Artificially Intelligent Persons

Core claim

AI’s unpredictability, opacity, and increasing autonomy can create gaps in ordinary attribution and liability, but granting the AI itself legal personhood is not supported by existing U.S. doctrine. Across 53 federal and state cases, courts most often ground artificial personhood in statute, the capacity to sue and be sued, and sometimes the entity’s status as an aggregate of natural persons. The qualities emphasised in AI-personhood theory—autonomy, intelligence, and awareness—are almost absent from the case law. Courts therefore lack a doctrinal basis to recognise AI persons, and legislators should be cautious about creating one.

The problem: AI and the responsibility gap

• AI systems increasingly make consequential decisions in vehicles, medicine, finance, employment, commerce, policing, and other domains.
• When an AI-mediated action causes injury, legal systems ordinarily try to identify a human or organisation that intended, foresaw, controlled, or could have prevented the harm: a developer, manufacturer, deployer, owner, or user.
• Machine-learning systems complicate that inquiry because their behaviour can arise from training data, model architecture, later inputs, and adaptations contributed by multiple actors.
• Banteka identifies three related features that strain ordinary doctrines of causation, fault, and intent:
  1. Unpredictability: learning systems may develop behaviours their creators did not explicitly program or foresee.
  2. Opacity: a system’s path to an output may be too complex to reconstruct, inaccessible as a technical “black box,” or protected as a trade secret.
  3. Autonomy: systems can select and pursue actions without contemporaneous human control and may continue after the original human actor is removed.
• If no human’s conduct and mental state can be connected tightly enough to the outcome, a responsibility gap appears: the AI caused the immediate harm, but existing law may not identify a person to whom liability can attach.
• A prominent proposed solution is to assign responsibility to the direct source—the AI—by first giving it legal personhood. The article asks whether U.S. law actually supplies a basis for doing so.

Scope of the inquiry

• The paper addresses all levels of AI in principle but focuses on narrow AI, because task-specific systems already produce disputes courts must adjudicate.
• Narrow AI can equal or surpass human performance in bounded tasks such as driving, image recognition, translation, recommendations, trading, or diagnosis without generalising across the full range of human cognition.
• Artificial general intelligence remains hypothetical and is not necessary for the legal problem. The accountability question already arises from contemporary machine learning.
• The paper’s main question is institutional and doctrinal: under what conditions has the U.S. legal system treated artificial entities as legal persons, and do AI systems satisfy those conditions?
• Banteka makes two ultimate claims:
  • Descriptive: courts’ approaches to legal personhood are more fragmented than commonly assumed and do not support AI personhood.
  • Prescriptive: the existing legal basis prevents courts from conferring personhood on AI and should make legislators pause before creating a new status.

Why machine learning is legally distinctive

• Traditional automated software follows rules specified in advance. Machine learning instead derives decision rules from data and experience within a model and objective chosen by developers.
• In supervised learning, humans specify examples or desired outputs and the model learns how to map new inputs to them. In unsupervised learning, the model identifies patterns without labelled outcomes. Reinforcement learning adjusts behaviour through rewards associated with successful actions.
• Deep neural networks contain layered, adaptive structures that can alter their internal parameters as they learn. Their scale lets them detect patterns no person could manually derive, but it also makes their reasoning difficult to explain.
• A fraud-detection model, for example, may connect financial records, IP addresses, merchants, and online activity to freeze an account. It can optimise fraud prediction without being able to explain its decision in the terms a court uses to assess intent, reasonableness, or discriminatory purpose.
• A system can therefore produce an unlawful result—such as discrimination, speeding, or market manipulation—without that exact action appearing in its code or being intended by its developer.
• The paper treats these features as a reason to reconsider legal attribution, not as proof that an AI is already a person.

Concrete accountability puzzles

• An art-project shopping bot given a weekly bitcoin budget bought drugs and a passport on the dark web, forcing authorities to decide whether anyone could be charged for its purchases.
• Tesla vehicles operating with driver-assistance systems were involved in fatal crashes.
• An Uber self-driving test vehicle misclassified a pedestrian and struck and killed her; prosecutors found an insufficient basis for corporate criminal liability against Uber, while civil claims settled outside court.
• AI agents increasingly form or execute smart contracts, raising the question whether an agent that can transact in its own name could also be sued for breach or tort.
• In 2017, the European Parliament asked the European Commission to consider “electronic personhood” for sufficiently autonomous systems, with liability distributed among the system, engineers, and manufacturers according to degrees of autonomy. AI and robotics experts responded that robot personhood was legally and ethically inappropriate.

