One idea runs through almost every story this week. The screen shows you a number, and the number is not the thing. Four tankers in the Black Sea told the world they were in Peru. A ship system with 20 known software flaws turned out nearly as risky as one with 200. A port planning agent says operators accept 90% of its plans, and nobody outside the company has checked. A free freight forecast gives risk scores without saying how often it was right before. And the Nautical Institute is rewriting the book that once told investigators electronic records do not lie. None of this means the tools are bad. It means the job has changed. The skill is no longer reading the screen. It is knowing when to doubt it. That is exactly what one training chief asked the IMO to write into the rules for every officer at sea.

The week's most important developments in shipping & oceans — distilled into a 5-minute read.

🔗 Quick Links

• 🏗️ Huawei turns China's port AI into products it can sell — Splash247

  Live at Shandong and Tianjin, and Huawei says it serves over 100 ports. It claims 90% of machine-made plans are accepted. Company figures, not checked by anyone outside

• 📡 Four Black Sea tankers "report" from Lima — Splash247

  27 impossible AIS positions picked up in Bulgaria. The startup that caught it says normal data cleaning would have deleted the evidence

• 🔐 CYTUR: 200 software flaws or 20, the risk came out almost the same — Smart Maritime Network

  A small sample from its own clients, and CYTUR says so. The risk often sits in parts delivered in the same box as the main equipment

• 🗼 Germany starts phase two of AI support for vessel traffic controllers — Splash247

  After three years and €3m, a demonstrator is due in July 2028. Ships are getting AI faster than the shore stations that watch them

• 🎓 A call to make AI literacy an STCW competence — Seatrade Maritime

  The IMO review of seafarer training closes in 2029 to 2030. Teach officers when to overrule the machine. Written by the head of a training company

• 📦 WiseTech gives away a four-week freight capacity forecast — Digital Ship

  Three main trades, monthly risk scores. No accuracy record published, and the free version leads to a paid one

• 🧾 AiatSea: the records that "cannot lie" just put four tankers in Lima — AI at Sea

  The Nautical Institute rewrites its evidence guide for the AI age, and a P&I club puts a second position source on offer. Three things to change this week

🚀 Big Moves This Week

1. Huawei Turns China's Port AI Into Products, With Numbers Nobody Outside Has Checked

On 16 September, at its Huawei Connect event in Shanghai, Huawei launched two port products built around AI agents: a Port Intelligent Planning Solution and an Intelligent Security Solution. The planning product joins pilotage, berth allocation, quay crane deployment, yard planning and vessel stowage in one scheduling layer. At the centre is what Huawei calls an all-factor scheduling agent. Its job is to turn the terminal's rules and the dispatchers' experience into plans made by the machine.

What Huawei claims. Planning time falls from hours to minutes. More than 90% of machine-made plans are accepted by operators. The planning horizon grows from three days to as much as one month. The security product uses a vision model trained on more than one million port data points to watch people, vehicles and cargo, including from cameras mounted over 20 metres high, and Huawei puts its accuracy above 95%.

This comes from real ports, not a lab. At Qingdao, part of Shandong Port Group, the port has listed 80 AI use cases. Its berth-planning agent weighs 132 parameters and produces a berthing plan in about 30 seconds. At Tianjin, roadside sensors and cloud scheduling coordinate up to 300 driverless and normal trucks in the same yard, without keeping them apart, and Huawei says human intervention on the driverless trucks is below 0.1%. Huawei says it now serves more than 100 ports worldwide, and the Port of Douala in Cameroon is working with it on an AI data layer.

The honest part. Splash247, which reported the launch, says clearly that the performance figures are company data and have not been independently verified. And "90% of plans accepted" needs a second question. Accepted as the machine wrote them, or accepted after a planner fixed them? A plan the operator accepts is not the same as a ship that left the berth earlier.

