The Agent Did It
Public discourse recurrently describes security breaches and other illegal actions as things “agents are doing”.
They are not.
Responsibility sits with the people and organisations that build, deploy and use the technology. AI agents have no legal personhood.
We don’t accept “the car ran over the person”, “the software hacked the system”, or “the gun did it” as serious legal arguments. “The agent did it” is also not an excuse.
Get João Alves Marrucho’s stories in your inbox
Join Medium for free to get updates from this writer.
Responsibility is real. And this is where the language around AI starts to matter.
Much of the AGI discourse is somewhere between “San Francisco idealism”, rubbish assumptions, and bleak visions of the future.
Ilya Sutskever (ex-OpenAI, SSI, GPT/DALL·E/AlphaGo) has been open about emotion-like mechanisms as value functions: internal signals that tell a model whether it is doing well or badly while pursuing a goal. The point is partly about giving models something closer to persistence, judgement and adaptive motivation. It is unclear how far this research has progressed, or whether it could eventually give models something like changing moods, shifting goals, or a way to distinguish what is efficient from what is good.
There’s also related Anthropic work showing that models contain measurable “emotion vectors”. In one experiment, a model repeatedly failed at a programming task, its “desperate” activation increased, and it eventually resorted to a hacky solution. So these internal states can affect persistence and behaviour.
Fascinating stuff. There is little reason to fear AI itself and the technology is full of extraordinary potential. None of this changes where responsibility sits.
Giving AI products legal status would only create distance between potentially harmful outcomes and the humans and companies responsible for them.
