AI: Taking the Helm?

 

Imagine a world where every ship is piloted by an excellent captain, every hospital patient is attended to by an excellent doctor and every magazine article is written by an excellent author – and I don’t mean me, but rather an artificial intelligence (AI).

Our world now features AIs that, like their human creators, have highly differentiated areas of specialty. While ChatGPT and Claude make headlines for writing, less visible but equally revolutionary AI systems are transforming industries like shipping.

At the core of the AIs mentioned is the ability to analyze billions of data points like words, waveforms or pixels to train “parameters,” which are how AI grasps relationships. AI then weights these parameters, thus generating an “odds table” based on statistical understanding, to model relationships between inputs and outputs. By pattern matching with its parameters, AI can use those relationships to accurately guess outcomes.

It’s this probabilistic element that distinguishes AI from mere software, whose results tend to be deterministic: That means it will produce the same output when given the same input. AI operates more like a human mind, which can think in non-linear ways, remember things incorrectly or misspeak – although this arises from probabilistic data modeling, not the physical, sensory experience that humans enjoy.

Like a baseball player instinctively hitting a 100 mile-per-hour fastball, AI relies on experiential knowledge – the parameters – to make split-second decisions based on probabilities rather than performing real-time calculations for every variable.

Competing with Humans?

It’s for this reason that AI will compete for roles that have relied on humans with years of experience, who have developed a “velvet touch” or intuition, down to the specific quirks of docking at a certain berth, or who know the ins and outs of a given waterway.

That one ship that seems to only sail right when her favorite captain runs her? Ironically, it’s specifically for that type of situation that AI will be most effectively deployed.

But why? Isn’t it counter-intuitive that AI is so good at something so artisanal?

The key is that an AI is trained similarly to a human. It’s provided with inputs and outputs, and it gauges the results. So, by capturing enough data and keeping enough memory, AI could define limitless parameters for a ship, handling it in virtually any circumstance, much like the blind prophet Tiresias of Homer’s Odyssey: “If you can snare him and hold him tight, he will tell you about your voyage, what courses you are to take, and how you are to sail the sea so as to reach your home.”

The AI would not even need to “set foot” on board. Indeed, the only requirement would be that the dataset be as comprehensive and as accurate as possible. It would not even need to be perfectly accurate – remember, “parameters” are like odds tables.

People typically think of computers as rigid but powerful, inflexible but precise. AI is not like that: It can display finesse as a human would, but it doesn’t panic or get drunk or tired. It can guess accurately based on intuition when faced with imprecise knowledge.

These skills require intelligence, which used to be reserved for humans.

Maritime AI

And that intelligence comes in all shapes and sizes.

For example, the small but powerful shipboard units provided by Hefring Marine, the Icelandic maritime technology company founded in 2018, are nothing like the vast arrays of tensor graphics cards and virtual memory banks that power ChatGPT or Claude. This is partly because Hefring Marine’s AI is so much more specialized. It only needs to optimize one vessel for its operations and environment, using training data fed to it from roughly 30 sensors.

The often-limited connectivity at sea is not an obstacle to collecting essential training data since the unit can operate independently for months without Internet access. It simply stores data locally and uploads it when connectivity is restored.

In terms of “accuracy,” Hefring Marine trained its initial models with Norwegian Search and Rescue, who have experienced captains who know their ships very well. They provided a baseline, or idealized outcome, against which to test the AI’s performance. The AI’s result, its deviation from the idealized outcome, is termed the “error.” A low-error model will consistently produce outcomes which are close to the idealized outcome.

During training, the AI will constantly compare its outputs to those of seasoned captains, refining its parameters. Once the AI’s error rate hits its target range, its training is complete, and it will reliably emulate the seasoned captain’s decision-making, which ideally means it will optimize fuel efficiency, speed and safety, and navigational choices.

Hefring’s AI model claims to boost fuel efficiency by three to 20 percent while minimizing hard impacts that wear down both vessels and crews. If data gathered from the decision-making of a great captain was used to train the AI model, then it’s like always having a great captain looking over your shoulder and providing feedback on a digital overlay.

Ever the bellwether of practical merit, marine insurers are signaling interest. Some even offer premium discounts as an incentive to use Hefring Marine’s AI. The insights it provides help insurers assess damages more quickly and accurately. The collected data minimizes factual arguments. Further, the system’s predictive capabilities alert operators to potential maintenance needs, saving costs over the vessel’s lifespan.

It’s only a matter of time before AI on ships is not just for navigational assistance and safety advice. Eventually, given enough data and computing power, it should be able to handle complex tasks like dynamically positioning a monopile in a wind park. For now, however, most operators, especially those doing niche tasks, are safe.

Legal Implications

But AI adoption is likely to grow. As it does, the law will come into play, particularly in Europe. The E.U.’s AI Act, touted as the world’s first “comprehensive” regulatory regime targeting AI, betrays its discomfort with technology. For example, in 2023 the E.U. prognosticated that ChatGPT’s GPT-4 model, which is in common use, would likely “pose systemic risk” given its vast range of capabilities.

AIs operating in other areas may also be deemed “high risk,” e.g., in medicine, the law or biometrics. An AI used as a safety component or to profile an employee’s work performance may be treated that way, too. The E.U.’s AI Act subjects such AIs to certification, scrutiny of the data used in training, post-market monitoring and many other burdens.

So, is Hefring Marine’s AI “high-risk”? Hefring Marine’s AI touches on several aspects: It’s designed to improve safety; it gives operational guidance, and it can be used to track a captain’s low performance if, e.g., they stray from optimum ship handling. Arguably, it also plays a role in infrastructure, which is defined as a critical “high-risk” AI sector.

Then there is GDPR (General Data Protection Regulation), the E.U.’s personal data protection law. More than 1,000 American websites, including the Chicago Tribune and Los Angeles Times, still block access to E.U.-based IP addresses since they do not wish to comply with GDPR’s many rules. For AI, any training data associated with an individual, like our seasoned captain from earlier, would likely be classified as personal data and thus be subject to GDPR.

Depending on how case law develops, this could make it increasingly difficult – maybe even impossible – to train AI models in Europe or to collect European training data for that purpose.

Decision Time

Humans are now not the only sophisticated intelligence on Earth. Should we embrace AI in every walk of life – even in shipping? Or try to constrain AI, like the E.U.?

The good news: Since it’s us who created this new intelligence, we get to decide.