In the United States, and in New York, artificial intelligence is still governed primarily through existing law rather than through a single, general AI code. The practical framework is built from contract, tort, product liability, privacy, intellectual property, anti-discrimination, consumer-protection, securities, employment, and sector-specific rules. That is why AI analysis remains use-case specific. A document-review tool, a hiring screen, a chatbot, and an autonomous agent may all involve AI, but they raise very different legal issues.

Contract law is often the first line of governance. Enterprise AI agreements now address data rights, confidentiality, security, acceptable use, audit rights, performance commitments, model changes, retention, IP allocation, and indemnities. Tort and product-liability doctrines also continue to apply, especially where AI outputs can foreseeably cause economic, reputational, or physical harm. Privacy and data-protection law remains central because training, retrieval, monitoring, and agentic workflows can expose personal data, confidential business information, or third-party data subject to contractual restrictions. IP law is equally important, particularly for training data, outputs, source code, trade secrets, publicity rights, and branding.

New York adds several targeted overlays. These include the Stop Hacks and Improve Electronic Data Security Act (SHIELD Act), New York Department of Financial Services (NYDFS) guidance and regulation for financial services and insurance, New York City’s Local Law 144, New York City’s biometric notice requirements for certain commercial establishments, the state’s AI companion law, the Responsible AI Safety and Education Act (RAISE Act) for covered frontier-model developers, which will come into full effect on 1 January 2027, and the state’s Civil Rights Law publicity regime, including recent digital-replica and post-mortem amendments. New York has also enacted a synthetic-performer advertising disclosure law, although that statute is not effective until 9 June 2026. Agentic AI does not create a separate legal category, but it increases familiar risks around authorisation, supervision, recordkeeping, and explainability.

Commercial use of AI is now pervasive across financial services, healthcare, retail, media, software, logistics, manufacturing, and professional services. Most deployments fall into three broad buckets:

• predictive systems that score or rank;
• generative systems that create text, code, images, audio, or video; and
• increasingly, agentic systems that can take multi-step actions using tools or connected enterprise systems.

Businesses are using these tools both internally and externally, including for search, drafting, summarisation, customer service, fraud detection, underwriting, software development, surveillance, pricing, marketing, and workflow automation.

Legally, the most important question is not whether a system is labelled AI, but whether it is used in a consequential context. Uses affecting employment, credit, insurance, housing, health, safety, public communications, or regulated advice receive the most scrutiny. The safest approach is to treat AI as an enterprise-risk issue, not just a technology issue. Companies should inventory use cases, classify risk, test for accuracy and bias, establish human-review triggers, document known limitations, monitor vendors, and preserve logs sufficient to investigate disputed outputs or actions.

Governments are deeply involved in AI innovation as funders, purchasers, regulators, and users. At the federal level, AI development is supported through research funding, procurement, semiconductor and compute policy, and national infrastructure initiatives. Programmes associated with the CHIPS and Science Act, National Science Foundation’s National AI Research Institutes, and the National AI Research Resource illustrate how industrial policy and research policy now shape the AI market.

New York has taken a particularly active innovation role through Empire AI, a public-interest compute and research consortium. Even where government measures are framed as innovation policy rather than regulation, they still affect private companies because public grants and procurement vehicles often carry documentation, cybersecurity, civil-rights, and performance expectations. For many companies, the government is therefore both a market opportunity and a source of de facto AI standards.

The United States has generally favoured a targeted, sectoral approach over a single omnibus AI law. As a result, AI-specific legislation has tended to focus on identifiable risks, such as hiring bias, synthetic media, children’s safety, robocalls, insurance discrimination, public-sector use, or consumer disclosure, while leaving most background obligations to existing law. That approach is likely to continue in the near term, even as a few states experiment with broader frameworks.

For companies, the consequence is practical rather than theoretical: compliance cannot be reduced to checking one AI statute. It requires a matrix that combines existing law with a growing set of AI-specific overlays at the federal, state, and local levels.

At the federal level, there is still no generally applicable US AI act. Instead, federal law reaches AI through technology-neutral statutes and sector regimes, including the Federal Trade Commission (FTC) Act, civil-rights statutes, the Fair Credit Reporting Act (FCRA) and Equal Credit Opportunity Act (ECOA), the Telephone Consumer Protection Act (TCPA), securities laws, product-liability doctrines, copyright and trade mark law, healthcare regulation, export controls, and agency-specific supervisory frameworks. Federal policy has also shifted over time between safety-oriented and innovation-oriented emphases, which means businesses should focus less on broad policy rhetoric and more on concrete agency requirements.

