137AI > Governance > Criminal Law & AI Unsettled Categories

Criminal Law & AI Categories

Criminal law was written for human actors with the capacity to form intent, take action, and bear culpability. The categories that organize the field — robbery, theft, assault, trespass, fraud, conspiracy, computer crime — were defined with human actors in mind. The application of these categories to autonomous and ambient AI agents raises questions that are unsettled across most jurisdictions, and the answers are being worked out through litigation, regulatory commentary, and the early case law that is beginning to accumulate as incidents reach the courts.

This page catalogs the unsettled questions systematically by agent category. The questions are real, the answers will affect operators, manufacturers, and the people affected by autonomous and ambient agents, and the absence of settled doctrine is itself a governance problem. The broader analytical frame for autonomous physical agents as a regulatory category appears in Autonomous Physical Agents as a Regulatory Category.

The Structural Problem

Most criminal statutes assume the actor whose conduct is at issue is a human. The elements of an offense typically include conduct, intent, and circumstances, with each element evaluated for the human defendant. Autonomous AI agents complicate every element.

The conduct is partly the agent's behavior, partly the operator's instructions, and partly the design and training choices made by the manufacturer and software vendors. Allocating the conduct element across these parties is not straightforward.

The intent element is harder still. Conventional criminal law treats intent as a human mental state. An AI agent does not have intent in the legal sense. The operator may have intent, but the operator's intent may not extend to the specific act the agent performed. The manufacturer may have intent in the design choices but not in any specific act.

The circumstances element often turns on facts that an agent perceives differently than a human would. Whether a building is occupied, whether a person consented, whether a vehicle was abandoned — these are facts that humans evaluate through context and judgment. An agent evaluates them through perception, models, and policy, with consequences that may differ from the human-trained legal evaluation.

Each of these structural complications shows up in specific unsettled questions across agent categories.

Physical Agent Unsettled Questions

Physical agents create the largest population of unsettled criminal law questions because their conduct can directly cause physical harm, property damage, or unauthorized transport.

Question	What Is Unsettled	Why It Matters
Robbery versus theft for humanoid-conducted property crime	Whether a humanoid's physical presence and manipulation capability constitutes the force or threat that distinguishes robbery from theft	Sentencing exposure, prosecutorial discretion, and insurance coverage all turn on which characterization applies
Assault liability for autonomous-agent-conducted physical harm	Whether and how assault statutes apply when a physical agent applies force under operator direction, hijacked control, or autonomous decision	Allocates criminal responsibility between operator, attacker (in compromise cases), and manufacturer
Trespass and unauthorized entry by autonomous agents	Whether autonomous entry into restricted areas constitutes trespass when no human is present in the agent and the agent's behavior was directed remotely	Affects how facility access controls, residential security, and public-space law respond to autonomous-agent presence
Vehicle as instrumentality versus vehicle as defendant	Whether autonomous vehicles used in crimes are treated as tools of the operator or as separate evidentiary or forfeiture subjects	Determines civil forfeiture exposure, evidence preservation requirements, and operator liability allocation
Reckless and negligent homicide for autonomous-vehicle fatalities	When an autonomous vehicle kills a person, whether reckless or negligent homicide charges can be sustained against the operator, manufacturer, or anyone in the chain	The Tempe Uber case resulted in a negligent homicide plea by the safety driver; broader application is uncertain
Weapons offenses for autonomous-agent-mediated violence	When an autonomous agent uses an ordinary object as a weapon, whether weapons enhancement statutes apply	Affects sentencing exposure for both operators directing autonomous agents and attackers compromising them
Aggravated offenses involving vulnerable victims	Whether enhancements for offenses against children, elderly, or disabled victims apply when an autonomous agent is the immediate actor	Vulnerable-victim enhancements substantially increase penalties; their application to autonomous-agent cases is not yet established

Personal and Ambient Agent Unsettled Questions

Personal and ambient agents raise distinct criminal law questions because they capture continuously and operate in private contexts where recording-consent and surveillance statutes overlap with criminal law.

