AI for IMO Compliance: Where Assistance Stops and Company Responsibility Begins
AI can help a maritime company read, draft, compare and explain. It cannot hold an obligation, and it cannot answer for the evidence.
"AI for IMO compliance" covers several different tasks. The International Maritime Organization develops instruments addressing safety, pollution prevention, seafarer training and watchkeeping, security and ship management. Their duties and implementation depend on the instrument and applicable national law. The Maritime Labour Convention, 2006 sits alongside that framework, but it is an International Labour Organization convention, not an IMO instrument.
An AI tool can help explain requirements or compare supplied evidence. Its answer is only as reliable as the sources, vessel information and controls behind it. A company should not treat a fluent response as a verified applicability decision. Using a tool does not transfer the responsibilities of the company, master, flag Administration or other duty holders to the model.
This article sets out where the assistance is genuinely valuable, where it creates exposure, and what a controlled use case looks like for a maritime team.
First, which obligation are we talking about?
IMO compliance is not a single standard. The ISM Code provides a management-system framework. SOLAS addresses ship safety, MARPOL pollution prevention and STCW training, certification and watchkeeping. The ISPS Code operates under SOLAS chapter XI-2 for maritime security. The ILO's MLC, 2006 addresses seafarers' working and living conditions. These instruments need separate applicability checks rather than one undifferentiated compliance score.
Those instruments interact in practice. A MARPOL deficiency may also expose an ISM weakness. An hours-of-rest problem may engage both MLC and STCW requirements. Security and safety processes must work together, but a ship security plan has its own approval and confidentiality requirements; it is not simply an unrestricted SMS document to upload into an AI service.
A broad instruction such as "check our compliance with IMO requirements" leaves essential inputs unstated: vessel type, flag, tonnage, trade, certificates, exemptions and applicable amendment dates. A system with verified retrieval and structured vessel data can help identify relevant provisions. The reviewer still needs to establish that the source, version and applicability logic are correct.
Four uses that add real value
1. Explaining and drafting internal guidance
A model can turn a primary source into a plain-language explanation, suggest a structure for an internal briefing, or draft a first version of a procedure for review. The output is a draft. The reviewer must check every regulatory statement against the current text, confirm that the interpretation matches the flag State's position and adapt the language to the vessel.
2. Preparing for gap analysis
Comparing an existing procedure set against a list of requirements is slow, repetitive work. AI can help produce a first-pass comparison, highlight apparent gaps and generate questions for the reviewer. The value is in the questions. A generated gap list is a prompt for investigation, not a finding of non-compliance.
3. Clustering findings and themes
Inspection reports, audit findings, near misses and complaints contain patterns that are hard to see across hundreds of records. A model can group them by theme, produce a timeline or identify recurring language. This is genuinely useful for management review, provided the clustering is checked against the original records and the reviewer does not treat a theme as a proven cause.
4. Translation and crew briefing
Mixed-language crews are a reality of ship management. AI can produce a first translation of a procedure or briefing for review by a competent speaker. That can improve understanding, but the translation must be checked for technical accuracy - particularly where a term has a specific operational meaning - and the controlled document remains the authoritative version.
Four uses that create exposure
1. Producing the compliance conclusion
A generated statement that a vessel "complies with" an instrument is not sufficient evidence. A tool may support analysis, but cannot issue statutory approval or replace the authorised body's judgement. An internal recommendation also needs a competent, accountable reviewer and a traceable evidence basis.
2. Creating records of activity that did not happen
If a model invents a drill, maintenance activity or inspection observation and the output is filed as a factual record, the record is false. A human pressing save does not cure that. This differs from controlled transcription or summarisation of real source records: those uses can be acceptable when the original evidence is preserved, accuracy is checked and the applicable attribution, signing and approval process is followed. Missing facts must remain missing until established.
3. Asserting flag, class or port State expectations without verification
Interpretation varies. A flag State may issue guidance that goes beyond the convention text; a class society may apply its own rules; a port State may take a particular view in practice. A model trained on general material may present one interpretation as settled. The reviewer has to check the applicable source, and the answer may depend on the ship's certificates and trading pattern.
4. Handling information without controls
Incident reports, complaints, medical information, crew data, commercial terms and investigation material can all be sensitive. Pasting them into a consumer AI tool may disclose more than the task requires and may breach contractual or legal obligations. The question is not only whether the tool is approved, but whether the specific material is permitted for that task. The free information-handling guide sets out a five-step decision process for exactly this situation.
The amendment problem
IMO instruments change. Amendments enter into force on defined dates, with entry-into-force and application arrangements that differ between ships and flags. A model's training data has a cut-off, and even a current model may mix versions of a requirement, quote a superseded figure or miss a recent amendment.