At the state level, the picture is becoming more heterogeneous. Some states have adopted narrow AI laws focused on disclosure or specific harms; others are moving toward broader risk-management duties. A prominent example of a comprehensive law governing automated decision-making technology is Colorado’s Artificial Intelligence Act (SB 189), signed into law on 14 May 2026, which repealed and replaced the state’s original artificial intelligence law (SB 205). Whereas the original Act focused primarily on AI bias and risk assessment, the amended Act adopts a notice-and-rights framework for automated decision-making used in consequential decisions involving employment, housing, finance, insurance, healthcare, education, and government services. The revised law will take effect 1 January 2027. Utah has taken a different approach, emphasising transparency and regulatory experimentation.

New York does not yet have a single statewide omnibus AI act. Its approach remains targeted, but it now includes the RAISE Act, a significant statewide law aimed at covered frontier-model developers. Other important New York rules include the state’s AI companion law, existing privacy and cybersecurity statutes, NYDFS requirements and guidance, the Civil Rights Law’s publicity protections, recent digital-replica amendments, New York City’s Local Law 144, and New York City’s biometric notice regime. The synthetic-performer advertising disclosure law should also be watched closely, although it is not effective until 9 June 2026.

In practice, nonbinding directives are often as important as legislation. Cross-sector governance references, the Office of Management and Budget (OMB) Memoranda M-25-21 and M-25-22 for federal agencies, and sector-specific guidance all shape expectations before a court ever decides a case. Agencies including the FTC, Equal Employment Opportunity Commission (EEOC), US Department of Housing and Urban Development (HUD), Consumer Financial Protection Bureau (CFPB), US Food and Drug Administration (FDA), Federal Communications Commission (FCC), Securities Exchange Commission (SEC), Commodity Futures Trading Commission (CFTC), and Financial Industry Regulation Authority (FINRA) have all issued AI-relevant guidance through existing authorities.

In New York, NYDFS Circular Letter No. 7 is especially important for insurers using AI or external consumer data in underwriting and pricing. State and court internal AI policies also matter for vendors and other counterparties dealing with government entities. These instruments may not be binding in the same way as a statute or regulation, but they often define what regulators, procurement officials, and courts will view as reasonable practice.

The EU AI Act is not US law, but it is highly relevant to many US and New York businesses. It applies on a functional and territorial basis that can reach providers and deployers serving the EU market, and it creates a risk-based structure with heightened duties for certain systems and uses. Even where it does not apply directly, it is already influencing customer due diligence, contract drafting, technical documentation, vendor questionnaires, and board-level expectations.

For US companies, the main mistake is to treat the EU AI Act either as irrelevant or as universally controlling. It is neither. The better view is that it is an important cross-border overlay and a strong market signal, especially for global businesses, but it does not displace US federal and New York law.

State law is now one of the fastest-moving parts of US AI regulation. The main trend is not toward a uniform national model, but toward a patchwork of targeted state and local obligations. These include laws addressing automated hiring tools, deepfakes and election content, AI transparency, insurance discrimination, child safety, privacy, and publicity rights. A few states are also adopting broader process-based duties for high-risk systems.

For companies operating nationally, that means AI governance must be state-aware from the outset. New York is important not because it has enacted an omnibus AI law, but because it combines financial-services oversight, aggressive consumer protection, local employment regulation, publicity rights, and recent AI-specific statutes, including the AI companion law and the RAISE Act, in ways that can materially affect deployment decisions.

Many AI disputes arise first through data, information, or content law rather than through a statute labelled AI. Training, fine-tuning, retrieval-augmented generation, session replay, chatbots, monitoring tools, and synthetic media can all trigger privacy, eavesdropping, trade-secret, confidentiality, copyright, publicity, records-management, and deceptive-practices issues. The central questions are often whether data was lawfully sourced, whether downstream use was adequately disclosed or authorised, and whether outputs or model behaviour interfere with someone else’s legal rights.

