This is the question the whole autonomy debate reduces to: can an AI be trusted with the collision-avoidance decision — the one where a wrong answer costs hulls and lives? The honest answer is neither 'yes' nor 'no' but 'under conditions' — and understanding the conditions is the difference between knowing the technology and repeating the hype.
The evidence has genuinely moved: RL policies now demonstrate COLREGS-compliant avoidance in simulation, and the failure modes are increasingly understood. What follows is both cases — the demonstrated competence and the limits that remain — because the honest verdict requires both.
Key findings
- AI collision avoidance is demonstrated in simulation — RL policies handle COLREGS-compliant encounters at scale. Research demonstration
- The safe cases are the trained ones — open-water, clear-weather, in-distribution encounters where the policy is strong. Established
- The dangerous cases are the edge ones — unusual encounters, degraded sensors, and dynamics outside the training set. Established
- Safety isn't the policy's accuracy — it's the architecture around it: the envelope, the fallback, the human override. Established
- The honest position is conditional: AI can make the decision safely inside a bounded envelope under supervision, not unconditionally. Maritime-GPT analysis
The case for: the competence is real
The demonstrated capability is genuinely impressive — RL policies trained on millions of encounters achieve COLREGS-compliant avoidance rates approaching human performance in simulation, multi-agent coordination handles ship-to-ship scenarios, and the decision quality on the common cases is arguably better than a fatigued human's. In the trained distribution, the AI's collision competence is real and improving — this isn't hype. Research demonstration
The case against: the failure modes are specific
- <strong>Out-of-distribution.</strong> The encounter type the policy never saw — a vessel behaving irrationally, a scenario outside training — produces unpredictable behaviour. Established
- <strong>Degraded input.</strong> Sensor noise, spoofed AIS, or a corrupted feed produces a decision on false premises. Established
- <strong>The human-factor edge.</strong> The novel situation that requires judgement rather than policy — the crew in the water, the vessel in distress, the exceptional case. Established
- <strong>Opacity.</strong> A policy that can't explain its decision can't be checked — the explainability problem is a safety problem. Established
Safety is the architecture, not the policy
The honest answer reframes the question: 'is the AI safe' is the wrong framing; 'is the system safe' is the right one — and the system's safety comes from the architecture around the policy. The operating envelope bounds where the AI may decide; the fallback defines what happens when it can't; the human override provides the final authority. The AI can make the collision-avoidance decision safely <em>inside a system designed for its failure modes</em> — and that's a different claim than 'the AI is reliable'. Established
The conditional verdict
So can AI make the decision safely? Conditionally — yes, inside a validated envelope, with an instrumented fallback and a human override. Unconditionally — no, because the failure modes (out-of-distribution, degraded input, novel situations) remain live. The honest industry position is the bounded one: the AI's decision competence is real and the supervision structure is what makes it safe, which is exactly the MASS architecture the regulation is writing. Maritime-GPT analysis
What we know
- AI collision competence is demonstrated in simulation on trained encounters.
- The failure modes — out-of-distribution, degraded input, edge cases — are specific.
- Safety comes from the system architecture (envelope + fallback + override), not just the policy.
What we don't
- Field-validated collision-avoidance safety on operating vessels at scale.
- Whether edge-case coverage can ever be 'complete enough' for unsupervised use.
- The liability allocation when an AI's bounded decision still causes a casualty.
Being tested
- Envelope-bounded autonomous collision-avoidance trials.
- Fallback-behaviour certification.
- Explainable decision auditing for safety cases.
Maritime-GPT analysis
The question 'can AI decide safely' deserves an honest two-sided answer: the competence is real and the failure modes are real, and the system that contains both is the one that's deployable. The AI is good enough to trust inside its bounds — and the bounds, the fallback, and the human override are what make the trust earned rather than assumed. Safety isn't what the AI can do; it's what the architecture does when the AI can't. 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 demonstrated competence evidence.