Question	What Is Unsettled	Why It Matters
Wiretap statute application to AI wearables and assistants	Whether always-on AI agent capture in shared space constitutes interception under federal and state wiretap law	Wiretap violations carry criminal penalties; the applicability to AI wearables and smart home assistants is being worked out
Recording-consent law for ambient AI	In states requiring all-party consent for recording, whether AI wearable or assistant capture of bystanders meets the consent requirement	Recording-consent violations are criminal in several jurisdictions; the standard of consent for ambient AI capture is not yet defined
Stalking and harassment via AI agents	Whether using AI agents to track, monitor, or contact a victim constitutes stalking or harassment under existing statutes	Many stalking and harassment statutes were drafted assuming direct human conduct; AI-mediated patterns may not fit cleanly
Non-consensual intimate imagery from AI capture	When AI agents capture intimate material that is then distributed, whether NCII statutes apply to the operator, the platform, or the capturing party	NCII statutes typically require intent and identifiable conduct; AI-mediated capture and distribution complicate both
Biometric privacy law enforcement with criminal penalties	In jurisdictions with criminal penalties for biometric privacy violations, whether AI wearable and assistant capture meets the criminal threshold	Illinois, Texas, and other biometric privacy frameworks have civil penalties; expanded criminal applicability is being considered
Conversational AI and harassment by automated systems	Whether automated harassing or threatening messages from AI agents constitute criminal harassment under statutes designed for human conduct	Several Character.AI and similar platform incidents have raised this question; case law is early

Software Agent Unsettled Questions

Software agents with autonomous action authority raise criminal law questions at the intersection of computer fraud, financial crime, fraud, and emerging agentic AI accountability.

Question	What Is Unsettled	Why It Matters
Computer Fraud and Abuse Act application to agentic AI	Whether unauthorized access via prompt injection or agent compromise constitutes CFAA violations, and who is liable when the agent itself was the immediate accessor	CFAA is the primary federal computer crime statute; its application to AI agent compromise is not yet established
Fraud liability for AI-conducted transactions	When a software agent is induced to complete fraudulent transactions through prompt injection, whether fraud statutes apply to the inducer, the operator, or both	The Air Canada chatbot case held the operator accountable for the agent's representations; criminal fraud application is broader and less settled
Wire fraud and AI-mediated deception	Whether AI-generated deceptive content used in fraud schemes triggers wire fraud enhancements, and what intent showing is required	Wire fraud is a major federal charge; AI-generated content as the wire vehicle is increasingly common
Identity theft via AI impersonation	Whether AI voice cloning, video synthesis, or persona impersonation meets the elements of identity theft statutes	Identity theft statutes typically require use of identifying information; AI synthesis of identifying characteristics may or may not qualify
Money laundering through AI-mediated transactions	When AI agents are used to structure or conceal transactions, whether AML criminal statutes apply to operators and users	AML criminal penalties are substantial; AI-mediated structuring creates new evasion patterns that statutes may not address
Securities law and algorithmic manipulation	When AI trading agents engage in market manipulation, whether the operator, the AI vendor, or both face criminal securities liability	SEC and CFTC have civil enforcement experience; criminal allocation is less developed for AI-mediated activity

Cross-Cutting Doctrinal Questions

Several criminal law questions cut across all agent categories rather than belonging to any single one.