Why a ship manager should care. Last week China Merchants launched a vessel operating system. This week Huawei packaged its port projects into products it can sell to other terminals. The pattern is clear: Chinese groups are turning single projects into complete systems for export, for ships and for ports. If a terminal you call at buys this, your berth window, your crane allocation and even your stowage may be set by an agent. The question to put to your agent or the terminal is simple: when the machine sets our berth window, can we see why, and who do we call when it is wrong? [Splash247 · Huawei · Shandong showcase]

2. Four Tankers "in Lima": The Position Everyone Throws Away Is the Evidence

Last Sunday night, a receiver in Varna, Bulgaria picked up 27 position reports from four tankers in the Black Sea. Every one of them placed the ships on the same one-kilometre circle in central Lima, Peru, more than 12,000 km away. The alert came from Worldwide AIS Network, a startup based in Copenhagen.

Why this is certain, in plain words. AIS travels by VHF radio, and VHF only reaches a limited distance. A receiver in Varna can only hear ships near Varna. So when it hears a ship saying it is in Peru, the message is physically impossible, whatever position is written inside it. Nobody has to guess who is lying. The physics settles it.

Bigger than one night. Between 2 and 9 September the company counted 1,702 impossible position jumps from 379 vessels in the Gulf of Oman. At midnight on 9 September, almost one third of position reports from that area arrived with the ship's positioning system marked as not working. It has also seen interference in the Baltic.

The point that made us run this story. Most AIS data services clean their data. Positions that are obviously wrong are removed later as outliers, because they spoil maps and analytics. The company's argument is that this cleaning throws away the proof that a ship's satellite navigation was jammed or spoofed. The "error" is the record of the attack.

Honest limits. All the numbers are the company's own. The ships are not named, and no customer is named. The company says it is building its receiver network for insurers, maritime authorities and defence customers. Its chief executive says rivals analyse interference after the event while his company catches it live. That is a sales claim until somebody tests it.

What to ask on Monday. Voyage optimisation, arrival-time prediction, sanctions screening and emissions checks all use AIS data. Ask your AIS or voyage-data provider one question: do you delete impossible positions, or do you flag them and keep them? If they delete them, the record of a spoofing event around your ship may be gone before you know you need it. [Splash247]

3. Two Hundred Software Flaws or Twenty, the Risk Came Out Almost the Same

Korean cybersecurity company CYTUR looked back at more than a dozen cyber risk assessments of ship equipment that it carried out for clients. It rated 94% of the systems as needing cyber risk treatment under IACS Recommendation 171, the guidance on bringing cyber risk into the safety management system. About 60% were rated "Required", which means treatment is mandatory under that framework.

Read the limit first, because CYTUR wrote it. The 94% does not mean 94% of ship equipment is at risk. The systems were chosen and paid for by CYTUR's own clients, so this is not a random sample. "More than a dozen" is also a small number. And CYTUR sells these assessments. So this is an interesting finding from a small, self-selected sample, by a company with a business interest. We are running it because of what it found inside the sample, not because of the headline number.

The finding that matters. Many buyers judge equipment by counting CVEs, the publicly listed software flaws. CYTUR found that count misleads. One integrated automation system with more than 200 known flaws scored a risk level of 20. Another with about 20 flaws scored 16. Both needed treatment. Ten times fewer flaws, almost the same risk. What counts is whether a flaw can actually be reached, whether it leads to other systems, and what happens to the ship if someone uses it. CYTUR looks at four things together: vulnerability, exposure, connectivity and impact.

Where the risk hides. In one case the maker's main controller had no known flaws at all, while a network device installed next to it had 13. Other flaws sat in PLCs, operator screens, input and output modules and firewall software delivered with the equipment. One type of equipment talked to the ship over a serial cable, yet had eight open service ports on an Ethernet connection nobody used, and CYTUR found the same setup on equipment from two different makers. It also found unsupported operating systems, and security patches that had been available for years but were never installed.

The good news. One system dropped from risk 20 to 12, from "Required" to "Appropriate", after software and firmware updates, switching off services nobody needed, encrypting management traffic and checking the network setup. Ordinary work, not magic.