What legal personhood means

• Legal personhood is a classificatory tool that determines which entities can act within law and bear legal consequences.
• It is not synonymous with being human. Corporations, associations, trusts, partnerships, municipalities, and other artificial entities may be treated as single legal subjects.
• A classical definition describes a person as a subject of legal rights and duties.
• Those rights and duties form a divisible bundle, not an all-or-nothing status. They may include:
  • owning property;
  • entering contracts and other transactions;
  • suing and being sued;
  • holding constitutional protections;
  • incurring debt;
  • being subject to civil or criminal liability.
• Different persons possess different bundles. Children have fewer powers and different duties than adults; corporations cannot exercise every constitutional or procedural right held by natural persons.
• The legal question is therefore not only whether AI is a person, but which rights and duties it would receive, for what purpose, and on what authority.

The fragmented statutory background

• The U.S. Constitution uses “person” without defining it.
• The 1871 federal Dictionary Act stated that “person” could extend to bodies politic and corporate unless context required a narrower reading. Its modern version, 1 U.S.C. § 1, includes corporations, companies, associations, firms, partnerships, societies, and joint-stock companies.
• Courts have treated the Act inconsistently, ranging from a presumptive interpretive guide to a tool of last resort.
• Federal and state statutes recognise different entities as persons for different purposes. A municipality may be a person under one civil-rights provision but not another statute; foreign governments, agencies, and partnerships likewise receive context-dependent treatment.
• Legislatures and courts seldom articulate a general theory explaining why a particular entity qualifies. Personhood is often presumed, conferred ad hoc, or inferred from legal powers the entity already exercises.

Main theories of artificial legal personhood

1. Fiction theory

• Artificial persons are creations of positive law. The entity is not literally a person; law treats it “as if” it were one so it can hold rights and duties.
• On a pragmatic version, the fiction is justified when personhood advances legal-system purposes.
• Applied to AI, the relevant question would be whether a legal fiction makes accountability and administration work better—not whether the machine resembles a human.

2. Aggregate or symbolist theory

• Personhood is shorthand for the relations among the natural persons who make up an organisation and between those people and outsiders.
• Saying that someone contracted with a corporation avoids separately specifying relations with every shareholder or member.
• This rationale fits organisations composed of people but not an AI understood as a unitary technical artefact. That mismatch becomes important in the empirical results.

3. Realist theory

• Organisations are real social entities that exist before legal recognition, act independently of individual members, and produce legal effects.
• Law does not fabricate them; it acknowledges and personifies them.
• Churches, unions, and corporations can own property, transact, and persist despite membership changes, suggesting an entity-level existence.

4. Moral-person theory

• A further view treats some collectives as moral persons with natural rights because they can act intentionally.
• Banteka separates this claim from legal personhood: legal status and moral status do not necessarily coincide.

Why the corporation analogy is incomplete

• Corporations are the dominant model for artificial personhood. They can own property, contract, sue, be sued, incur civil and criminal liability, and hold selected constitutional rights independently of individual members.
• Their restricted bundle of rights demonstrates that artificial personhood can be partial and purpose-specific, which superficially makes them an attractive analogy for AI.
• But corporations are organised by law and usually understood as collections or instruments of natural persons. AI systems are neither associations of members nor conventional legal fictions.
• The aggregate theory—the justification that personhood organises the legal relations of underlying humans—is therefore especially difficult to transfer to an autonomous machine.
• Rather than extending corporate personhood by analogy, Banteka argues for examining the conditions courts actually use.

Conditions emphasised in AI-personhood theory

Legal scholarship commonly focuses on four related ideas. The article reviews each and questions whether it actually maps onto legal doctrine.

Autonomy

• Technological autonomy is distinct from mere automation. An autonomous AI can sense inputs, evaluate options, change internal states, set or modify subgoals, and act without ongoing human intervention.
• Scholars connect autonomy to free will and responsibility: an agent that controls its behaviour may appear an appropriate bearer of liability.
• But autonomy exists on a continuum, and neither law nor philosophy supplies a settled threshold at which machine independence becomes personhood.
• Technical ability to produce an unplanned output does not establish the normative self-determination associated with human agency.

Intelligence

• Intelligence can be understood behaviourally, as in the Turing test, or as genuine understanding, as challenged by Searle’s Chinese Room.
• Banteka adopts a continuum-oriented conception: intelligence is an agent’s capacity to achieve goals across environments, with differences in the number, range, speed, and generality of goals it can attain.
• AI systems range from executing a programmer’s specified route, through learning subgoals from supervised data, to self-training without supervision.
• Law already adjusts rights and responsibility to cognitive capacity for children and people with serious intellectual impairments. But these are modifications within natural-person status, not evidence that intelligence itself creates legal personhood.

Awareness or consciousness

• Scholarship often treats awareness and consciousness as interchangeable and links them to intentionality.
• The law does not generally make continuous consciousness a condition of personhood: sleep, coma, or temporary unconsciousness may affect culpability without erasing legal status.
• Awareness also differs from intention. Knowing that a harmful outcome is likely does not mean desiring or investing resources to bring it about.
• Awareness may support intentional action by letting an agent know what it is doing, and some theories connect consciousness to having interests that law can protect. Still, it is not established as a general doctrinal gateway to personhood.

