THE MOTION
“Actions carried out by AI agents should be legally required to be watermarked.”
Discuss.
Where the room broadly agreed
On 23 October, our first Campus Discussion brought together nine students with different perspectives on AI. The group broadly supported a legal requirement to watermark actions carried out by AI agents. The strongest concerns centred on feasibility: how such a requirement could work, who would be responsible, and whether it could be enforced.
The conversation brought together backgrounds including social data science, social science of the internet and energy systems, alongside other fields.
Who would carry responsibility?
Participants considered how responsibility for an agent’s actions should be split between the person deploying it and the company that created it. That raised a connected question: who would have to ensure an action carried a watermark, and who would be accountable when it did not?
What would a watermark look like?
The group asked how visible a watermark should be. Would transparency require a label that an ordinary person could immediately recognise, or could a trace that only software detects be enough?
Participants used statistical traces in text authored by generative AI as an analogy: something potentially detectable by an AI detector, yet invisible to a reader with the naked eye. This left open what meaningful transparency should look like in practice.
One standard, or many?
If watermarking became a legal requirement, should every company and model follow a common procedure? Participants questioned how such a standard could be agreed and implemented.
They also raised the alternative risk: without standardisation, companies might favour or discriminate against the actions of agents from particular providers. A system intended to make AI activity transparent could also shape which agents others accept or trust.
The challenge of open-weight models
Participants questioned how a watermarking requirement could be imposed on agents built with open-weight models. The difficulty of demanding compliance in that setting was a significant challenge to the proposal’s feasibility.
Declared identity, or external detection?
The group was concerned about relying on external detection software, or even another AI agent, to identify and mark agent activity.
Mouse-movement metadata on webpages offered a concrete example. If a detector identified agents through distinctive mouse movements, participants argued that engineers at frontier labs could develop agents whose movements more closely resembled those of humans. A detection method could therefore create an incentive to evade it.
A QUESTION TO CARRY FORWARD
How could we make agent activity identifiable without building a system that is easy to evade or unfair to some providers?
The room broadly supported the principle. The technical design, allocation of responsibility and enforcement remained questions for further discussion.
This is Harry’s summary of the discussion, rather than a verbatim transcript or a formal vote. It describes the views explored by this group, not a representative view of Oxford students. Individual contributions are not attributed.
THE CONVERSATION CONTINUES
What do you think?
Add a perspective, challenge a proposal or return to a question we left open. You don’t need to have attended.
Opening the discussion…