ARTICLES ·
Why does an AI assistant get my boat wrong?
A general AI assistant knows marine systems in general and nothing about what is fitted on your boat. Asked a specific question, it fills the gap with a plausible guess. Give it a record of your own yacht, with its sources and its unknowns marked, and it can answer from that and say when it does not know.
What goes wrong
One owner on a boating forum asked an assistant for the fuel filter part number for his 1990 boat. He asked three times and received three different part numbers. Others in the same discussion found assistants a useful place to start, and some were uneasy about relying on them for anything that mattered. These are owners’ accounts, not a study. The first shows the problem: three confident answers and no way to tell which is right.
The companies that build the assistants describe the same thing. Anthropic’s guidance for its Claude models says that even advanced models can produce text that is factually wrong or inconsistent with what they were given. Researchers at OpenAI argue that models guess because the way they are trained and tested rewards a guess over an admission of uncertainty.
Why a boat is a hard case
- It has never seen your boat. The builder, model and year say little about what is aboard after years of ownership, refits and repairs.
- Manuals describe equipment, not your installation. A manual for the pump that was fitted at launch is no help if that pump was replaced.
- It is built to answer. The OpenAI researchers’ point applies here: unless it is told otherwise, a model is more likely to offer a plausible answer than to say it does not know.
What the AI companies recommend
The advice from the companies that build these assistants is consistent.
- Give it the documents, and tell it to use them. Anthropic suggests telling the model to work only from the documents provided and not from its general knowledge. Google’s guidance for Gemini gives a similar instruction: rely only on the facts supplied, and say so when the answer is not there.
- Allow it to say it does not know. Anthropic’s guidance is to give the model explicit permission to admit uncertainty, and says this can greatly reduce false information.
- Ask for the source of each claim. The same guidance suggests having the model support each claim from the text it was given, and withdraw any claim it cannot support.
None of this makes an assistant infallible. Anthropic says plainly that these techniques reduce wrong answers and do not remove them. A 2024 Stanford study of legal research tools that answer from documents found that two of them gave incorrect information more than 17% of the time and a third more than 34%, while still making fewer errors than general chatbots. Working from documents is an improvement, not a guarantee.
What that means for your yacht
An assistant can only work from documents if the documents exist in a form it can use. For a boat, that means one record which says:
- what is fitted, system by system;
- where each fact comes from: a numbered paper, your own word, or not yet confirmed;
- what was replaced, kept as history so the old manual is not mistaken for the new one;
- what has happened: work, changes and what you have noticed;
- what is still unknown.
The pack made from the record, which is the file you load into the assistant, carries instructions on the same lines: answer from the record first, name the source, say when the record does not hold the answer, label anything from general knowledge as general, and treat what an owner has noticed as an observation and not a diagnosis.
We tried this with Northlight, a fictional yacht whose record shows that the domestic water pump was replaced and that the new pump’s make and model were never written down. We gave the pack made from her record, and nothing else, to one AI model, Anthropic’s Claude Sonnet, and asked three questions an owner might ask. One was the pump’s service interval according to its manual. It declined to use the original pump’s manual, said the record did not identify the new pump, and asked for its make and model. That is one test with one model, not a study. It is what the pack’s instructions ask for.
What a record does not do
It does not turn an assistant into a marine engineer, and it is not navigation or safety advice. It is meant to make the assistant’s answers traceable to your own papers and plain about the gaps. Anything that matters should still be checked against the equipment itself and, where it calls for one, a specialist.
Questions
- Can I just upload my manuals?
- It helps. Assistants such as Claude let you add files to a project and use them as background for your questions. But a manual says how a piece of equipment works, not whether it is the one fitted to your boat today. The record is what tells the assistant which manual applies.
- Does a record stop wrong answers?
- No. The AI companies’ own guidance is that working from supplied documents reduces wrong answers without removing them. A record should also make a wrong answer easier to catch, because each answer is asked to name the paper it rests on.
- Which assistants does this work with?
- The pack is a plain text file with its own instructions, made for assistants that accept files. So far we have tried it with Claude only.
Sources
- Anthropic, “Reduce hallucinations”: the guidance on working from supplied documents, admitting uncertainty and supporting each claim.
- Google AI Studio, “Prompting strategies”: the instruction to rely only on supplied facts.
- Kalai, Nachum, Vempala and Zhang, “Why Language Models Hallucinate”, OpenAI, 2025.
- Stanford HAI and RegLab, on legal research tools that answer from documents, 2024 (a preprint, on legal questions).
- Anthropic, “What are projects?”: how uploaded files are used.
- iBoats forum, “Artificial Intelligence”, June 2026: owners’ own accounts.