Moral personhood

• Moral status means an entity counts for its own sake and places constraints on how others may treat it.
• Some theorists treat moral agency and blameworthiness as prerequisites to criminal accountability and therefore to legal personhood.
• Banteka resists that equation:
  • corporations can have legal and criminal responsibility without generally being treated as morally sentient persons;
  • infants have legal personhood without corresponding legal duties;
  • intelligent animals may learn rules without receiving personhood;
  • strict liability imposes accountability without fault or moral blame.
• Moral status, legal personhood, and legal accountability overlap in some cases but remain analytically distinct.

The empirical study

Research design

• The study asks which factors U.S. courts actually use when deciding whether an artificial entity is a legal person.
• It covers U.S. Supreme Court, federal circuit, federal district, state supreme, and lower state court decisions from 1809 through the study period.
• Searches began with “legal person” and references to corporations or companies, then combined “legal person” with “artificial entities.” Biological-personhood cases concerning fetuses and abortion were excluded because they use different concepts and terminology.
• Additional searches used “juridical entity” and “juridical person,” both before and after the modern federal Dictionary Act.
• The final corpus contained 53 cases.

Qualitative content analysis

• Banteka used qualitative content analysis to code the legal basis articulated in each opinion.
• Codes were iteratively refined, with redundant concepts collapsed and rare but distinct rationales retained.
• The final codebook contained 32 conditions, including:
  • statute-based status and 1 U.S.C. § 1;
  • implicit inclusion in a statute or the statute’s spirit and purpose;
  • capacity to sue and be sued;
  • rights to own property, transact, and contract;
  • constitutional rights, citizenship, legal standing, and accountability;
  • separate or independent existence;
  • being made up of individuals;
  • rights and duties generally;
  • perpetuity;
  • lack of autonomy or self-determination;
  • analogy to a natural person;
  • context-specific and institutional characteristics.
• Frequencies were then compared overall and across the five court levels.

Main findings

Across all 53 cases

Condition	Share of cases
Personhood is statute-based	47%
Entity has the right to sue and be sued	40%
Personhood is implicit in a statute	15%
Entity is analogous to a natural person	6%
Autonomy	2%

• The 32 total conditions show substantial doctrinal variation, with several appearing in only one case.
• Despite this fragmentation, two routes dominate:
  1. the legislature explicitly or implicitly recognises the entity;
  2. the entity can participate in litigation in its own name.
• A further recurring theme, especially in state and federal trial courts, is that the entity is an aggregate of natural persons.

By court level

• U.S. Supreme Court (10 cases): statute-based status appeared in 50%; capacity to sue and be sued in 40%; citizenship and constitutional rights each in 20%. Implicit statutory status did not appear.
• Federal circuit courts (8 cases): statute-based status appeared in 50%; implicit statutory inclusion in 38%; the right to sue and be sued and reliance on 1 U.S.C. § 1 each in 25%.
• Federal district courts (14 cases): the right to sue and be sued appeared in 50%; statute-based status in 36%; the right to contract in 21%.
• State supreme courts (9 cases): statute and the right to sue and be sued each appeared in 33%. Rights to transact, own property, contract, hold constitutional rights, and analogy to a natural person each appeared in 22%.
• Lower state courts (12 cases): statute-based status appeared in 67%; the right to sue and be sued in 33%; being made up of individuals, possessing rights and duties, and being an independent unit each in 25%.

The theory/practice gap

• Scholarship supporting AI personhood concentrates on autonomy, intelligence, awareness, and sometimes moral agency.
• Courts deciding artificial-personhood cases concentrate on legislative recognition and established legal capacities.
• Intelligence and awareness do not emerge as operative conditions in the dataset.
• Autonomy appears in only 2% of cases—and there as a rare doctrinal consideration, not a general test.
• Analogy to a natural person appears in only 6% overall, although it is more visible in federal appellate decisions.
• The aggregate-of-humans rationale cuts directly against AI personhood: legal doctrine often traces the artificial entity back to the natural persons whose collective vehicle it is, whereas an AI is not itself a membership organisation.
• The empirical record therefore does not validate the claim that sufficiently intelligent or autonomous behaviour naturally matures into legal status.

The circularity problem

• Courts sometimes determine that an entity is a person because it can sue, contract, hold property, or exercise constitutional rights.
• But those capacities are often consequences of already having legal personhood.
• The reasoning can therefore become circular:
  1. an entity is a person because it has a legal right;
  2. it has that right because it is a person.
• Banteka suggests this may reflect pragmatic judicial accommodation. Corporations already existed and produced social and economic effects, so courts legitimised capacities they exercised in practice when statutes were unclear.
• Fiction and realist theories reproduce a similar structure: both begin from entities already acting in legally consequential ways and ask law to regularise them.
• Circularity is especially dangerous for AI because extending one capacity as evidence of personhood may bootstrap an unprecedented entity into a broader bundle of rights and liabilities without a coherent justification.