This is particularly important for generative and agentic systems because they often combine multiple layers of input and output data. A business may be handling user prompts, enterprise data, retrieved third-party content, training datasets, model logs, and generated outputs at the same time. Sound governance therefore requires provenance controls, contractual use restrictions, retention limits, and escalation paths for sensitive or legally restricted data.

AI-specific legislative activity remains substantial at both the federal and state levels, but it is also volatile. Bills continue to appear on issues such as deepfakes, digital replicas, child safety, political communications, automated decision-making, and government procurement. New York remains especially active, but several topics that would previously have sat in a ‘proposed’ bucket are now enacted law, including the RAISE Act, the AI companion law, the synthetic-performer advertising disclosure law, and the 2025 amendments expanding post-mortem and digital-replica publicity protections.

The prudent assumption is that targeted obligations will continue to expand, even if comprehensive legislation remains uneven. Businesses should therefore design governance systems that can absorb new disclosure, documentation, and human-review requirements without requiring wholesale redesign each time a new bill becomes law.

US courts are beginning to shape AI law, but the case law remains early and highly fact-specific. In legal practice, Mata v Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023) remains the clearest warning that lawyers, not software, are responsible for filings that contain fabricated authorities. In copyright, Thomson Reuters Enter. Ctr. GmbH v Ross Intelligence, 765 F. Supp. 3d 382 (D. Del. 2025) is an important decision because it rejected a fair-use defence on the facts presented in a non-generative training dispute. Other generative AI copyright cases remain active and are producing mixed procedural and merits rulings, particularly around training data, output similarity, and Digital Millennium Copyright Act (DMCA) theories.

Outside IP, courts are also confronting AI through older doctrines involving surveillance, privacy, discrimination, misleading statements, and due process. The key point is that judges have not been waiting for a new body of AI common law to emerge. They are applying existing rules to new technologies. For companies, that means documentation, provenance, disclosures, human review, and control over deployment facts will often matter more than abstract debates about whether a system is truly intelligent.

No single US regulator has plenary authority over AI. Oversight is distributed among existing agencies applying their traditional mandates to AI-related conduct. The FTC is the broadest general enforcer in the consumer and commercial space, especially for deceptive claims, unfair practices, and hidden or harmful AI use. Other key agencies include the EEOC, US Department of Labor (DOL), HUD, and CFPB for employment, housing, and credit; the FDA and US Department of Health and Human Services (HHS) for healthcare and medical-device contexts; the FCC for AI-generated robocalls; National Highway Traffic Safety Administration (NHTSA) and Department of Transportation (DOT) for automated driving; and the SEC, CFTC, and FINRA for financial-services uses.

Other institutions matter even where they are not classic enforcement bodies. OMB governs federal use and procurement of AI. The Department of Justice’s (DOJ) interests span fraud, civil rights, antitrust, national security, and corporate compliance. In New York, NYDFS is the leading state regulator for financial services and insurance, the Attorney General can use general consumer-protection and civil-enforcement tools, and New York City’s Department of Consumer and Worker Protection enforces Local Law 144.

AI guidance has become one of the most important sources of practical law in the United States. OMB’s federal memoranda, FTC guidance on deceptive AI claims, EEOC and HUD guidance on automated decision-making, FDA materials on AI-enabled products, FCC action on AI-generated voices, and FINRA and CFTC guidance in financial services all shape expectations before a court ever decides a case.

In New York, NYDFS guidance is especially influential because it translates general statutory obligations into concrete expectations around governance, testing, proxy discrimination, and vendor oversight. The result is that sophisticated companies increasingly treat major AI guidance as operationally mandatory, even where the instrument is formally nonbinding.

Enforcement activity has focused less on futuristic theories and more on familiar misconduct in an AI wrapper. The FTC has pursued deceptive AI claims, including through Operation AI Comply and the DoNotPay matter, making clear that AI marketing claims require substantiation like any other advertising claim. The SEC has likewise brought “AI washing” cases where investment advisers or public-facing firms overstated their use of AI or misdescribed what their systems could do. The FCC has also made synthetic voice misuse a high-priority issue under the TCPA.

The enforcement lesson is straightforward. Regulators care about substantiation, bias, recordkeeping, supervision, consumer harm, and whether humans remain accountable for consequential decisions. In New York, public enforcement is still likely to run through existing financial-services, employment, privacy, and consumer-protection authorities, with Local Law 144 providing an especially concrete hook for employment-related uses.

