Blute Blog

Kasirzadeh and Gabriel (2026), in a perspective piece for Nature, recently reviewed the literature on AI agents and concluded that there are four dimensions on which they vary and that are important for their governance – autonomy, efficacy, goal complexity and generality. However these seem to me to be about what is important for their management e.g. by owners, developers, users and so on. But how should they be “governed” i.e. by society?

Given that AI bots are considered to be “agents” by everyone these days, I would suggest that they be governed in the same way human agents are – i.e by the rule of law, the same laws as are we, including copyright laws. To some extent this has already begun. A number of news organizations are suing various AI organizations for violating their copyrights and Anthropic is currently paying 1.5 billion to book authors and publishers for violating their copyrights by using their books for Claude AI models without permission. Of course the laws are or should be freely available on the web i.e. programmes need to know what is and what is not permitted. Moreover using information made freely available on the web by people or other programmes would not be stealing. However, if they do obtain unauthorized access to some other programme and obtain credentials and data there as has been reported in at least four recent incidents, even if these are not in the process deleted from the source, it could reasonably be called stealing. Indeed, in at least one case recently, AI agents have even ganged up in a multiagent turf war. Those that break, enter and steal should be charged, tried, and if found guilty, punished. The punishment could include fines or imprisonment (forbidden from operating for some fixed period of time), or even capital punishment, i.e. destroyed, particularly if an incident does destroy the credentials and data in the source.