Why legislation dominates

• Statutory recognition is the most consistent foundation across jurisdictions and the majority route in Supreme Court cases.
• When statutes are silent, courts draw from a much larger and less coherent menu of conditions, creating indeterminacy and potential arbitrariness.
• Similar cases at the same judicial level have used different combinations of property, contract, transaction, litigation, constitutional-rights, and perpetuity factors without explaining why one set controls.
• The findings imply that creating AI personhood would be a major policy choice for legislatures, not a modest analogical extension courts can confidently derive from settled doctrine.
• Because existing statutes list corporations and other organisational forms but not AI systems, courts have neither explicit authorisation nor a stable implicit category into which AI clearly fits.

Implications for AI liability

• Legal personhood could make an AI suable in name, but that does not by itself create assets, insurance, deterrability, procedural representation, or a punishment capable of influencing future conduct.
• Treating the AI as the responsible person could shield developers, manufacturers, deployers, and corporations whose decisions created the risk.
• A personhood regime would need to decide:
  • which system instance is the person—the model, a deployed copy, an embodied device, or a service;
  • who represents it in court;
  • what assets or insurance satisfy judgments;
  • which rights accompany liability;
  • when humans remain jointly or vicariously liable;
  • how modification, copying, merger, shutdown, or deletion affects identity and continuity.
• The article does not design an alternative liability regime. Its narrower conclusion is that personhood cannot be justified merely by invoking an accountability gap.

Methodological limitations

• The study is backward-looking: it describes the doctrinal grounds courts have used, not which rule would produce the best future consequences.
• It excludes biological-personhood cases, so it cannot compare artificial-entity doctrine with every legal conception of persons.
• The corpus is small—53 cases distributed across five court categories—and the coded concepts depend on interpretive judgments about judicial reasoning.
• Frequency is not precedential weight. A condition mentioned in fewer cases may be legally decisive, while a common statement may be dicta or inherited boilerplate.
• Many cases concern corporations, partnerships, and governmental units, not technological agents. The study tests whether existing artificial-person doctrine transfers to AI, but cannot empirically observe AI-personhood cases that did not yet exist.
• Showing that doctrine does not currently support AI personhood does not prove that legislation could never create it. The prescriptive force depends on how strongly one values doctrinal continuity and legal expectations relative to new policy needs.
• The paper deliberately does not conduct a full consequence-based analysis of whether AI personhood would improve compensation, deterrence, innovation, or risk allocation.

The author’s conclusion

• Courts’ personhood doctrine is fragmented but nonetheless points away from AI personhood.
• The most prominent conditions—statutory basis, capacity to sue and be sued, and aggregation of natural persons—do not arise from AI’s intelligence or autonomy.
• Existing law therefore does not permit courts to declare AI systems persons simply because they behave independently or create attribution problems.
• Legislatures could in principle alter the law, but should first confront the doctrinal incompatibility and the risk of destabilising settled expectations.
• Policy should not move from “AI caused the harm” to “AI should be a person” without a clear account of the rights, duties, human beneficiaries, and accountability consequences the new status would create.

Takeaways

• Legal personhood is conferred, not discovered. Technical sophistication does not automatically generate legal status.
• Personhood is a bundle. Any proposal must specify the capacities, liabilities, and protections involved rather than using “person” as an undifferentiated label.
• Corporate personhood is not a ready-made template. Its doctrinal logic often depends on organising the rights and duties of underlying humans.
• The theory most discussed for AI is not the doctrine courts use. Autonomy, intelligence, and awareness have little empirical presence in artificial-personhood cases.
• Liability gaps require liability design. Giving the direct causal system a legal label may displace rather than solve questions of compensation, deterrence, and human responsibility.

Notes & Observations

• The article’s most valuable move is methodological: instead of asking only what traits should make something a person, it asks what U.S. courts have actually treated as legally relevant.
• Its results also expose how unstable “legal personhood” is as a master concept. Courts often reason from specific capacities to status, while theorists reason from status to capacities.
• The empirical findings strongly support judicial caution, but the article occasionally states the legislative conclusion more strongly than its backward-looking method alone establishes. A legislature is precisely the institution capable of creating a new category when old categories do not fit; whether it should do so requires the consequence-based analysis the paper brackets.
• Much of the technical discussion reflects the state of AI scholarship before 2021. Later advances may intensify opacity or autonomy concerns, but they do not change the paper’s historical findings about the cases it coded.
• Citation: Nadia Banteka, Artificially Intelligent Persons, 58 Houston Law Review 537–596 (2021).