Gains:
- Ability to scan and summarize class rule, SOLAS/MARPOL and survey requirements in a structured format with AI
- Ability to produce survey reports, technical documents and compliance files in draft form with AI support
- Ability to verify every rule clause, date and requirement given by AI from official and up-to-date source
A ship also exists on paper: class rules, contract surveys, flag state legislation, certificates, technical manuals and inspection reports. This world of documents establishes the legal and technical legitimacy of the ship; One missing certificate or a misinterpreted rule will result in the ship being detained in port (Port State Control detention) or its insurance invalidated. Class rules (class rules; technical requirements of the classification society regarding the design, construction and maintenance of the ship) and legislation texts are long, intertwined and constantly updated documents. Artificial intelligence (AI) makes it faster to summarize, compare and draft reports. But maritime legislation is one of the areas where AI is most hallucinatory.
Absolute rule: No rule clause, clause number, limit value or the phrase "that SOLAS/MARPOL/class requirement is" given by AI shall be used without being confirmed verbatim with the current and official text of the relevant document. AI can make up item numbers and believable but false values that do not actually exist; In legislation, this has serious consequences in terms of legal and security.
Concepts: Classification society: The organization that inspects and certifies the ship's compliance with the rules (ABS, DNV, LR, BV, Türk Loydu, etc.). SOLAS: International convention for the safety of life at sea. Survey: Periodic inspection of the ship's compliance with the rules. Port State Control: Inspection of foreign ships by the port state. Certificate: Official document certifying a certain conformity. Hallucination: AI fabricating non-existent matter/value as real.
Why Is Regulation a Dangerous Area for AI?
Large language models produce the “most probable continuation” of the text; It is not linked to a definitive source. This gives rise to three traps in maritime legislation. First, spurious attribution: AI can produce a reference that looks realistic but is false or non-existent, such as “SOLAS Chapter II-1, Rule 8.3.” Second, legacy value: rules are updated frequently; The threshold in the AI's memory may have been superseded. Third, context confusion: one may speak of a rule valid for one ship type/size as if applying it to another ship for which it is invalid.
So use AI as a “search and draft” tool in legislation, not as an “authority”. AI can point you to which document you need to look at, which section, summarize a long text, sketch the difference between two versions; but every concrete item, value and reference is verified by reading it from the original current text. The safest use is to give the relevant rule text as context to the AI (if possible) and instruct "only tell me if its source is unclear, based on the text I give."
Caution: The believable form of an item number given by the AI is not proof of its accuracy. You can't tell an inspector or class "AI said so"; Each reference must be based on the relevant page of the official document in force. Any substance whose source cannot be shown is not included in the report.
Survey and Technical Report Drafting
AI is really useful in improving the structure, language and consistency of survey and technical reports: turning a list of inspection findings into an organized report, reminding of missing headings, ensuring the same finding is written consistently in different places. However, each technical value of the report (measured thickness, detected corrosion rate, limit value) must depend on the measurement and the source. While AI produces a smooth report, it tends to fill in an unmeasured value as “reasonable”; This is unacceptable in the survey report.
A good workflow is this: you provide the actual measurements and findings, the AI structures and verbalizes them, then you match each value with its source (meter, date, location) and extract each statement of uncertain origin. The report is a legal and technical document; The responsibility lies with the signing surveyor/engineer.
Quest
Safe use of AI
Risky/prohibited use
Rule summary
Give the text and summarize
Requesting an item/value from memory
Finding citation
"Which document should I look at?"
Using the item number without verifying it
Version difference
Compare two texts
Assuming current version
Report draft
Structuring the findings
make up an unmeasured value
Certificate list
Reminder of required documents
Not verifying the validity date
Mini Cases
Case 1 — Made-up SOLAS clause. An engineer asks AI to support a safety equipment requirement; AI gives an exact value with a specific "SOLAS Chapter/Rule number". When the engineer looks at the current SOLAS consolidated text, he sees that that article number belongs to a different subject and the value given is not included in the text. The correct requirement is in another rule and a different value. Lesson: every legislative reference is verified verbatim from the original current text.
