When AI meets reality: Systematic’s defence AI provides support for military decision-makers in critical situations
AI can seem deceptively promising at first glance, but may not quite deliver on closer inspection. Consequently, the solutions being developed by Systematic’s AI team need to function effectively with non-ideal, “chaotic” data before they move forward from prototype to ready-to-go functionality.
AI often looks great “from a distance”, but the reality may differ, and that is a problem. In particular when it comes to technology for armed forces, where human lives may be at stake.
“It is not sufficient for it to work in laboratory conditions,” says Søren Løfquist Stiil, AI Architect. He holds a PhD in Computer Science from Aarhus University and is part of Systematic’s 12-strong AI team, working on Systematic’s SitaWare command-and-control system, and more specifically the system’s data and analysis tool SitaWare Insight.
Insight is a type of military version of a commercial search engine, specially developed to create an overview of the multiple data sources at the disposal of military decision-makers. The AI team is made up of data scientists plus experts in testing and validation and MLOps (framework for quality assurance of AI systems).
Tank identification
AI and image recognition are two areas Søren highlights in terms of challenges they face in their efforts to develop defence software. Even though AI-based image recognition performs well with open data sets downloaded from the internet, the technology does not fare so well when dealing with military data, e.g. photos of tanks taken out in the field.
“There are too many nuances in the image material. Even if you had a complete photo index of all tanks in the world and had trained the perfect AI model to identify them, tanks out in the field can look totally different. Their crews attach all kinds of gear to them, ranging from shovels to specialist nets for protection against drones and anti-tank missiles,” he explains.
The more realistic and “chaotic” the data, the greater the risk, according to Søren, that the AI will make a mistake if it has only been tested and trained on “ideal” data, which is not representative of real-world scenarios. The AI needs running in. In order to address the AI’s inherent weaknesses, Systematic’s AI team are focusing on testing and quality assurance, to be able to retrain models over time using realistic data and optimising the interaction between the AI algorithms and the software into which they are incorporated.
Suspicious shipping traffic
The AI team has already integrated a number of new AI-powered features into SitaWare Insight. These include detection of irregularities in shipping traffic, automatic recognition of text in images, which has been specially developed for a military context, and cross-language document searching.
“But in actual fact, the AI itself is only a tiny element of getting AI-based services up and running in production systems like SitaWare. It is much more about the framework you set up around the AI and ensuring optimum interaction when it is incorporated into other software,” explains Søren.
Talk to your data
At present, Systematic is also working on developing talk to your data solutions that make it possible for operators to interact with the systems via a chat interface. In addition, they are looking into retrieval-augmented generation, making it possible for language models to react to data they have not been trained with.
“We are focusing specifically on decision support and optimisation of workflows. It is about supporting the soldiers in making the right decisions on the basis of the data processed by the systems. It should always be a human and not the AI making the final decision.”
