Maritime law is built on a person — the master, the officer of the watch, the pilot — whose judgement the regulation addresses and whose licence the state holds. A MASS changes the addressee: the decision-maker is now software, and the question becomes how a regulation written for human judgement certifies a machine's.
The MASS Code's answer, taking shape through the IMO's non-mandatory framework, is to regulate the system around the decision-maker — the operating envelope, the assurance process, the human fallback — rather than the algorithm's internals. This article is about that regulatory move and its consequences.
Key findings
- The Code regulates the system, not the algorithm — certification attaches to the operational-design-domain, assurance evidence, and fallback, not the model's internals. Established
- The operating envelope is the core instrument — the AI's authority is bounded by the conditions it was validated for. Established
- Software assurance is the certification problem — how to evidence that an adaptive system's decisions are trustworthy. Established
- Human responsibility doesn't disappear — it relocates to the remote operator, the company, or the shore-based supervisor. Established
- The Code is goal-based, not prescriptive — it specifies outcomes (safe navigation), leaving the technical means open. Established
The regulatory move: certify the system, not the mind
You can't licence an algorithm the way you licence a master — there's no examination, no experience requirement, no person to hold accountable. The Code's approach is to certify the system's claims: the operating conditions it's validated for (its operational design domain), the evidence that it performs within them (the assurance case), and what happens when it leaves them (the fallback). The AI isn't certified as a good decision-maker; the system is certified as a bounded one. Established
The envelope is the licence
The operational design domain — the sea states, traffic densities, visibility conditions the system was validated for — functions as the AI's licence terms: inside it, the autonomy is authorized; outside it, the system must alert and hand off. The certification question isn't 'is the AI good' but 'is the ODD defined, is it instrumented, and does the system know when it's outside it'. The bound is the mechanism the trust attaches to. Established
The assurance problem
The hard regulatory question is evidence: how do you demonstrate that a learned system's decisions are trustworthy across its operating envelope? Traditional software is verified by inspection and test; a learned system's behaviour emerges from data, which can't be inspected the same way. The assurance answer combines simulation coverage, operational monitoring, and the system's own uncertainty reporting — but the field is still converging on what 'assured' means for an AI navigator. Established
Where responsibility lives
The Code doesn't remove human responsibility — it relocates it. The remote operator, the shore-based supervisor, the company that deployed the system: someone must be accountable for the decisions the software makes, because a casualty needs a defendant. The legal architecture this implies — how far up the chain responsibility flows — is one of the Code's genuinely unresolved areas, and the reason 'who is the master of an autonomous ship' is a question, not a punchline. Established
What we know
- The Code regulates the system's bounded operation, not the algorithm's internals.
- The operating envelope is the core regulatory instrument.
- Human responsibility relocates rather than disappears.
What we don't
- The final form of the mandatory Code's decision-making requirements.
- How far responsibility flows up the chain (designer? operator? owner?).
- Whether adaptive/learning systems can be assured under current frameworks.
Being tested
- Assurance-case frameworks for learned navigation systems.
- ODD boundary-instrumentation standards.
- Responsibility allocation in remote-operation legal frameworks.
Maritime-GPT analysis
The MASS Code's quiet revolution is that it doesn't try to make the AI a person — it makes the system's bounds the regulable object. The operating envelope is the licence, the assurance case is the evidence, and the human responsibility is relocated rather than removed. It's the most sensible architecture available, but it leaves the field's hardest question open: how do you prove a learned mind is trustworthy inside its bounds? The answer to that is the certification frontier the code depends on. Maritime-GPT analysis
This is a research feature, not operational guidance. Simulation results are reported as simulation; demonstrations are not deployments; company claims are attributed as such. Verify safety-critical decisions against official nautical publications and flag-state requirements.
Sources & evidence
The official IMO MASS work this article interprets.