What Matters
The urgent legal question is not whether AI agents should become “persons.” It is who authorizes them, who owns their outputs, who is bound by their contracts, who pays when they cause harm, and who can shut them down. Existing law already recognizes automated contracting in important contexts, which makes attribution and accountability more immediate than speculative personhood.
The six questions that matter before personhood
| Question | Practical issue | Default DN position |
|---|---|---|
| Authority | What is the agent allowed to do? | Explicit mandate, limits and revocation. |
| Attribution | Whose act is the agent’s act? | Attribute to the deployer or party that authorized the system unless law clearly provides otherwise. |
| Ownership | Who owns outputs, assets or IP? | Define contractually and by applicable law rather than assuming agent ownership. |
| Liability | Who pays when the agent causes harm? | Keep accountable human or legal persons in the chain. |
| Appeal | Can affected people challenge an agent action? | Yes for consequential decisions. |
| Shutdown | Who can stop the agent? | Always maintain an independent kill, revoke or suspend path. |
DN Policy Position Simulator
Automated contracting already exists
AI agents do not need legal personhood for automated transactions to have legal effect. UNCITRAL’s Electronic Communications Convention recognizes contracts created through automated message systems, including when no natural person reviews each action. Its 2024 Model Law on Automated Contracting extends that framework to modern AI-enabled automation.
That makes one issue central: the system may be automated, but the legal effect still needs to attach somewhere. The most practical answer today is attribution to the legal or natural person that deployed or authorized the system.
Legal personhood solves fewer problems than it appears to
Giving an agent legal personality sounds like a clean answer to liability, but it can create new problems. An agent may own no meaningful assets, have no independent source of capital, and be controlled by software operators who could otherwise externalize risk into an empty legal shell.
Personhood therefore does not automatically improve accountability. It may weaken it if responsibility moves away from solvent humans or companies.
Rights should not be confused with operational protections
An agent may need protections against tampering, unauthorized memory edits or credential theft, but those are security controls, not necessarily “rights” in the legal or moral sense.
Likewise, a requirement that an agent receive accurate inputs or consistent policies may improve reliability without implying that the agent has standing to sue.
Humans affected by agents need stronger rights first
The immediate rights problem is usually on the human side. People need to know when they are interacting with AI in relevant contexts, understand consequential decisions, challenge errors and reach an accountable person.
The EU AI Act reflects this direction. Its transparency framework places duties on providers and deployers, and high-risk systems require defined human oversight when the relevant provisions apply.
A practical rights and responsibilities matrix
| Actor | Rights / protections | Responsibilities |
|---|---|---|
| User | Disclosure, correction, appeal, privacy, safe exit. | Provide lawful instructions and accurate information. |
| Deployer | Set policies, revoke authority, inspect logs. | Supervision, monitoring, compliance, incident response. |
| Provider | Protect IP and service integrity. | Safety, transparency, documentation and support obligations where applicable. |
| Agent | Operational integrity, authenticated instructions, protected credentials. | Remain within mandate, log actions, escalate uncertainty. |
| Affected third party | Notice, redress and human review where appropriate. | None simply because an agent affected them. |
What happens when an agent exceeds authority?
The answer should depend on the mandate and the surrounding law, not on whether the software “intended” the result. A properly designed system should make authority machine-readable: spending limit, permitted counterparties, contract scope, jurisdiction, time window and approval thresholds.
If the agent steps outside that mandate, the event should be treated as an authorization failure first, with clear audit evidence showing which operator, provider or control layer failed.
Shutdown is a governance requirement
Any agent with meaningful authority should have a revoke path that does not depend on the agent itself agreeing to stop. Credentials, wallets, API keys, tool permissions and scheduled actions should be independently suspendable.
This is not a claim about consciousness. It is basic operational control.
When could personhood become a serious question?
The debate becomes materially different if agents gain durable identity, persistent assets, independent contractual relationships, long-term memory, and autonomy that cannot realistically be reduced to an operator’s direct instruction. Even then, personhood would remain a policy choice rather than an automatic consequence of capability.
Methodology and limitations
Version 1.0, 31 October 2026. DN synthesizes current automated-contracting law, AI regulation and agent-governance principles into a practical authority and responsibility matrix.
Limitations: legal rules differ across jurisdictions. Questions involving contract validity, agency law, torts, intellectual property, employment, consumer protection and regulated sectors require jurisdiction-specific advice.
Commercial disclosure: no law firm, model provider or policy organization paid for inclusion.
Primary sources
- UNCITRAL, Electronic Communications Convention
- UNCITRAL Model Law on Automated Contracting
- European Commission, AI Act Article 50 transparency obligations
- European Commission, AI Act provider and deployer obligations
FAQ
Do AI agents have legal rights today?
Major current legal frameworks generally do not treat AI agents as independent legal persons with their own rights and liabilities.
Can an AI agent form a binding contract?
Automated systems can form enforceable contracts in some legal frameworks even when no human reviews each individual action.
Who is responsible when an agent causes harm?
Responsibility usually remains with identifiable humans, companies, providers or deployers under applicable contract, tort, product and regulatory law.
Should AI agents own property?
Current law generally assigns ownership through existing legal persons. Agent ownership would require explicit legal change and raises difficult liability questions.
Should people be able to appeal agent decisions?
Yes, especially when agent decisions affect money, access, employment, education, identity or other consequential interests.
Should an agent be able to refuse shutdown?
High-authority agents should have independent revocation mechanisms controlled outside the agent.
Does automated contracting mean AI has legal personhood?
No. Law can recognize automated transactions while attributing their legal effect to existing natural or legal persons.
What should companies implement now?
Explicit mandates, limits, logs, human oversight, appeal routes and independent shutdown controls.