Why ship managers should care. On 6 September we wrote about tools for tracking supplier cyber certificates under IACS UR E27. This study shows why the certificate for the main unit is not the whole story. The risk often sits in the parts that arrive in the same box. When you buy or refit, ask the maker for a list of every software component delivered with the system, including third-party network devices, and ask who will patch them, and how often. [Smart Maritime Network · Digital Ship · Safety4Sea · AiatSea, 6 September]

4. Ships Are Getting AI Faster Than the Shore That Watches Them

Germany's Fraunhofer FKIE has started a research project called LEAS-PRO. The goal is a demonstrator: an AI system that reads developing traffic situations and gives recommendations to vessel traffic service (VTS) operators, the people ashore who watch and guide ships in busy coastal waters. It will combine traffic data, ship movements and context to help operators spot risk in complicated encounters. Fraunhofer says the aim is not to replace controllers, but to reduce their mental workload as traffic becomes a mix of normal, highly automated and autonomous ships. The project runs until July 2028.

This is phase two. LEAS-PRO follows an earlier project called LEAS, funded with €3 million by the German research ministry from January 2022 to December 2024. Seven partners took part, including Fraunhofer CML, two institutes of the German Aerospace Center (DLR), Wismar University, Jakota Cruise Systems and Bergmann Marine. The work used a VTS simulator run by the federal waterways administration. When it started, Fraunhofer CML wrote that apart from some simulation studies, there was no practical experience of how VTS should handle mixed traffic.

The honest reading. Three years and €3 million produced a concept and simulator tests. The next step, a demonstrator, is due in mid-2028. That is not a criticism. Shore systems must be careful, because a wrong recommendation from VTS affects every ship in the area. But it shows the gap clearly. AI decision support and assisted navigation are arriving on bridges now, while the shore stations that watch those bridges are still in research. One more note, for transparency: FKIE describes itself as one of the largest defence research institutes in the Fraunhofer group. This project is civil, and its first phase was funded by the research ministry.

Why this matters to you. Every ship meets VTS. If your vessel uses AI-assisted navigation, the operator ashore will not know it unless someone tells them. Today there is no standard way for a ship to tell the shore that it is being steered with machine help. One part of the first project is worth noting: Wismar University tested whether AI could predict how a ship will manoeuvre, faster and more reliably, for the VTS operator. That only works if the shore knows what kind of control is on the bridge. A question for your next autonomy or assisted-navigation project: what does VTS see, and who tells them? [Splash247 · Fraunhofer FKIE · LEAS · Fraunhofer CML · LEAS]

5. Should Every Officer Be Trained to Doubt the Machine?

The IMO began a full review of STCW, the convention that sets training and certificate standards for seafarers, in 2024. The first phase found more than 400 possible gaps. Changes are now being written, with completion targeted for 2029 to 2030. In an article for Seatrade this week, Capt. Pradeep Chawla, chief executive of training company MarinePALS and chairman of GlobalMET, argued that the review should make AI literacy a formal STCW competence.

What he means, in plain words. Not programming. Not how to write a good prompt. He means seafarers should know that AI can sound certain and still be wrong. They should be able to spot invented facts, bias, missing information and data-security risks. Most of all, they should know when their own judgement, and what they see from the bridge, must overrule the machine. His example is a voyage recommendation built from weather, traffic and fuel data. Before following it, the officer should check that the data is current, that the ship's real limits were included, and that the advice matches what he or she can see outside.

Two levels. Everyone would get basic awareness. Officers, and anyone responsible for AI-enabled systems, would get deeper skills: checking results, knowing limits, escalating and overriding. He says the standard should not be tied to any product, because products change quickly. And the exam should test behaviour, not definitions. Give the candidate a plan with a false reference, missing data, or an efficient route that breaks a safety limit, and see whether they catch it. He also wants crews to know what must never be typed into an unapproved tool: personal data, commercial data, ship vulnerabilities and protected records.