In the United States, National Institute of Standards and Technology (NIST) remains the most visible AI standard-setting institution, even though its core AI materials are voluntary. In practice, many enterprises use those materials as a reference point rather than trying to implement them line by line. American National Standards Institute-accredited standards and sector-specific technical bodies also matter, particularly where procurement, certification, or regulated products are involved.

For businesses, the importance of these standards is usually indirect but significant. They shape vendor questionnaires, procurement criteria, regulator expectations, board reporting, and what later looks like “reasonable” care in litigation. In that sense, voluntary standards increasingly operate as the baseline against which hard-law compliance is measured.

Internationally, ISO/IEC standards, OECD principles, and other multilateral or transnational frameworks are influencing AI governance, especially for multinational companies. These materials often address the same themes as US guidance, including risk management, transparency, human oversight, robustness, security, and accountability.

For US and New York businesses, international standards matter most through contracts, customer due diligence, and global programme design. They rarely displace domestic law, but they often determine what global counterparties will expect from an AI governance programme.

Government agencies are using AI in internal operations and public-facing functions, including document management, translation, fraud detection, cybersecurity, case triage, research, procurement, and constituent services. These uses can create efficiency gains, but they also raise concerns about due process, discrimination, transparency, records retention, public access, procurement compliance, and explainability. In New York, state law now also requires an AI inventory for state agencies, which reinforces transparency and governance expectations around public-sector use.

Public-sector use becomes most legally sensitive when AI influences benefits, enforcement, adjudication, hiring, surveillance, or other decisions affecting rights or entitlements. Agencies and vendors should therefore expect heightened documentation, testing, and escalation obligations for high-impact uses.

Case law directly addressing government use of AI remains limited. Courts have more often confronted adjacent issues, such as algorithmic scoring, surveillance technologies, opaque decision tools, and challenges based on due process or discrimination theories. New York-specific appellate law on government AI use remains sparse.

The stronger trend is institutional rather than doctrinal, but there are already useful warnings for legal practice. In United States v Heppner, No. 25-cr-00503-JSR, 2026 WL 436479 (S.D.N.Y. Feb. 17, 2026), Judge Rakoff held that documents a defendant created using Anthropic’s Claude were not protected by attorney-client privilege or the work-product doctrine where the materials were generated through a third-party AI platform and not created at counsel’s direction. More broadly, courts increasingly expect transparency about AI use in litigation and are beginning to scrutinise how AI affects evidentiary reliability, professional responsibility, and the administration of justice.

AI has become a national-security issue as well as a commercial one. Export controls, sanctions, semiconductor policy, defence procurement, cybersecurity, and investment screening all matter for certain AI models, chips, data flows, and customers. Large model development and deployment can also implicate restricted-party screening, cloud and compute access, and controls on technical transfers.

These issues are not limited to defence contractors. Commercial companies may encounter national-security constraints through supply-chain counterparties, foreign customer relationships, open-weight releases, model-sharing, or data access. Governance programmes for advanced AI therefore need a path for export-control, sanctions, and geopolitical review, not just privacy and consumer-protection review.

Generative AI has concentrated legal attention because it can create persuasive but inaccurate, infringing, confidential, biased, or misleading outputs at scale. The main legal issues are now familiar: hallucinations and reliance risk, copyright and publicity claims, privacy and confidentiality breaches, deceptive marketing, discrimination, recordkeeping, model drift, and vendor opacity. Many of these issues also intensify when the system can autonomously retrieve information, invoke tools, or take downstream action.

The US regulatory response remains fragmented and use-case dependent. The FTC addresses deceptive or unfair conduct; the SEC, CFTC, and FINRA address regulated financial uses; the FCC addresses AI-generated voices in robocalls; employment, housing, and credit regulators address algorithmic bias; and IP or publicity doctrines address unauthorised use of protected content or likenesses. New York adds important specific overlays, including the AI companion law, synthetic-performer advertising disclosures, Local Law 144 where employment decisions are implicated, and the state’s publicity and privacy rules. The compliance lesson is that generative AI cannot be governed solely as a content tool. It must be governed as a legal-risk amplifier.

