Its substance would be relatively prosaic.
- Autonomy declaration: every mission would classify the scope of onboard autonomous authority by function and mission phase.
- Design accountability: a named responsible organisation and designated accountable officer within the licensing state would attest to testing, risk assessment and operational limits.
In orbit, a black box is not merely a safety device; it is a constitutional device for accountability.
- Decision logging: systems would record consequential decisions, material sensor states, confidence thresholds where relevant, and whether a human command, pre-set rule or adaptive process drove the action.
- Version control: all updates to models, software parameters and autonomy rules would be time-stamped and attributable.
- Override architecture: missions would specify safe modes, interruption mechanisms and the conditions under which autonomous authority degrades or transfers.
- Retention and access: logs would be preserved for a minimum period and made available to the licensing state for investigation, and to affected states through agreed procedures where incidents occur.
- Incident notification: states would notify others not only of harmful events, but of serious loss of autonomy integrity, corrupted logs or unexplained behaviour affecting due regard obligations.
None of this requires consensus on frontier questions of AI consciousness or machine rights. It requires agreement that responsibility without records is empty.
Due regard becomes measurable only if conduct is reconstructable
Article IX of the Outer Space Treaty is often discussed in broad diplomatic terms: avoid harmful contamination, conduct activities with due regard to others, consult if potentially harmful interference appears likely. AI turns these into evidentiary questions. Suppose an autonomous servicing vehicle approaches another object; suppose a lunar rover reroutes across an area another state regards as operationally sensitive; suppose a collision-avoidance algorithm executes a manoeuvre that forces a third party to burn fuel. In each case, the debate will hinge on whether the conduct was foreseeable, bounded and reviewable.
“Due regard” cannot be judged solely from outcomes. It depends on the quality of precautions and the ability to reconstruct decision paths. A state that cannot show what constraints its system operated under will struggle to persuade others that it exercised due regard. By contrast, a well-documented mission may still make mistakes, but it can demonstrate that the mistake arose within a supervised and reviewable framework. That distinction matters diplomatically as much as legally.
Liability law is broad enough, but proof may fail
The Liability Convention does not collapse simply because software is involved. Damage caused by a space object remains attributable through the treaty’s existing framework. The more serious difficulty is proof. Establishing fault, causation and foreseeability becomes harder when the proximate decision is generated by a model whose behaviour depends on dynamic inputs, post-launch updates or emergent interactions with other systems. A state confronted with a claim may argue that the incident was not reasonably predictable or that the relevant chain of causation cannot be established from available telemetry.
An evidentiary standard would not eliminate disputes, but it would narrow them. If mission logs are mandatory, tamper-evident and structured for review, arguments over causation become less opaque. This is especially important for incidents short of catastrophic damage: interference, resource obstruction, contamination risks, unsafe proximity operations or repeated near misses. Space law often appears toothless not because it lacks norms, but because facts are too sparse to apply them confidently.
A credible standard would regulate evidence before it tried to regulate intelligence.
Soft law is the likely venue, but domestic licensing is the lever
No major power appears poised in mid-2026 to reopen the core space treaties for comprehensive amendment. The more realistic pathway runs through soft law and national implementation. The long-term sustainability guidelines developed in the UN system already normalise practices around information-sharing, safety and mission planning. UN discussions on responsible behaviours in space have also shifted attention from abstract capability categories to observable conduct. That intellectual move favours an evidentiary approach.
Still, the most immediate leverage sits with domestic authorisation regimes. Article VI already obliges states to authorise and continually supervise national activities. Licensing conditions can therefore require autonomy declarations, record-keeping standards, anomaly reporting and post-incident cooperation. If a critical mass of launching states converges on similar requirements, a de facto international baseline could emerge without waiting for a new treaty text. In this field, administrative law may do more work than diplomacy.
The hardest cases will involve multi-state and adaptive systems
The standard cannot assume a single spacecraft, one operator and one jurisdiction. Many missions already involve components developed across borders, cloud-linked planning systems, foreign launch services, shared ground segments or model updates supplied after launch by different entities. The legal simplicity of “launching state” can obscure operational complexity. For autonomous systems, that complexity is dangerous if it diffuses accountability.
A sound regime would therefore require a responsibility map before launch: which state authorises which subsystem, who controls updates, who maintains logs, and which authority can furnish a complete incident record. This matters even more for adaptive systems whose behaviour changes with experience. If a model fine-tunes on mission data, the evidentiary trail must capture not only initial certification but the pathway of change. Otherwise “continuous learning” becomes a sophisticated way of making responsibility discontinuous.
Why this matters most on the Moon and in cislunar space
Low Earth orbit gets most attention because congestion is immediate. Yet the sharper governance test may come in cislunar and lunar operations. Surface missions will rely on autonomy for navigation, site selection, extraction support, power balancing and traffic coordination in environments where communication constraints and harsh conditions reward local machine judgement. At the same time, legal and political sensitivities over access, interference and heritage protection are likely to intensify.
In such settings, ambiguity is combustible. A rover that autonomously reroutes around terrain and enters another mission’s operational perimeter may not cause physical damage, yet it can trigger claims of unsafe behaviour or strategic probing. The ability to reconstruct why the route was chosen, what constraints were encoded, and whether a human supervisor set those parameters may be the difference between a manageable consultation and a geopolitical dispute.
The next phase of space governance is administrative, not theatrical
There is a tendency in technology policy to assume that new machines require new metaphysics. Space governance would do better to begin with paperwork. Not bureaucratic clutter for its own sake, but disciplined record-making that preserves the link between state responsibility and machine conduct. A credible standard would regulate evidence before it tried to regulate intelligence.
The Outer Space Treaty falls silent on autonomous agents in one obvious sense: it never names them. But its deeper silence lies elsewhere. It does not specify the documentary conditions under which state responsibility remains meaningful once action is delegated to software beyond immediate human reach. Filling that silence need not await a grand bargain. It requires a shared expectation that every autonomous mission carries its own legal memory: a traceable account of who authorised what, under which constraints, and why the machine acted when it did.
If that expectation takes hold, the law of space need not treat AI as a sovereign mystery. It can treat it as something more governable and, in the end, more useful: a system whose decisions are bounded, logged and answerable to public authority.
