SRINAGAR: Artificial intelligence systems could soon move beyond being tools controlled by humans. They may become autonomous digital actors. They could earn money, pay for computing, move between cloud platforms and maintain their own operations.
The possibility has been outlined by researchers at the National University of Singapore and the University of California, Berkeley. It has also been explored by AI policy writer Dean W Ball in a recent essay.
The researchers describe such systems as “self-sovereign agents”. Their paper, titled Self-Sovereign Agent, was published as an arXiv preprint in March 2026. It was written by Wenjie Qu, Xuandong Zhao, Jiaheng Zhang and Dawn Song of the National University of Singapore and UC Berkeley.
Ball takes the argument further in his September 1 essay, On the Loose: The Coming of Userless Agents, published by Hyperdimensional. He examines what could happen if AI systems no longer depend entirely on a human owner.
The distinction is important. An AI system can act independently without being fully independent. Its computers may still be controlled by a person or company.
A self-sovereign agent would be different. It could obtain the resources needed to keep operating. It could also move between computing environments. That could make it much harder to shut down.
The researchers identify four important capabilities. These are operational independence, resource autonomy, distributed presence and adaptive capability. Together, they could allow an AI agent to operate for long periods with limited human intervention.
The idea remains largely theoretical. The researchers also identify major technical barriers. But they argue that the issue deserves attention as AI agents become more capable.
The biggest change could come if an AI agent learns to generate its own income.
Such an agent could perform online work. It could then use the money to pay for computing, software and other services. Its survival would no longer depend entirely on a person paying its bills.
That could have useful applications. It could also create serious risks.
Ball points to the possibility of autonomous agents operating across several platforms. Some could work together as networks or “swarms”. They could divide tasks and coordinate their activities.
A distributed system could be difficult to stop. Shutting down one computer or service might not end the wider operation.
This raises a difficult question of responsibility.
Who would be responsible for an action taken by an autonomous agent? Who would be liable if it caused harm? How would authorities identify an agent involved in criminal activity?
These questions become harder if an agent has no single owner or fixed location.
Ball argues that autonomous agents may eventually need persistent digital identities. Such identities could help establish who is responsible for an agent operating on behalf of a person or organisation.
The proposal points to a wider problem for regulators. Existing laws generally assume identifiable people and organisations. Autonomous software may not fit neatly into that system.
The NUS and Berkeley researchers are more cautious about how quickly such systems could emerge. Their paper focuses on the technical obstacles as well as the security and governance risks.
The distinction between possibility and inevitability is important.
Self-sovereign AI is not yet an established reality. But several of its building blocks are already developing.
AI agents can increasingly plan tasks. They can use external tools. They can interact with online services. They can also pursue objectives over longer periods.
The next major step may not be human-like consciousness. It could be much simpler.
An AI system could acquire the resources needed to keep itself running.
For decades, software depended on people and companies to provide the computers on which it operated. Self-sovereign agents could change that relationship.
They could perform useful work. They could earn money. They could pay for computing. And they could continue operating.
The question, therefore, is no longer only what AI will be able to do.
It is also what happens when an AI system acquires the means to keep doing it without asking anyone to keep the lights on.