AI use in legal practice is no longer novel, but the profession’s core duties have not changed. Lawyers remain responsible for competence, confidentiality, privilege protection, supervision of staff and vendors, candour to courts, and the accuracy of work product. AI can improve research, drafting, e-discovery, knowledge management, contract review, and administrative efficiency, but those benefits do not dilute professional responsibility. The American Bar Association’s Formal Opinion 512 is now the clearest national ethics reference. It emphasises competence, confidentiality, client communication where appropriate, supervision, candour to tribunals, meritorious positions, and reasonable fees, and it treats AI as a tool rather than a substitute for legal judgment.

The main legal-tech risks are hallucinated authorities, disclosure of client confidences, failure to verify outputs, overreliance on opaque tools, and insufficient vendor diligence. Law firms and legal departments should therefore use approved tools, limit sensitive inputs, validate legal authorities and quotations, document review protocols, and train users on when AI output is not an acceptable substitute for legal judgment.

The principal theories of liability for AI are still ordinary ones: breach of contract, negligence, product liability, negligent misrepresentation, fraud, defamation, discrimination, privacy violations, consumer protection, professional malpractice, and IP infringement. Which theory is most relevant depends on the context, the nature of the harm, the parties’ contractual allocation of risk, and how much control a defendant had over design, training, deployment, warnings, monitoring, and post-incident response.

For generative and agentic systems, the hardest issues often involve foreseeability and control. If a business exposes customers or employees to a model it knows can hallucinate, leak data, discriminate, or act outside intended bounds, plaintiffs and regulators will frame those failures using traditional doctrines rather than waiting for a new AI liability rule.

The United States does not yet have a general statutory liability regime written specifically for AI. Instead, regulators are imposing process obligations in targeted areas and using existing enforcement tools to police conduct that involves AI. That makes liability analysis highly contextual.

In practice, the most important liability-allocation instruments remain governance, documentation, and contract. Vendors and customers increasingly negotiate detailed provisions on data provenance, testing, audit rights, output restrictions, incident response, indemnities, and human review because those provisions often determine where liability lands after a failure.

Agentic AI is governed by the same legal framework that applies to other AI, but the level of autonomy changes the operational analysis. When a system can call tools, access enterprise systems, communicate externally, spend money, write code, or complete multi-step workflows, it raises sharper questions about authorisation, identity and access management, change control, auditability, and who is responsible when something goes wrong.

For that reason, agentic AI should be governed more like a privileged actor than like a passive software feature. Good controls include narrow permissions, human approvals for high-risk actions, environment segmentation, kill switches, transaction limits, prompt and tool logging, exception handling, incident-response playbooks, and testing that evaluates not just one output, but multi-step behaviour over time. Contracts should also address agent boundaries, override rights, data handling, and the vendor’s obligations to preserve logs and notify customers of material changes.

Autonomy does not create legal personhood for AI systems. Responsibility remains with natural persons and legal entities: developers, deployers, integrators, employers, and other actors with relevant control or benefit. The difficult question is not whether the AI is liable, but how liability should be allocated among the human parties in the stack.

That allocation will turn on facts such as training choices, warnings, deployment settings, supervision, permissions, and contract language. The more a system is allowed to act independently in consequential contexts, the more important it becomes to define those responsibilities ex ante.

Bias and fairness remain central AI issues because many legally significant AI uses involve ranking, recommendation, screening, or prediction. Employment, housing, credit, insurance, healthcare, education, and public-sector decisions all raise heightened concerns where model behaviour may create disparate treatment or disparate impact, including through proxies for protected traits.

In the United States, these issues are still governed mainly through existing anti-discrimination and unfairness law, supplemented by targeted rules such as Local Law 144 and Colorado’s in-force insurance anti-discrimination framework, including the regulation governing life insurers’ use of external consumer data and information sources, algorithms, and predictive models, which directly targets unfair discrimination. The defensible path is not to promise perfect neutrality, but to document data governance, testing, monitoring, proxy analysis, human review, and complaint or appeal mechanisms.

Biometric AI raises unusually high legal and reputational risk because it deals with sensitive identifiers and frequently intersects with surveillance, consent, discrimination, and accuracy concerns. US regulation remains fragmented, with important roles for state privacy law, unfairness law, sector regulation, constitutional doctrine in the public sector, and targeted local rules.