Case 2 — Stale threshold. A team writes an environmental requirement into a report with an old limit value given by the AI. That value was tightened a few years ago; The ship is behind the current limit. The report contains an incorrect "suitable" statement. The error is caught when compared to the current MARPOL text. Lesson: regulatory boundaries are updated; The value in the AI's memory may have become obsolete.
Case 3 — Contrived measurement. A draft thickness gauging report prepared by AI includes thickness values for some points that appear "reasonable" but are not actually measured. Once the surveyor matches each value with the field measurement record, he finds and deletes lines that have no source. Lesson: every technical value in the report should be based on actual measurement; fluent text is not data.
Copiable Prompt Templates
Template 1 — Rule text summary (depending on source):
Role: You are the maritime legislation reading assistant.Context: I am pasting the relevant rule text below.Task: Based ONLY on the text I provided:1) Summarize the requirement of this article in plain language.2) If there is a numerical limit/value, quote it verbatim and with its unit.3) Do not add anything that is not in the text; If it is unclear, say "not in the text". Restriction: DO NOT make up the item number or value from your own memory.
Template 2 — Citation verification plan:
Point out which official documents and sections I should look at when trying to answer the following regulatory question: [question]. 1) Possibly relevant document(s) and section headings. 2) "Verify from original text" note for each. Constraint: Exact article number and value; I will read these from the official text. Mark "unclear" where you are not sure.
Template 3 — Survey report configuration:
Convert the following raw inspection findings into an organized survey report:[findings and actual measurements].1) Structure with standard headings; mark missing headings.2) Match each technical value with the measurement I give.Constraint: DO NOT make up any value I have not measured or fill it in as "reasonable"; mark every statement of uncertain source with the [VERIFY] tag.
Template 4 — Regulatory output hallucination control:
Table all references and numerical values in the following legislation/rule response:[paste response]Each row: (citation/value) | (based on: given text/assumption/uncertain)| (how do I verify it). Highlight any line whose basis is "uncertain" or "memory" and state that it should not be used without verification from the official source.
Weak prompt / Strong prompt
Weak prompt:
What is the requirement for this equipment according to SOLAS? Say it with the item number.
Powerful prompt:
Role: You are the legislation reading assistant. Context: I am pasting the current text of the relevant SOLAS section below. Task: Summarize the equipment requirement based on this text only and quote it verbatim with its numerical value unit, if any. Constraint: DO NOT make up the item number or value that is not in the text; If it is not in the text, say "it is not in the text you provided". I will separately verify each reference against the official consolidated statement.
Weak prompt assumes the AI is the authority and asks for items from memory; Strong prompt text gives context and prohibits fabrication.
Common mistakes
- Mistaking AI for regulatory authority. Item numbers and values may be fictitious; AI is a search/draft tool, not a source.
- Falling for the citation style. A convincing “Chapter/Rule” format is not proof of accuracy.
- Assuming timeliness. The rules are updated frequently; The threshold in memory may have become obsolete.
- Making up a value in the report. Unmeasured technical values should not enter the report silently in flowing text.
- Skipping ship type/size context. A rule valid for one ship may be invalid for another; coverage should be checked.
In summary
Classification rules, legislation and survey documents establish the legal-technical legitimacy of the ship, and AI makes it faster to summarize, compare and structure these texts. But legislation is the area where AI hallucinates the most: every clause, reference and threshold is verified verbatim from the current official text; Every technical value in the reports is linked to the actual measurement. AI is a search and drafting tool, not a legal-technical authority; The responsibility lies with the signing engineer and surveyor.
Application task
Select a piece of rule text (a representative or publicly available example) and have the AI summarize it based solely on the text you provide with the “rule text summary (source dependent)” template. Then ask the AI a memorized item number on the same topic and test the output with the “regulatory output hallucination audit” template; Report how many citations are unverifiable. Finally, construct a raw survey findings list and mark any values of uncertain origin.
checklist
- [ ] I have verified each legislation/class reference verbatim from the current official text.
- [ ] I used AI as a search and draft tool, not an authority.
- [ ] I confirmed that the rule limit values are up to date.
- [ ] I linked every technical value in the report to the actual measurement/source.
- [ ] I have checked that the ship type/length scope refers to the correct rule.
- [ ] I left the final report and declaration of conformity with the surveyor/engineer who signed it.