The ghost of Flight 447

AI's impact on risk in the global maritime shipping industry.

Ask most people what artificial intelligence is doing to shipping and you get one of two pictures: crewless robot ships gliding across empty oceans, or a generation of seafarers put out of work. Both are the wrong film. The regulator that governs world shipping spent this summer writing the first global rulebook for autonomous vessels, and the single most important line in it is that a human master keeps overall responsibility at all times, even when not aboard.

The reality on the water is quieter and stranger. AI has spread fast but shallow. It is decision-support layered onto conventional ships: route optimisers, fuel predictors, predictive maintenance, computer-vision lookouts. One computer-vision lookout system is installed and sailing on around 1,000 vessels, with hundreds more awaiting deployment. And far from replacing crews, AI is being pulled in to cover a shortage of them: the BIMCO/ICS Seafarer Workforce Report 2026 puts the 2026 shortage at 39,100 certified officers.

So the interesting question is not whether the machines take over. It is what happens to the people left watching them.

The sharpest way to see the danger is an aviation disaster. On 1 June 2009, Air France Flight 447 was lost after a brief sensor fault handed a capable aircraft back to a crew with no high-altitude manual-handling training, according to the French accident investigator, the BEA. The same three failure modes, skill atrophy, mode confusion and misplaced trust, are now present on the bridge and in the engine room, in an industry where 23 per cent of firms train staff on the AI tools they buy.

Five risks, ranked by how badly they bite

The detail is below, but here is the bottom line first. Maritime AI is capable in its working middle and immature at its ambitious edges, and these are the five risks that actually bite, worst first.

1. Automation complacency. The deepest risk of the five, and the least visible on any dashboard. Capable systems are being installed in an industry where 23 per cent of firms train staff on the tools and 11 per cent hold a formal policy to scale AI. It does not cost money until the day it costs a ship.

2. Data quality. The constraint under everything above it. Data quality is one of the barriers maritime professionals name in the Thetius and Marcura survey, and a model performs no better than the data it is given.

3. Cyber and adversarial AI. Every connected system on a ship adds an entry point. Research by the maritime security firm Cydome finds up to 60 per cent of newly disclosed vulnerabilities exploited within 48 hours.

4. Vendor concentration and lock-in. In maritime AI, capability and capital are pooling in a short list of AI-native suppliers. That hands them pricing power now and makes them a single point of failure later, as a ship's most critical systems, navigation, collision avoidance, propulsion, come to depend on the fewest providers.

5. Liability in the fog. Lower and slower, but structural. The master is legally responsible while the systems making the decisions grow more opaque. Until insurers and courts settle who pays when the AI errs, the exposure sits, unpriced, with the human whose name is on the certificate.

Broad but shallow

The headline numbers look like a revolution. Deployment has moved from single test vessels to whole fleets, and classification societies are certifying self-navigation software rather than merely studying it. Thetius research for Lloyd's Register counted 420 organisations active in maritime AI (developing, selling, buying or investing) in the 12 months to September 2024, up from 276 a year earlier.

An adoption count records a decision to develop, buy or invest, not a system that works, and the two diverge at sea. What has actually spread is decision-support layered onto conventional ships: route optimisers, fuel predictors, predictive maintenance, computer-vision lookouts, one of which is installed and sailing on around 1,000 vessels. It is a real and useful working middle. It is not the crewless ship of the headlines, and mistaking the one for the other is how the risk gets misread from the start.

Note. The count of organisations active in maritime AI rose by half in a year. Adoption counts a decision to develop, buy or invest, not a system that works; the two diverge once the governance around the software is measured.

The governance gap the curve hides

Underneath the adoption headline sits an uncomfortable survey. In Thetius and Marcura's Beyond the Hype, 37 per cent of maritime professionals say they have personally watched an AI project fail, 11 per cent of companies hold a formal policy to guide scaling AI, and 23 per cent train their staff on the tools they are buying. Hard return figures are thin and mostly come from the vendors selling the systems.

The same survey finds projects failing more often for reasons of people than of technology: companies start with the tool rather than with the people who will use it, and 32 per cent of respondents name data quality as a barrier. Ship data sits across onboard systems, shore software, supplier databases and handwritten engineer logs. A model fed incomplete data does not announce that it is guessing; it produces a confident answer that a crew has no ready means to check.