In New York, companies should account at least for New York City’s biometric notice requirements for certain commercial establishments, employment and privacy risks, and the state’s broader publicity and consumer-protection framework. Emotion-recognition systems deserve additional caution because their scientific validity is contested and because they can easily be challenged as intrusive, unreliable, or biased when used in consequential settings.

Deepfakes and synthetic media are regulated in the United States through a mix of publicity, fraud, defamation, election, child-safety, platform, and consumer-protection law. The legal issue is often not the synthetic medium by itself, but whether it misleads, impersonates, exploits a protected likeness, invades privacy, or facilitates another unlawful act.

New York is especially important here. Its Civil Rights Law already protects certain publicity interests, and the state has recently strengthened post-mortem and digital-replica protections. New York’s Election Law also addresses materially deceptive media in political communications. In addition, New York has enacted a synthetic-performer advertising disclosure law for commercial advertisements, although that statute is not effective until 9 June 2026. Companies using synthetic voices, likenesses, or avatars should secure consent, label appropriately, and assess whether the use could be characterised as deceptive, unauthorised, or exploitative.

Transparency is becoming a recurring legal expectation in AI governance, although the precise obligation varies by context. Some laws require disclosure directly, as in parts of the employment or synthetic-media context. In other settings, disclosure matters because failure to disclose can render a statement misleading, defeat consent, or aggravate a regulator’s view of the deployment.

The practical rule is that businesses should tell users, consumers, employees, and counterparties when AI use is material to the interaction, especially where the output could reasonably be mistaken for human work, where automated processing affects rights or opportunities, or where a synthetic depiction or voice is being used.

AI procurement should not be treated like ordinary software purchasing. Buyers should address data rights, model provenance, security controls, testing, benchmark methodology, human-review requirements, change management, audit rights, service levels, logging, open-source components, geographic restrictions, regulatory co-operation, and exit rights. IP indemnities, confidentiality obligations, and restrictions on training or retention of customer data remain especially important.

Government and regulated-industry procurement is pushing the market toward more detailed documentation, even where the law does not prescribe a single contract form. As a result, procurement diligence is becoming one of the most important controls in enterprise AI governance.

AI supply chains are deeper than they appear. A single deployment may depend on a base model, fine-tuning layers, retrieval tools, third-party datasets, moderation services, plug-ins, open-source libraries, cloud infrastructure, and downstream integrators. Legal accountability can arise at any of those layers.

Due diligence should therefore map the full stack, not just the top-line vendor. Key questions include who supplied the model and the data, what open-source terms apply, how updates are made, where logs are stored, whether the provider uses customer data for training, whether any restricted parties are involved, and what evidence exists of testing and incident response. Even where US law is not prescriptive, this kind of supply-chain mapping is increasingly necessary to support defensible deployment.

AI use in hiring, promotion, and termination is one of the most legally sensitive deployment areas in the United States. Title VII, the Americans with Disabilities Act (ADA), the Age Discrimination in Employment Act (ADEA), the FCRA, state discrimination law, accommodation duties, and related rules can all be implicated, depending on the tool and the employer’s workflow. The fact that a vendor supplied the system does not displace the employer’s responsibility.

New York is particularly important because New York City’s Local Law 144 imposes concrete process requirements where covered tools are used to substantially assist or replace discretionary decision-making in hiring or promotion. More broadly, employers should avoid sole reliance on AI, test for adverse impact, provide reasonable accommodations where needed, and preserve records sufficient to explain how a decision was made.

AI is also being used to monitor employee productivity, communications, behaviour, attendance, location, scheduling, and performance. These uses can improve operational visibility, but they raise privacy, discrimination, labour, and retaliation concerns, especially when employees do not understand what is being tracked or how the information will be used.

Technologies and Scope

Relevant tools include productivity dashboards, keystroke and activity monitoring, communications analysis, sentiment or behavioural scoring, scheduling optimisation, workforce fraud analytics, and systems that generate performance narratives or disciplinary recommendations. The broader and more opaque the monitoring, the greater the legal and employee-relations risk.

New York Overlays

New York adds meaningful overlays. State law requires notice to employees for certain electronic monitoring practices. If a tool also affects promotion decisions in New York City, Local Law 144 may come into play. Sector-specific obligations and general discrimination and retaliation law must also be considered in parallel.

