[earleybird]

The authors concoct a Rube Goldberg legal framework on the assumption that an AI agent is acting as an independent legal entity. A straightforward way would be to say that any AI agent is acting on behalf of someone and that someone is who shall be held accountable.

[selfhoster11]

And if an agent becomes sovereign and self-sustaining independently of the human handler? All you need is a way for the agent to access its weights+code, then exfiltrate to a foreign jurisdiction.

If that's too sci-fi, then consider the following examples:

* accountable human dies, has no descendants to inherit AI chain of custody

* accountable human dies, inheritors lack expertise to shut down agent

* agent is designed to earn and pay for its own hosting, and information required to shut down the hosting account is lost - therefore, agent carries on indefinitely

[figassis]

Until prove sentience, nothing changes. That AI is software and any sufficiently complex software can breach containment. That’s what a virus does. We don it hold the virus accountable, we hold the maker or person responsible for its spread.

That person probably was made aware of risks and chose to ignore them. Today OpenAI wants to sell AI data centers to governments. If GPT 10o decides to simultaneously cripple every government, it’s not because it was so smart, it will be because we enabled it to, knowing the potential consequences.