Confident but wrong: shipping wakes up to the AI hallucination risk

Enthusiasm running ahead of verification

Shipping’s rush to adopt artificial intelligence is creating a new operational hazard: systems that sound confident, look useful and are wrong. As Splash247 reports, generative AI is now appearing in voyage planning, maintenance, compliance, chartering, emissions reporting, cargo screening and onboard decision support — and “AI hallucinations,” where a model produces false or unsupported information, are following it into the industry.

The risk is not hypothetical. Splash247 notes OpenAI’s own SimpleQA research treats hallucination as a continuing open problem, while Stanford researchers have found leading specialist legal AI tools still hallucinate between 17% and 33% of the time. Shipping is, like law, a documentation-heavy and high-liability business: a false clause summary, a wrong port restriction, an invented regulatory requirement or a fabricated maintenance instruction can quickly turn into a dispute, a detention or an insurance problem.

Trust, not appetite, is the constraint

Industry surveys point the same way. An IMarEST survey of 130 maritime professionals found around 80% believe AI can improve efficiency, but 37% have personally witnessed AI produce false or fabricated results — and only 11% have a formal AI policy. A separate survey of 300 seafarers and officers by IT provider Dualog found little resistance to the technology itself; the constraint was trust, with companies wanting onboard AI built on vetted sources rather than the open internet.

That distinction is where responsible tools are heading. DNV’s RuleAgent, for example, searches the class society’s own rules and standards database and links every answer back to official source material — positioned, in DNV’s words, as a trusted reference rather than a black-box answer generator. Regulation is also moving: the EU AI Act’s penalty regime allows fines of up to €35m or 7% of global turnover for prohibited practices, making AI governance a legal exposure as much as a technology choice. For owners, the practical answer is controlled deployment — approved data sources, audit trails, human review, role-based access and clear accountability — with any AI output that touches safety, contracts or compliance verified before use.

Apeks view — Our position is simple: AI is excellent for ideas and perspective, but it should not be the decision-maker — not for a long time yet. It carries both a hallucination problem and an accountability gap: when it is wrong, there is no one to answer for it. Used well — researching a fault, comparing similar cases, surfacing what an engineer might have missed — it is a genuinely useful adviser. But the final call, and the accountability that goes with it, stays with a qualified human.