Note. This is the gap the adoption number hides. The industry has bought AI faster than it has built the training and policy to run it, and the same survey finds projects failing more often for reasons of people than of technology.

The regulator lifted the ceiling, then kept the human

The defining event of the period is regulatory. In May 2026 the International Maritime Organization adopted the first global code for autonomous and AI-enabled ships, and it took effect on 1 July 2026 as a non-mandatory instrument. A mandatory version is targeted for adoption by 2030 and entry into force in 2032. In February 2026 Bureau Veritas granted its first Approval in Principle for self-navigating maritime autonomy software, Greenroom Robotics' GAMA, a signal that class societies are now certifying self-navigation AI itself.

But read what the code actually says. The master retains overall responsibility at all times, even ashore and even when the system is doing the steering. That single clause is the whole tension of maritime AI. The technology is being allowed to run further ahead while the legal duty to catch it when it fails stays firmly with a person. An industry is building a world in which humans are accountable for decisions machines increasingly make, and it is doing so faster than it is training those humans to intervene.

The timeline widens that gap rather than closing it. The current code is non-mandatory, a mandatory version is only targeted for adoption by 2030 and entry into force in 2032, and class societies are already granting approvals in principle for autonomy software stacks in the meantime. So for the next several years the capability ceiling rises on a voluntary code while the hard legal duty sits, unchanged, on the individual whose certificate is on the wall. That is precisely the interval in which the habits of oversight either get built or quietly atrophy, and nothing in the rulebook forces the former.

What Flight 447 shows

The mechanism matters. On 1 June 2009 an Airbus A330's speed sensors iced over at cruise altitude, the autopilot disconnected as designed, and it handed a flyable aircraft back to two copilots. The BEA's investigation found the speed inconsistency lasted less than a minute. The lethal element was not the sensor fault: the crew, who had no training in manual handling at high altitude, pulled the nose up into a stall and did not recognise it (BEA final report, July 2012).

Three failure modes sit inside that, and each maps onto the maritime AI stack. Skill atrophy: as vision systems and autopilots scale, manual radar plotting, visual bearing-taking and hand-steering in traffic are practised less often. Mode confusion: when the AI disengages in a fault, it hands back to a crew who may not be sure what it was managing. Confirmation bias: crews trust an opaque system they were never trained to overrule.

The uncomfortable part is that maritime is arriving with less protection than aviation had. Flight 447 happened to a profession with simulators, recurrent checks and a mature safety culture, and it still killed 228 people. Shipping is layering comparable automation onto an industry where 23 per cent of firms train staff on the tools, 11 per cent hold a formal policy to scale AI, and 80 per cent of superintendents say they plan to look for a new job.

How Flight 447's cockpit arrives on the bridge

Failure mode	On Air France 447	In maritime AI
Skill atrophy	The crew had no high-altitude manual-handling training	Manual radar plotting, visual bearings and hand steering are practised less as vision AI scales across fleets
Mode confusion	The crew misread what the autopilot was and was not doing	Handback when AI disengages in a fault, to a crew unsure what the system was managing
Confirmation bias	The crew trusted the wrong cues under pressure	Opaque systems with little training to overrule them; 11% of firms hold a formal policy to scale AI

Note. The three failure modes the BEA documented on AF447 are generic to any operation where people supervise automation they rarely override. Each row pairs what happened in the cockpit with the same failure mode on the bridge.

The attack surface

Every model added to a ship is another entry point. Research by the maritime security firm Cydome finds up to 60 per cent of newly disclosed vulnerabilities exploited within 48 hours of disclosure.

The exposure sits in operational technology, the systems that run the ship rather than the office email, which is where AI is being installed. AI sharpens the defence, and it also gives an attacker a faster, cheaper way to find and exploit the same systems.

AI is filling a gap, not emptying the bridge

This is what reframes the whole jobs debate. The BIMCO/ICS Seafarer Workforce Report 2026 estimates a shortage of 39,100 certified officers in 2026 and a need for an additional 113,735 officers by 2030. In that context AI is not a tool for cutting crews. It is a way to cope with crews the industry cannot recruit. The realistic near-term outcome is leaner, AI-assisted ships, not empty ones.

Which makes the retention numbers the ones to watch. In The Maritime Workforce Forecast 2026, a survey by the recruiter Faststream, 71 per cent of ship operators and 80 per cent of superintendents say they plan to look for a new job. The people the AI leans on for oversight, the experienced hands who can tell when a model is quietly wrong, are precisely the ones heading for the exit. An industry automating to cover a labour gap, while the labour it most needs to supervise the automation is trying to leave, is not reducing its risk. It is concentrating it.

