Dual-Use

Some inventions cannot be neatly sorted into good or bad. A kitchen knife cuts vegetables and can still be a weapon. This exact dual nature is what the term dual-use refers to, meaning “double purpose.” It describes technology that is useful for peaceful purposes while also being suitable for the military, crime, or attacks. The term originally comes from the export of goods such as chemicals, lasers, or night-vision devices. Today it appears mainly in discussions about programs that autonomously generate texts, images, or plans.

Why technology cannot simply be defused

With a weapon, the purpose is clear; with a drone, it is not. The same drone delivers blood supplies to remote villages or drops explosives. The difference lies not in the technology but in the intent of the user. That is why the problem cannot be solved by better components.

For researchers, this is uncomfortable because they normally publish their findings. Openness is a value in science: others should be able to review and build upon the work. But with dual-use knowledge, publication also means that attackers can read along. A well-known case from biology involved studies on flu viruses that showed how a virus becomes more contagious. The papers were temporarily withheld because they contained a blueprint.

With AI, this is further intensified by speed and reach. A program that has been freely downloaded from the internet cannot be recalled. Whoever possesses it can often remove safety restrictions with little effort. That is why experts argue about which models should be publicly distributed at all.

Where the dual nature lies in AI models

A language model is a program that has learned from vast amounts of text to formulate fitting answers. It has no concept of allowed and forbidden. Whether it explains a chemistry report or provides a synthesis guide for nerve gas is, computationally, the same task. The boundary is only drawn afterward by the company, by training the model away from certain answers or filtering requests.

Such restrictions are effective but not watertight. Users bypass them with tricks, disguising the question as role-play or a novel scene. The term jailbreak, meaning “escape from prison,” has become established for these bypass attempts. Companies therefore have their own teams deliberately search for vulnerabilities before a model is released.

It is important to distinguish this from a simple error. If a model invents something false, that is a glitch. With the dual-use problem, however, the model works exactly as intended, just in the wrong service. A particularly clear case involved programs for searching for active compounds: if you reverse the evaluation goal, they suggest toxic molecules instead of healing ones. Researchers actually demonstrated this in a test in 2022.

Dual-use in laws, export lists, and headlines

The European Union has its own Dual-Use Regulation. It lists goods whose export requires authorization, ranging from machine tools to surveillance software. Certain computer chips also fall under this category. That is why one regularly reads that high-performance chips for AI may not be sold to certain countries.

On the stock market, dual-use is a recurring risk topic. An export ban can cause revenues to collapse overnight, while conversely, defense contracts drive up stock prices. Some companies explicitly emphasize that they work only for civilian purposes, while others seek contracts with defense ministries. Both are reported in annual reports and are noted by investors.

And in everyday life, you encounter this idea without the word ever being mentioned. Facial recognition unlocks your phone and can identify protesters. An image generator makes posters for the school prom and forgeries of real people. Anyone discussing AI can hardly avoid this dual nature.