The EU AI Act and Liability for Shipping: What Applies, When, and Who Pays

Two sets of EU rules now decide what happens when AI on a ship goes wrong. The AI Act sets duties for the companies that build AI and for those that use it. The new Product Liability Directive decides who pays when software causes injury. Both reach shipping, and both have dates that sales people like to quote wrongly.

This guide explains, in plain words, what applies to a ship manager, when, and what to do about it now. It is not legal advice. Use it to ask your lawyer and your insurer better questions.

The dates that matter

The two high-risk dates moved with the EU's "Digital Omnibus", agreed in May 2026 and in force since 27 July 2026 as Regulation (EU) 2026/1744. These dates are now final. If a supplier tells you the AI Act "takes full effect this year", ask which article they mean. We explained the change in The EU AI Act deadline you're being sold against has moved, including a correction to our own earlier report.

Does the AI Act apply to my company?

Probably yes, in one of two roles.

• Provider: you build an AI system, or have one built and put your name on it.
• Deployer: you use an AI system in your business. This is what most ship managers are.

It applies if you are in the EU, or if the AI's output is used in the EU. A ship manager in Cyprus, Greece, Germany or Denmark is a deployer for every AI tool its office and ships use.

The Act sorts AI into four levels of risk. Most tools in shipping today sit in the lower two: weather routing, performance dashboards, chatbots, document readers. For those the main duty is transparency, plus helping staff understand the tools they use. That training duty ("AI literacy") has applied since February 2025. The Omnibus made it softer: companies must now take measures to support their staff's AI knowledge, rather than guarantee a level. It did not remove it.

What is banned already

Since February 2025, some uses of AI are simply not allowed. The one that matters most at sea: AI that tries to read workers' emotions from faces, voices or typing, except for narrow medical or safety reasons.

That matters because crew monitoring is a growing market. As Wolfgang Lehmacher wrote, and we reported in May, if crew think a system is scoring their mood, they learn to look calm instead of saying "I'm too tired for this watch". You get a clean dashboard and less safety.

The line in practice:

• Usually fine: detecting a clear safety signal, such as microsleep on watch, a fall, or a distress alarm, designed with the crew involved.
• Not fine: "crew sentiment", "attitude" or "engagement" scores used to rate people or decide contracts.

The fines for banned practices are the highest in the Act: up to €35 million or 7% of worldwide turnover. Most other breaches are capped lower, at €15 million or 3%.

Which shipping AI could be "high-risk"

High-risk systems carry the heavy duties: risk management, documentation, logs, human oversight and testing. For ship managers, two areas stand out.

1. AI that manages people. The Act lists AI used to recruit, select, assign tasks to, monitor or evaluate workers as high-risk. Crewing platforms that rank candidates, or tools that score seafarer performance, are the clearest example in our industry. As a deployer you would have to use them as instructed, keep their logs, have a trained person overseeing them, and tell workers' representatives before you start.

2. AI built into ship equipment. Navigation aids, engine control and autonomy software are mostly covered by their own product rules. The EU Marine Equipment Directive is one of the product laws listed in the AI Act, and the Act changes that directive so that EU standards and testing for wheelmarked equipment must take the AI requirements into account. In practice the requirements will reach you through type approval and class, not through a separate AI Act certificate, and the timing follows the marine equipment rules. (The August 2028 date applies to AI in other regulated products, such as machinery.)

Autonomous ships are a separate track. The IMO's MASS Code became available on a voluntary basis on 1 July 2026, as an "experience building phase" before any mandatory rules. See The rulebook switches on.

9 December 2026: who pays when software fails

This is the date most ship managers have missed. From 9 December 2026, the EU's new Product Liability Directive treats software as a product. We covered it in If software on board fails, who pays?.

What changes:

• If faulty navigation, engine control or autonomy software injures someone, the maker can be liable without anyone proving fault.
• A court can order the maker to hand over its evidence.
• There is no cap on the amount.
• A security hole can count as a defect.
• An update can count as a new product, so the date of the last big update may matter more than the date the system was installed.

