Hallucination — the generation of fluent, confident, fabricated content — is the foundational safety property of large language models, not a bug to be patched. In general use it's a productivity tax; aboard a ship it's the difference between decision support and a liability. A model that invents a COLREGS rule, a buoy characteristic or a port approach depth is worse than one that says 'I don't know'.
Maritime is unusually exposed: the domain is rule-bound, numerate and citable — exactly the properties that make hallucination both dangerous and, uniquely, checkable. This article maps where the risk lives and what the defences buy.
Key findings
- Hallucination is a property of the model class — a probability-based generator that does not distinguish true from plausible. It is not fixable, only manageable. Established
- Maritime's risk surface is asymmetric: fabricated rules, wrong positions, invented citations — each is a direct safety or compliance hazard, not a cosmetic one. Established
- Retrieval-grounding measurably reduces factual hallucination by forcing the answer to quote a real source — the most practical maritime mitigation. Research demonstration
- Constrained outputs — forcing the model to answer only from a closed set of tools/actions — is the architecture the verified agentic work uses to bound hallucination. Research demonstration
- The residual risk is the confident-close-enough answer: a number that's plausible, sourced to a real document, but wrong in the specific case. Maritime-GPT analysis
The three hallucinations that matter aboard
- <strong>Regulatory hallucination.</strong> An invented or stale rule cited as current — the model confidently states a MARPOL limit or MASS requirement that's wrong, or no longer in force. The citation looks right because the model knows what a citation looks like. Established
- <strong>Factual hallucination.</strong> A position, depth, range or capability stated wrongly — the model that 'knows' a port's approach depth it was never given. Established
- <strong>Reasoning hallucination.</strong> A chain of logic that is internally fluent but wrong — the COLREGS analysis that selects the right rule and misapplies it. The most dangerous kind because it survives a skim. Maritime-GPT analysis
Why maritime is uniquely exposed
Three properties make maritime the worst case for hallucination and, paradoxically, the best case for containing it. It's <strong>rule-bound</strong> — the right answer is checkable against a document. It's <strong>numerate</strong> — positions, speeds, depths are exact, so a wrong number is detectable if it's checked. And it's <strong>safety-cascaded</strong> — a hallucinated rule propagates into a manoeuvre. The same properties that make a hallucination dangerous make a retrieved, cited, checked answer verifiable — if the system is built to check. Maritime-GPT analysis
The defences that actually work
- <strong>Retrieval grounding.</strong> The model must quote the document, not recall it — cutting fabricated-fact hallucination at the cost of retrieval errors. Research demonstration
- <strong>Constrained action space.</strong> The LLM selects from a closed set of tools/queries rather than generating free-form output — the architecture that made agentic diagnostics safe aboard. Research demonstration
- <strong>Uncertainty surfacing.</strong> The model must say 'insufficient information' rather than default to a plausible guess — a trained behaviour, not a free property. Maritime-GPT analysis
- <strong>Human verification of anything consequential.</strong> The MASS-consistent position: the model informs, the human confirms — hallucination is tolerated only where its blast radius is zero. Established
The irreducible floor
No defence eliminates hallucination — grounding reduces factual errors, constraint reduces action errors, verification reduces propagated errors, and the combination still leaves the confident-nearly-right answer that survives all three. The honest design treats the LLM's output as <em>recommendation requiring confirmation</em>, not fact — which is precisely the relationship Voyage Intel Pro's decision-support positioning already asserts. The safety case is built on never trusting the fluent answer alone. Maritime-GPT analysis
What we know
- Hallucination is inherent to LLMs; it's a property to manage, not a bug to remove.
- Maritime's rule-bound, numerate, citable nature makes hallucination checkable — if the system checks.
- Retrieval + constrained action + human confirmation is the effective mitigation stack.
What we don't
- A quantified hallucination rate for maritime-domain LLMs on real operational queries.
- How well grounding works when the retrieved corpus itself has errors or gaps.
- Whether uncertainty-flagging can be made reliable enough to trust alone.
Being tested
- Retrieval-grounded benchmarks for maritime regulatory QA.
- Constrained-tool architectures vs free-form generation.
- Calibrated-uncertainty training on domain data.
Maritime-GPT analysis
The maritime AI systems that survive the safety review are the ones built assuming the model will confidently lie. That isn't cynicism — it's the correct prior. Ground it in retrieved documents, constrain its action space, flag its uncertainty, and never let it act unverified. The systems that fail are the ones that treat fluency as correctness. Hallucination isn't the reason maritime AI is dangerous; it's the reason it must be designed as decision support rather than authority. 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 canonical taxonomy and mitigation catalogue for LLM hallucination.
The grounding architecture that is the principal maritime mitigation.
The retrieval architecture as a hallucination defence.
The constrained-tool pattern that bounds what the model can get wrong.