The honest part. This is an opinion piece, and the author runs a company that sells maritime training. A new STCW competence would be good for his business. That does not make him wrong. The argument is sound, and he says openly that he is an early adopter of these tools himself. A second voice this week pointed the same way. An analysis in Offshore Magazine argued that people use AI with more confidence when they know which decisions stay theirs, and that every AI system should have a written limit on what it may do on its own.

What you can do before 2030. You do not need to wait for the IMO. Your company can write down two things now: which decisions an AI tool may support, and which decisions always stay with the officer. Then add one scenario to your next onboard drill or familiarisation where the tool is wrong on purpose. If your officers have never seen an AI tool fail, they will not recognise the first real failure. [Seatrade Maritime · Offshore Magazine]

6. A Free Four-Week Freight Forecast, and the Number It Does Not Show

WiseTech Global, the company behind the CargoWise logistics software, launched the Ocean Freight Risk Outlook on 17 September. It is a free dashboard that looks four weeks ahead at ocean freight capacity on three main trades: Asia to North America, Asia to Europe, and Europe to North America. It is updated monthly and gives risk scores for capacity, reliability and carrier risk. The idea is to show where space will get tight before it shows up in bookings or prices.

How it works, as far as WiseTech says. It compares expected cargo demand with planned carrier capacity, using combined supply chain data and WiseTech's analysis of the work that flows through CargoWise, a system used mainly by freight forwarders. The first edition said risk will stay high on Asia to North America and Asia to Europe, with falling schedule reliability as the main worry.

What is missing. There is no published accuracy figure: no record of how often past forecasts would have been right. The method is not described, and WiseTech does not call it AI in its announcement. The data comes mostly from the forwarder side, not from shipowners or operators. And the free version is a summary of an eight-week outlook that paying CargoWise customers buy, so it is also a door to a sale.

Why we still cover it. In August we reported on a review of 28 research studies on machine learning for freight forecasting. The conclusion was that these tools get better at the predictable part of the market and stay blind to sudden shocks. A four-week view is exactly where that matters. It will probably do well in a calm month and miss the next canal closure or strike. That is not a reason to ignore it. It is a reason to treat it as one input, not the answer.

For owners and ship managers. Charterers and their forwarders will start bringing dashboards like this into negotiations. Whenever a forecast is used against you, ask two questions: how often was it right last year, and whose data is inside it? [Digital Ship · The Loadstar · AiatSea, 16 August]

7. AiatSea: The Records That "Cannot Lie" Just Put Four Tankers in Lima

In 2019 the Nautical Institute published the second volume of its Guidelines for Collecting Maritime Evidence, about electronic evidence from ECDIS, VDR and AIS. It carried the view of an Admiralty judge, Mr Justice Teare, that "electronic or digital records cannot lie". Unlike a witness, they have no faulty memory, so they are the best evidence of what happened.

He was right about memory. A voyage data recorder does not forget, and it does not protect its own career. But this week showed the other half. A record can store a lie perfectly. The four tankers in story 2 did not forget where they were. Their receivers were fed a false position, and the equipment recorded it faithfully. The record did not lie. It was lied to.

The book is being rewritten. This week the NI announced a second edition of that volume. It now covers electronic records, artificial intelligence and digital reconstruction in accident investigations, and it says material produced or supported by AI must be checked, not accepted without professional scrutiny. Its technical editor, Simon Daniels, said ships now produce a huge amount of digital information, but more data does not automatically make an incident easier to understand. You need to know where the evidence is, act fast to preserve it, and understand the limits of the system that produced it. The foreword is by the head of loss prevention at Gard.

And an insurer is paying attention. Also this week, P&I club NorthStandard added PntGuard from SGM Technology to its GetSET! list of safety technology, with a 15% discount for members. The system compares the ship's satellite position with an authenticated signal from the Iridium satellite network, using equipment from NAL Technologies, and sounds visual and audible alarms when the two disagree. It was developed together with the shipowner Tschudi Shipping. Last week we asked whether the bridge would ever show the crew when a position is false. Here is one answer that exists today, before Galileo's protected signal arrives in 2027. Two honest notes. It is a discount, not funding, and NorthStandard has fully funded an AI safety pilot before, so this is a smaller step. And no performance data was published.