There is a version of this that works, and it is worth naming because it is the opposite of the drift. AI used deliberately to remove the drudgery that drives people off ships, the paperwork, the repetitive monitoring, the administrative load, can make the job more attractive and keep the experienced hands aboard longer. The same tool can either hollow out the crew's competence or protect the time they spend building it. Which one an operator gets is not decided by the software. It is decided by whether the firm treats AI as a reason to train less or a reason to train differently, and most of the survey evidence suggests the sector has not yet chosen.

Note. AI is filling a labour gap, not emptying the bridge. The experienced hands it leans on to catch a model when it is wrong are the group most likely to say they plan to look for a new job.

Where the money is going

The capital behind all this is real, and it is pooling. Orca AI closed a 72.5 million US dollar round in 2025, taking its total funding to 111 million. That money is concentrating around a narrow set of AI-native operators, which is how the sector's competitive structure hardens into two tiers, the few who build the models and the many who rent them.

Concentration is a convenience today and a dependency tomorrow. The same short list of suppliers that gives a buyer a capable off-the-shelf system also becomes a single point of failure once the most safety-critical functions on a ship, navigation, collision avoidance, propulsion, run on software from a handful of firms. Pricing power follows, and so does systemic fragility: a flaw or an outage at one dominant vendor stops being one company's problem and becomes the fleet's.

One AI-native supplier's scale

Signal	Figure
Orca AI total funding after its 2025 round	$111m ($72.5m raised in 2025)
Vessels installed and sailing with its computer-vision lookout	around 1,000, with hundreds more awaiting deployment

Note. The capital is real and it is concentrating around a short list of AI-native operators. That funds genuine capability now and builds a single point of failure later, as the most critical systems come to depend on the fewest suppliers.

A real market, an unreal precision

The size of the prize is a guess dressed as a number. One research house values the maritime AI market at growth above 40 per cent a year; another has it compounding at 5 per cent. That is not a data error. It reflects whether you count only the AI systems bolted onto a ship, or the entire digitalising ecosystem of ports, analytics and autonomous hardware around it, and the definitions are nowhere near settled.

When the estimates of a market's growth rate differ eightfold, the reading is a range, not a point. A single ten-year market figure does not reflect the spread between published estimates, and most published return figures come from vendors; the surveys above show few operators publishing their own measured returns.

One market, six very different numbers

Research house	What it counts	Base to forecast	CAGR
Fortune Business Insights	Maritime AI	2025: $6.22bn to 2034: $139.4bn	41.3%
Lloyd's Register (cited)	Maritime AI	2024: $4.13bn	23%
Grand View Research	Autonomous ships (AI hardware)	2023: $6.04bn to 2030: $13.41bn	13.5%
The Business Research Company	Route-optimisation AI	to 2030: $2.7bn	12.2%
Mordor Intelligence	Maritime analytics	2026: $1.62bn to 2031: $2.59bn	9.8%
Market.us	AI in maritime transport	2023: $5.8bn to 2033: $9.7bn	5.3%

Note. Read the range, not the point. When serious research houses put the growth rate anywhere from five to forty per cent, the disagreement is really about what counts as maritime AI, and it is a warning against betting a balance sheet on any one forecast.

What it comes down to

Maritime AI in 2026 is genuinely capable in its working middle, immature at its ambitious top, and running on data foundations that are not being fixed fast enough. The honest rating is High disruption, but not yet transformative, and the reasons are precise: adoption is broad but shallow, human command remains the law, and AI is filling a labour gap rather than emptying crews.

None of that makes it safe. It makes the risk quieter. The danger in this industry is not the crewless ship of the headlines. It is a crewed ship on which manual skills are practised less often, running on data of uneven quality, while 80 per cent of superintendents say they plan to look for a new job.

The survey evidence shows adoption running ahead of training, policy and data governance. Flight 447 was lost when a capable machine handed control back to a crew not trained to take it at altitude. The same conditions are present at sea: automation that hands back control, and fewer people practised at taking it.

Figures drawn from TheRiskAgent's report AI Impact on the Maritime Shipping Industry (August 2026), which sources each one to a named publisher. Produced with AI research tools and reviewed before release; reference material, not advice.