Who it protects: people. Crew and passengers can claim for injury and death, and private owners for things like a passenger's car on a ferry. The shipowner cannot claim for the ship, lost hire or salvage under these rules. That still depends on your contract with the supplier and on your hull cover.

The catch: suppliers facing unlimited injury claims will try to push that risk back to you in their contracts. Read every software renewal with 9 December in mind.

What insurers are doing

Insurers have noticed. On 22 October 2026 the chief executive of Chubb said AI liability is a risk the industry has not yet priced. He expects clearer wording first, then exclusions that take AI losses out of normal cover, then separate AI policies. In Korea, Hanwha General Insurance and the cyber firm CYTUR plan cyber cover where the terms depend on how secure the ship is. And Lloyd's of London has backed policies from Armilla that pay out when an AI tool performs below an agreed benchmark.

The direction is clear: an AI or software loss that is covered today may not be covered after your next renewal unless you ask.

What to do now

You do not need a compliance programme to start. You need a list and some letters.

1. Make a list of the AI you use. Office and ships. Include the tools staff use on their own accounts; we explained why in Who pays for the AI your staff already use?
2. Mark anything that touches people decisions. Crew selection, performance scoring, monitoring. These are the likely high-risk ones.
3. Check for emotion reading. Ask every crew-monitoring supplier in writing whether their product infers mood, attitude or emotional state.
4. Tell people when they deal with AI. Chatbots for charterers, AI-drafted emails, automated replies.
5. Read software contracts before 9 December. Look for clauses where you must protect the supplier against injury claims, and for your right to decision logs and update records.
6. Write to your P&I club and hull insurers. Ask whether a loss linked to AI or software is covered, and get the answer in writing.
7. Train the people who use the tools. The AI Act expects it, and it is good practice anyway. Short and practical is enough: what the tool does, where it fails, who to tell.

Questions to ask your AI supplier

• Under the AI Act, are you the provider, and which risk level do you put this system in? Why?
• Does it infer emotions, mood or attitude from crew in any way?
• What documentation will you give us as deployer: instructions, limits, human oversight measures?
• Does the system keep logs, how long, and can we get them?
• When you update the model, will you tell us, and what changes?
• After 9 December, who carries the risk if your software injures someone? Show me the clause.
• Who has tested this system apart from you, and who paid for that test?

That last question matters more than it looks. As we wrote in Who pays the inspector?, shipping has 150 years of experience with checks that the checked party pays for, and built flag, port State control and insurers around them. AI has only the paid layer so far.

Common questions

Does the EU AI Act apply to ships outside EU waters? It applies to companies based in the EU, and to AI whose output is used in the EU. For an EU-based ship manager, where the ship sails does not change much.

Is weather routing or performance software high-risk? Usually not. It is likely to be in the lower risk levels, with transparency duties. AI that controls the ship or makes decisions about people is where the high-risk rules start.

Is the AI Act being enforced yet? Yes, in part. On 14 October 2026 the European Commission said its AI Office "has started enforcing", with formal requests to large AI developers. See AI agents got out of their test boxes. Shipping companies are not the first target, but the bans already apply.

Can a shipowner use the new liability rules to claim for a damaged ship? No. The Product Liability Directive is for injury, death and private property. Damage to the ship is still a matter for your contract and your insurance.

The ship-side view

The AI Act is not the immediate risk for most ship managers. Its heavy rules are a year or more away, and most shipping AI sits in the lower levels. The immediate risks are two: AI that reads crew emotions, which is already banned, and 9 December, when the people on board gain strong rights against software makers and the owner gains none. Make the list, read the contracts, write to the insurers. That is a week's work, and it is worth doing before the year ends.

We update this guide when the rules change. The latest stories on the EU AI Act and liability from our weekly digest are listed below.