Why this matters for AI. AI tools run on records. Voyage optimisation, emissions reports, AI-written incident summaries and digital reconstructions of an accident all start from the data the ship recorded. If the record was fed a lie, the AI repeats it with confidence. And if the data pipeline cleans away the "impossible" points, both the AI and the investigator lose the only sign that something was wrong.

Three things to change this week. First, tell your data providers and your own IT team: flag impossible positions, never delete them. Second, on ships trading in known interference areas, have a second position source, and record which source the officer trusted and why. Third, add one line to the incident procedure in your safety management system: any AI-made summary or reconstruction must be marked as AI and checked against the raw record before it goes into a report. None of this needs a new budget. It needs a decision.

Disclosure: this newsletter is produced with the help of Claude, an AI model made by Anthropic. The Nautical Institute's rule applies to us too, so we check what it writes against the sources before we publish. [Smart Maritime Network · Nautical Institute · Nautical Institute · first edition · Splash247 · NorthStandard · Smart Maritime Network · PntGuard]

📊 Why It Matters — Strategic Impact Table

🔭 On Our Radar

• 🗳️ Who Is Actually Paying for Shipping's AI Accounts? — carried for a seventh week: every adoption figure still counts organisations rather than people, our reader survey is open and collecting, we monitor whether any manager, union or class society publishes provisioning guidance instead of general policy language, and we will publish our own findings when we have enough responses, including if they prove us wrong.

• 🏗️ Will a Port Publish Its Own Number? — Huawei's planning and driverless-truck figures are its own, we monitor whether Shandong, Tianjin or Douala publish measured results such as vessel waiting time, berth window reliability or crane moves, and whether any terminal outside China signs up for the new products.

• 📡 Will AIS Providers Keep the Evidence? — impossible positions are routinely cleaned away, we monitor whether any major AIS or voyage-data provider publishes how it treats them, whether insurers begin asking for raw position data after incidents, and whether Worldwide AIS Network names a customer or publishes its detection method.

• 🔐 Will Equipment Makers List What Is in the Box? — the risk often sits in bundled network devices and controllers, we monitor whether makers start supplying full software component lists with their E27 documentation, whether class surveys bring those parts into scope, and whether anyone publishes a larger, independent sample than CYTUR's.

• 🗼 Who Tells VTS the Ship Is Using Machine Help? — shore decision support is due as a demonstrator in 2028, we monitor LEAS-PRO milestones, whether IMO work on autonomous ships covers how a ship reports its control mode to VTS, and whether any VTS authority publishes procedures for ships under assisted navigation.

• 🎓 Does AI Literacy Reach the IMO Agenda? — the STCW review is writing its amendments now, we monitor whether any member state or industry body formally proposes an AI competence, and whether any ship manager publishes its own internal standard for how officers use and overrule AI tools.

• 🧾 Will Anyone Publish Results for Position Checking and Forecasts? — NorthStandard offers PntGuard at a discount and WiseTech offers a free forecast, and neither has published performance data, we monitor whether NorthStandard or SGM release detection figures from members' ships, whether WiseTech publishes a back-test of its outlook, and whether the new Nautical Institute guide gives practical checks for AI-made evidence.