For any regulatory statement that will influence a decision, the review should include the primary source: the convention text as amended, the relevant IMO circular, the flag State's guidance and, where applicable, class requirements. If the tool cannot show where a statement came from, the statement is not yet usable. Citations are not a formality; they are the difference between a draft and a claim.
The same discipline applies to the vessel's own documents. AI summarising a procedure is summarising what the company wrote, which may itself be out of date. The controlled document, not the summary, is the reference point.
What a controlled use case looks like
The following is a suggested company control pattern, not a new universal IMO rule for AI. Adapt it to the SMS, applicable law, flag guidance and the risk of the task.
- Define one task. "Prepare a first draft of the enclosed space entry briefing for review" is a task. "Make us IMO compliant" is not.
- Name the inputs. Which sources may be used, and which material must not leave the company's systems. Record the decision.
- Name the reviewer. A competent person with the authority to reject the output. The reviewer's name belongs in the record.
- Define the output's status. It is a draft until reviewed. It does not enter the controlled document set, the official record or a regulatory submission without the normal approval route.
- Record what was checked. Which statements were verified, against which source, and what was changed. This is the evidence that the human review was real.
- Set stop conditions. Pause when a material claim lacks a verifiable source, the inputs fall outside approved information controls, or the output exceeds the reviewer's authority. Escalate the unresolved point rather than asking the model to sound more certain.
- Review the use case. After a defined period, decide whether to continue, change or stop. That decision should be recorded like any other management decision.
For related support, the AI Governance And Confidentiality Kit covers policy, tool approvals and a 90-day pilot. The SMS and ISM AI System supports integrating AI controls into an existing company SMS; it is not a complete SMS or a substitute for company review and approval. These current resources have different scopes from the earlier readiness and transition offers. Confirm scope and availability through their enquiry listings.
A small use-case test before wider rollout
Illustrative internal workflow, not a regulatory approval checklist. Suppose a team wants AI to summarise audit findings for management review. Test the following before adding more vessels or record types.
| Test input | Acceptable handling | Reason to stop and investigate |
|---|---|---|
| A finding with an open corrective action | Keep its source reference and open status visible | The summary treats a planned action as completed or effective |
| Two reports describing similar events | Suggest a possible theme and link both reports | The model asserts a common cause that neither investigation established |
| A missing date or unreadable quantity | Flag the missing or uncertain value for review | The output supplies a plausible date or number |
| A regulatory reference from an older report | Check it against the applicable current text before relying on it | A citation exists but does not support the statement or applies to a different ship |
Record the source set, intended use, reviewer, corrections and release decision. This gives management a basis to judge whether the time saved is worth the verification effort. A successful demonstration is a reason for a bounded trial, not evidence that every future output will be reliable.
Accountability does not delegate
The ISM Code allocates company responsibilities, the master's authority and the designated person's role. Audit and management review obligations are not displaced because software drafted a document. Keeping a record of how a consequential AI output was checked is a practical way to support those existing responsibilities. It should not be described as a separate blanket IMO requirement imposed on every use of AI.
For management, the immediate risk is a plausible output entering a decision or record without anyone checking it. Controls need to make the source and review visible, and give the reviewer enough time, competence and authority to reject the result.
For the relationship between AI and the safety management system specifically, the companion article on why AI policies fail explains why an approved tool list is not a control. ISM compliance software sets out the evidence a mapped clause still has to produce, and the analysis of false assurance covers the wider pattern of records that look complete without demonstrating control.
Conclusion
AI can assist reading, drafting, comparison and organisation. It can also support applicability analysis or preparation of a record when the inputs and workflow are controlled. It should not be the sole authority for a compliance conclusion, and it must not invent the evidence on which a decision depends.
For each use case, ask who will notice a material error, what evidence they will use and who may approve the result. If those questions cannot be answered, the use case needs more work before operational adoption.
AI can help a company prepare for an obligation. It cannot become the company that holds it.
References
- IMO: ISM Code and amendment history. Relevant areas include sections 1.2, 4, 5, 9, 11 and 12.
- IMO: SOLAS, 1974, as amended, including chapter IX on safe management.
- IMO: STCW Convention, covering training, certification and watchkeeping standards.
- IMO: SOLAS XI-2 and the ISPS Code, including mandatory Part A and recommendatory Part B.
- ILO: MLC, 2006, text and amendments. Regulation 5.1.1 sets general principles; 5.1.3 addresses certification and the declaration of maritime labour compliance; 5.1.4 addresses flag-State inspection and enforcement. Port-State responsibilities are addressed separately in 5.2.