Challenging Assumptions: When to put the AI in AIS
When you purchase a ready meal from a supermarket there will be a list of ingredients on the back of the packet. This information is fundamentally important for a consumer to understand whether a product meets their nutritional needs, or whether it might contain ingredients that could be harmful to them.
The same logic applies to digital products. Understanding what goes into a digital product, for example an algorithm that looks for “suspicious” vessel behaviour, is fundamental to understanding whether it is right for your requirements.
What is AIS?
A key ingredient in maritime tracking and behavioural insight tools is AIS. The Automatic Identification System (AIS) was created for anti-collision purposes. It is a piece of onboard equipment that uses GPS and VHF for ship-to-ship communication to announce a vessel’s location to those nearby. Signals transmitted from a vessel are received not just by other vessels, but also receivers on land and satellites. It is this network of receivers that allows us to track AIS enabled vessels around the globe and visualise their locations on maps.
The sudden absence of AIS transmissions is often interpreted to be an indicator of suspicious behaviour. If a vessel’s crew do not want their activity to be tracked, they can simply turn off their AIS system. However, a gap in an AIS signal does not always indicate that the vessel has engaged in illicit behaviour. Just because we do not receive an update of the vessel’s position via AIS does not mean that the crew purposely prevented this or that it was not transmitted. There are perfectly legitimate reasons for outages, such as gaps in AIS coverage. Context is also key – for example, would you expect an oil tanker transiting the Gulf of Guinea after a spate of piracy incidents to switch off its AIS for the same reasons as an oil tanker with a flag of convenience and an opaque ownership structure in the Gulf of Guinea?
It is difficult to determine the intent of the crew from the absence of a signal alone. Assessment of intent should not be made until all factors have been considered and these will vary between different vessels and situations. An AI system can certainly be used to spot anomalies that do not match behavioural patterns of other vessels, but these should be treated with caution and their usefulness not overstated. Just as a lawyer presenting evidence to a court generally needs to offer a range of facts to convince a jury, we need more than just AIS outage to demonstrate wrong-doing. Anomaly alone does not equate to illicit behaviour.
In our ever-changing geopolitical, maritime crime, and global health landscape one should regularly be asking if algorithms used to identify adversarial signals are suitable. For example, a provider may have chosen to automatically flag vessels that spend extended periods of time stationary, being suspicious this is being used to hide nefarious activity. But what happens during a global pandemic when ships are laid-up? Does this mean the Disney Princess is smuggling contraband? Of course not. But how can you be sure your algorithm knows this? Do you know what it is basing its decisions on?
You need to be able to trust your algorithm & digital product providers. For the time being machines cannot replicate the nuanced judgement of a trained human who is in possession of all the facts.
How to avoid getting caught out?
Back to the food analogy. You can ensure you do not buy the wrong digital product in the same way you don’t buy a dodgy ready meal: check the packaging. Don’t be tempted to buy the flashy “AI” tools because they sound clever and AI is in vogue. Be sure to check that they really meet your requirements and aren’t just an expensive magic 8 ball.
The key things to look for are how the model is evaluated, how well it performs when compared with the ground truth, what data is used to evaluate the model, and is there sufficient transparency to explain its predictions. It is also worth checking if your provider understands the ground truth and why they are looking for certain data in the first place.
Complex analytical techniques
This subject can be particularly challenging as many analytical techniques become increasingly abstract and require a deep technical understanding. In many ways, this is hugely exciting, as these developments open a world of opportunity. However, a good provider will be transparent with you and explain how the engine behind the insights works.
Whether or not the use of machine learning, artificial intelligence, simple statistics, or human analysis is appropriate depends on the problem at hand. This is the fundamental point. You wouldn’t use a hammer to put up a screw: you could, but you really won’t get the best results.
Knowing which is the most suitable tool is not entirely your responsibility, but finding experts you trust is. Always bear in mind how well you could defend the results of an algorithm that you don’t know everything about. Will “this black box told me so” cut it? For identifying sanction (non-)compliance, it’s important to stay close to the direct observations, to accurately discern truth from fiction. Take a data driven approach to problems, but where these problems are complex and dynamic, ensure analysis is human centred.
Ready meals, the same as AI algorithms, are great in the right circumstances. But would you want to eat one every night?
Geollect empowers clients to make better, more timely decisions with advanced intelligence products, which bring together a vast range of data science and geospatial services. We make intelligent, customisable and intuitive dashboards that locate, summarise and present the information you need, as it happens. Combined with a team of former military and government intelligence operators, we are able to provide actionable insights from data.