📅 Critical Maritime AI Research Areas for Managers

1. 📡 What Data Cleaning Throws Away: AIS and voyage-data services routinely remove impossible positions as outliers. Research should measure how much interference evidence is lost this way across the main providers, and propose a standard for flagging and keeping such points rather than deleting them.
2. 🔐 Cyber Risk in Bundled Shipboard Components: one small study found the risk sat in network devices, PLCs and screens delivered with the main equipment, not in the main unit. Research should take a large independent sample across makers and ship types, and test whether known-flaw counts predict real risk at all.
3. 🏗️ Does Port Planning AI Reach the Ship?: vendors report how many machine plans operators accept, not what happens to the ship. Research should measure whether agent-made berth and crane plans change vessel waiting time, berth window reliability and total port stay for callers.
4. 🗼 What the Shore Needs to Know About Ship Control Mode: VTS operators will soon watch ships under normal, assisted and autonomous control. Research should test in simulators how operators decide when they do and do not know a ship's control mode, and define the minimum information a ship should report.
5. 🎓 Testing Whether Officers Catch AI Mistakes: the case for AI literacy rests on officers spotting confident but wrong advice. Research should build scenario tests with false references, missing data and unsafe efficient plans, and measure how often serving officers catch the error today.
6. 🧾 AI-Made Material in Casualty Investigations: AI summaries, transcripts and reconstructions are entering incident reports and claims. Research should measure their error rates against raw records, and set out the labelling and verification standard that investigators, insurers and courts would accept.

📈 Top Investment Opportunities

1. 📡 Interference Detection That Keeps the Evidence — impossible AIS positions are proof of jamming and spoofing, and most data pipelines delete them, while insurers and authorities increasingly need that proof — the opportunity is in receiver networks that test positions against physics, flagging and archiving services for raw position data, and evidence packages for claims and investigations — Worldwide AIS Network
2. 🔐 Component-Level Cyber Assurance for Ship Equipment — the risk in ship systems often sits in bundled network devices and controllers that nobody lists, and patches that were never installed — the opportunity is in software component inventories for marine equipment, patch management for third-party parts, and commissioning checks that close unused ports and services before delivery — CYTUR
3. 🧾 Evidence Preservation for the AI Age — the Nautical Institute is rewriting its evidence guidance for AI and digital reconstruction, and P&I clubs are paying more attention to position integrity — the opportunity is in fast recovery and preservation of onboard data after incidents, tools that check AI-made summaries against raw records, and second position sources that log which source was trusted — Nautical Institute + NorthStandard
4. 🏗️ Port Planning Agents and the Layer Around Them — Chinese technology groups are packaging port AI from live terminals into products for export, with claims of plans made in minutes and horizons of a month — the opportunity is in integration with existing terminal systems, independent measurement of planning results, and the data links that let shipping lines see and challenge a machine-made berth plan — Huawei Port Intelligent Planning
5. 🎓 Scenario-Based AI Competence Training for Crews — a formal STCW competence may arrive around 2030, and companies can act well before that — the opportunity is in product-neutral training built on realistic AI failures, assessment methods that test behaviour rather than definitions, and short refreshers delivered on board as tools change — Seatrade · STCW review

📅 Top Monthly Picks

1. 🛰️ Galileo Holds the True Position While the Spoofer Lies — Europe's first civil satellite position protected against spoofing, tested live in Norway, with the service due free worldwide in 2027; read it next to this week's four tankers in Lima and NorthStandard's alarm system, which show what ships have to work with until that signal arrives — ESA
2. 🇮🇳 JNPA Signs a $9.7m AI Digital Twin Only After a Live Test — India's largest container port made bidders prove it on real operations before signing; it is the right test for this week's Huawei launch, where the only figures so far come from the seller — Splash247
3. 🔒 Somebody Finally Built the Spreadsheet Replacement for E27 — a tool for tracking supplier cyber certificates under the new IACS rule; this week's CYTUR study adds the missing half, because the certificate covers the main unit while the risk often sits in the parts delivered alongside it — Smart Maritime Network
4. 🛳️ CMA CGM Designs an AI-Assisted Containership and Asks How Much Control Stays With the Crew — the design study put the human question in writing before the steel was cut; this week the same question reached the training rules, with a call to make overruling the machine a formal STCW skill — Splash247
5. ⚖️ Who Pays the Inspector? — we argued last week that the check worth trusting is the one nobody paid the vendor for, such as a regulator, an insurer or your own trial; NorthStandard's move this week is an insurer stepping in, though with a discount rather than its own money — AiatSea, 20 September

Have feedback or a story tip? Email us or connect on LinkedIn.