AI is a great tool when used for non-safety-critical methods
It is clear now that artificial intelligence (AI) has become a significant player in our everyday lives. Today, people and companies use AI in various daily activities. For healthcare in particular, AI has provided lifesaving contributions, with, for instance, systems that can detect tumours in medical images. Transportation and navigation have also offered positive outcomes with AI by now, where it is, for example, used for traffic management to predict traffic flows and congestion. Even in aviation, one of the most safety-critical industries, there are areas where AI also works very well, such as weather prediction and administrative tasks.
However, the approach to AI is different in safety-critical environments such as unmanned traffic management (UTM), air traffic management (ATM), and space traffic management (STM). Skypuzzler’s CTO Morten Skov and Tech Lead Fredrik Holsten both deem AI an excellent tool for different non-safety-critical systems, but not when it comes to flight safety. The most common AI models, such as machine learning, are data-driven, meaning they are trained on historical data. This may be a good approach to weather prediction, but for aviation safety-critical decisions, a wrong prediction can have dire consequences. According to Morten and Fredrik, this is why we need algorithms that allow humans to understand their core components and how they function.
“With machine learning, you will need to retrain the entire model if an error occurs and you want to fix it. With Skypuzzler’s solution, we can inspect that exact component, determine where it went wrong, and modify that part of our system.” Says Fredrik.
Today’s definitions of AI are too comprehensive
There are several ways to define AI and even different types of AI. Morten and Fredrik use the definition provided by the European Artificial Intelligence Act, where AI is defined as “a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment (…) generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments” (artificial intelligence act). Based on this definition, Skypuzzler uses AI. The problem is, however, that there are different types of AI – especially types that Morten and Fredrik do not want Skypuzzler to be associated with.
“The definitions and categorisations of AI are too wide. There is a major difference between having an advanced automation system based on deterministic programming and then having it based on data-driven models.” Morten says.
Skypuzzler’s solution is an advanced automation system based on deterministic programming, a mathematical model that works in real time. This allows for human interaction and interference, meaning that Skypuzzler can isolate the responsible code and improve it directly if needed—an action that a data-driven AI model does not afford. For that reason, categorising Skypuzzler with AI firms that use machine learning is problematic, as it is also a matter of industry where it is applied, safety-critical level, and, most importantly, the type of technology.
Skypuzzler is no AI company
Both Morten and Fredrik predict a change in companies’ behaviour when using AI. Fredrik hopes that companies will realise the value in transparency when explaining their technology instead of describing it as AI. He believes there is a need for more simplicity in the approach to new technology. Morten expects companies to start considering how costly it is to use data-driven AI resource-wise, due to the constant need to train it.
Even though Skypuzzler’s technology can be categorised as AI, Skypuzzler should not be described as an “AI company”, but rather as a deep tech company that provides solutions based on mathematical algorithms.