Desai and Riedl: AI Agents Shouldn't Get Legal Personhood

A new arXiv paper by Deven R. Desai and Mark O. Riedl takes direct aim at the argument that AI agents, the autonomous systems that OpenAI, Google, Microsoft, and Salesforce are all promising will move from generating text to actually executing tasks like booking a trip or posting across social platforms, need a new legal category to keep them in check.

The authors' central claim is technical, not doctrinal. The way one piece of software interacts with another, they write, already creates mechanisms to discipline agent behavior, and in their view those mechanisms can constrain 'rogue, undesired actions' more effectively than rules built to govern human agents. Bolt that onto the computer-science literature on value alignment, they argue, and users get a practical way to prevent or correct an agent's behavior in the moment, rather than reaching for the courts after the fact.

The concerns they are answering are ones other legal scholars have raised: 'rogue commerce, human manipulation, rampant defamation, and intellectual property harms.' Desai and Riedl do not dispute the list. Their bet is that the fixes are engineering fixes, and that pushing for regulation before those fixes are tried risks locking in a heavier framework than the technology needs.

On the personhood question they are blunt. AI agents 'need not, and should not, be given legal personhood status,' and humans remain responsible for what their agents do.

The paper is a legal argument, not an empirical study, so it does not measure how well any specific software protocol actually catches an off-policy action in a deployed agent, and it does not name specific products beyond the four company names in its opening. Read it for the framing. For legal teams at agent-building companies, the useful signal is that at least one scholarly voice is offering a defense of the current liability regime, right as the calls to replace it are getting louder.