Why every AI answer on a ship needs a citation

Ask a general-purpose AI assistant a question about your engine's cooling system and it will answer. Fluently. Confidently. Sometimes correctly.

"Sometimes" is the problem.

Ashore, a wrong answer costs an awkward correction. At sea, the stakes are different. The officer acting on an answer about isolation procedures, lifting gear limits, or a chemical cargo's compatibility is not fact-checking against a search engine — they are often mid-task, mid-watch, sometimes mid-emergency. An answer that is wrong but confident is worse than no answer at all.

Why fluent AI gets things wrong

Large language models generate text by predicting what words plausibly come next. Trained on the broad internet, they hold an enormous amount of general knowledge — and no particular loyalty to your ship, your maker's manual, or your company's SMS. Asked something specific, a bare model will produce something specific-sounding. The industry calls this hallucination. A better word for the maritime context might be fabrication under confidence — and no amount of "better prompting" fully removes it, because it is how the underlying technology works.

The fix is architectural, not cosmetic

The approach we build on is retrieval-augmented generation — RAG. Instead of letting the model answer from its general training, the system first retrieves the relevant passages from your own documents — the maker's manual for your specific engine model, your SMS chapter, the relevant circular — and instructs the model to answer from those passages.

Then the part that matters most: the answer carries citations. Not "according to the manual" — the actual document, section, and paragraph, one click from the answer. Three things follow:

1. Verification takes seconds. The second engineer doesn't have to trust the AI. He taps the citation, sees the manual page, and confirms it himself. The copilot found the needle; the human checks it's the right needle.
2. "I don't know" becomes possible. If retrieval finds nothing relevant, a citation-first system can say so honestly instead of improvising. In our platform, an answer without a source is a refusal, not a guess.
3. Answers become auditable. When a decision is later reviewed — by the DPA, by an auditor, by class — there is a trail: this question, this answer, this source. That is language a compliance culture already speaks.

"But my crew won't check citations"

Some will, some won't — the same way some officers verify a colleague's recollection against the manual and some don't. The difference is that a cited answer makes verification a two-second habit instead of a twenty-minute search, and the culture can enforce it: act on cited answers; treat uncited claims as rumours. That is a trainable, auditable standard. An uncited chatbot offers no standard at all.

What to ask any AI vendor (including us)

• Does every answer show its exact source — document, section, paragraph?
• What happens when the system finds no source? (The only good answer: it says so.)
• Whose documents is it answering from — ours, or "the internet's"?
• Where does our data live, and who else's model learns from it?

If a vendor hesitates on any of these, keep your manuals closed.

Citations are not a feature we added to be thorough. They are the difference between an AI that performs knowledge and one your officers can act on. In this industry, that difference is the entire product.

See it on your own manuals — book a 45-minute demo, and send us a manual in advance. We'll answer from it, cited.