Question	What Is Unsettled	Why It Matters
Mens rea for AI-mediated offenses	When the immediate actor is an AI agent, how the mens rea requirement is satisfied for the human or organizational defendant	Mens rea is a fundamental element of most criminal offenses; its application in AI-mediated cases is not yet established
Causation and attribution in AI cases	Proving that the defendant's conduct caused the harm when the immediate cause was AI agent behavior	Causation is required for most offenses; AI agent behavior between defendant conduct and harm complicates the chain
Vicarious liability for AI operator conduct	When an organization deploys AI agents that commit acts the organization did not specifically authorize, whether the organization bears criminal responsibility	Vicarious liability doctrine varies by jurisdiction; its application to AI deployment is being worked out
Accessory and conspiracy liability	Whether AI vendors, training data providers, or platform operators face accessory or conspiracy liability when their components are used in criminal acts	Could extend criminal exposure substantially up the AI supply chain
Self-defense and necessity involving autonomous agents	Whether self-defense or necessity defenses apply when a person damages or destroys an AI agent that was threatening or harming them	Property damage statutes generally protect property; defenses for damaging a malfunctioning or weaponized AI agent are uncertain
Search and seizure for autonomous agents and their data	Fourth Amendment and equivalent doctrines governing law enforcement access to AI agent telemetry, internal state, and captured material	Investigation of AI-mediated crimes requires access to AI agent data; the constitutional framework is not yet settled

How the Law Is Being Constructed

The unsettled questions catalogued above are being resolved through several paths simultaneously, none of which is producing settled doctrine quickly.

Litigation is producing case-by-case precedent. The Air Canada chatbot tribunal ruling held operators accountable for agent representations in a civil context. The Tempe Uber case produced a negligent homicide plea by the safety driver. The various ChatGPT fake-citation sanctions established that attorneys remain responsible for AI-generated content in their filings. Each case produces a narrow holding that future cases extend, distinguish, or apply.

Regulatory commentary is shaping the framework before legislation arrives. NHTSA, FTC, EEOC, SEC, CFPB, and equivalent bodies in other jurisdictions are taking enforcement actions and issuing guidance that shapes the criminal-civil boundary. The EU AI Act provides framework but does not address criminal application directly.

State legislation is moving on specific topics where state authority is clear. NCII statutes have been amended in several states to address AI-generated content. Biometric privacy laws have been extended. State autonomous vehicle statutes address some criminal questions implicitly. The patchwork is uneven.

Federal legislation has been proposed but rarely enacted. Bills addressing AI-mediated fraud, deepfake identity offenses, and AI accountability appear regularly in Congress but rarely become law. The pace of federal criminal law reform is slower than the pace of AI deployment.

Academic and policy analysis is producing the analytical groundwork. Law schools and policy institutes have begun publishing systematic analysis of the unsettled questions, with varying influence on legislative and judicial development.

Practical Implications for Operators, Insurers, and Affected Parties

For operators, the unsettled criminal law landscape means that defensive planning has to anticipate adverse interpretations of statutes that have not yet been authoritatively applied. Operating practices that would be defensible under the most-likely interpretation may still create exposure under a more aggressive interpretation that a prosecutor or court may adopt in a specific case.

For insurers, the criminal exposure of operators and manufacturers is a coverage question that interacts with the civil liability landscape. Criminal acts are typically excluded from insurance coverage; the boundary between negligence and criminal recklessness in AI deployment will be contested.

For affected parties, the criminal law uncertainty means that available remedies are partial. Criminal prosecution of AI-mediated harm depends on prosecutors being able to fit the conduct into statutes that may not cleanly apply. Civil remedies may be more readily available but produce different outcomes than criminal accountability.

For law enforcement, the investigation and charging frameworks for AI-mediated offenses are still being developed. Prosecutors are working out how to charge cases, judges are working out how to apply statutes, and the precedent that results shapes practice going forward.

The Reframe

Criminal law is being reconstructed for autonomous and ambient AI agents one case at a time. The questions catalogued here are real, the answers will affect the entire ecosystem, and the absence of settled doctrine is itself a governance condition. The framework that emerges will combine extensions of existing statutes, new legislation, judicial interpretation, and regulatory enforcement, with the timing and direction uneven across jurisdictions. Operators, insurers, and affected parties navigate the current period with limited certainty about how specific conduct will be treated, and the construction of clearer doctrine is one of the substantial legal projects the autonomous and ambient AI agent ecosystem requires.

Related Coverage

Governance | Autonomous Physical Agents as a Regulatory Category | Liability & Product Law | Criminal Misuse & Autonomous Crime Economy