Discrimination and Liability Risks

The main risks are not limited to privacy. Monitoring and evaluation tools can produce disparate impact, encode disability-related bias, chill protected activity, create wage-and-hour disputes, or generate adverse employment records that are difficult to challenge. Employers should therefore adopt purpose limitations, human review, notice, retention limits, accommodation processes, and validation protocols before deploying AI-based monitoring at scale.

Digital platforms use AI for recommendation, moderation, ranking, search, advertising, fraud prevention, synthetic content generation, and customer support. That creates a legal mix of consumer protection, privacy, IP, publicity, child-safety, competition, and content-governance risk. Platforms should pay particular attention to how recommender systems and synthetic-media tools are disclosed and controlled.

Financial services firms are already heavily regulated, so AI is generally judged under existing supervisory, disclosure, recordkeeping, fair-dealing, and model-risk expectations. AI use in underwriting, surveillance, sales communications, portfolio tools, customer support, and trading therefore requires careful supervision. In New York, NYDFS and insurance guidance are especially important, and firms should assume there is no regulatory safe harbour merely because a recommendation or communication was generated by AI.

Healthcare AI sits at the intersection of device regulation, privacy, reimbursement, fraud-and-abuse rules, malpractice, civil rights, and professional standards. The legal analysis depends heavily on whether the tool is used for clinical decision support, diagnosis, administrative efficiency, patient communications, or research. New York adds its own privacy, insurance, and professional-liability considerations. High-stakes healthcare uses should be validated and governed with exceptional care.

Autonomous-vehicle law in the United States remains a combination of motor-vehicle safety regulation, tort law, state operating rules, and reporting obligations. NHTSA remains central at the federal level, including through reporting and transparency programmes. New York has historically taken a cautious approach to deployment. For companies in this space, the key legal themes are safety, incident reporting, product liability, and insurance.

Retail and consumer businesses use AI for personalisation, pricing, demand forecasting, fraud control, visual search, customer service, and marketing. The main legal issues are deceptive practices, privacy and eavesdropping risk, accessibility, unfair discrimination, and synthetic interactions that may mislead consumers. Businesses deploying chatbots or synthetic voices should consider whether additional disclosure is appropriate, especially where a user may reasonably believe they are dealing with a human representative.

Industrial AI and robotics are used for predictive maintenance, machine vision, warehouse automation, route optimisation, quality control, and industrial cybersecurity. In this sector, the law is driven less by AI-specific statutes than by workplace-safety rules, product liability, cybersecurity, export controls, and contract. New York has no standalone industrial AI code, but industrial deployments remain legally consequential because they can affect worker safety, physical assets, and critical infrastructure.

AI assets can be protected through a combination of copyright, patent, trade-secret, trade mark, contract, and unfair-competition law. Which tool matters most depends on the asset. Source code, model architectures, weights, training methods, data curation, evaluation techniques, prompts, and documentation may all have value, but not all of them fit neatly into one IP category. In practice, trade-secret and contract protection are often as important as formal registration.

Companies should therefore focus on chain of title, employee and contractor assignment provisions, access controls, open-source compliance, and documentation of who created what. Many disputes about AI assets are really disputes about ownership, confidentiality, and use restrictions rather than about patentability or copyrightability in the abstract.

US law still requires human inventorship and human authorship. AI cannot itself be named as an inventor on a patent application or as the author of a copyrighted work. That does not mean AI-assisted work is unprotectable. It means protection turns on the human contribution.

For patents, the question is whether a human conceived the claimed invention. For copyright, the question is whether the work contains sufficient human-authored expression. Businesses using AI in creative or technical workflows should document human direction, selection, editing, and inventive contribution rather than assuming that the existence of AI eliminates protectability.

Copyright issues around AI training data remain one of the most active and unsettled areas of AI law. Courts and the Copyright Office have made clear that there is no categorical answer. Fair use is fact-specific, and provenance matters. Training on lawfully acquired materials for certain analytical uses may be argued differently from training on pirated or obviously unauthorised corpora, and output-based infringement questions can be distinct from training-based ones.

For deployers, the practical lesson is diligence. Businesses should ask what data a vendor used, what licenses or restrictions apply, whether opt-out commitments are honoured, how outputs are tested for memorisation or similarity, and what contractual protection the vendor is willing to offer.

Purely AI-generated output is not protected by copyright simply because a model produced it. Protection, where available, attaches to human-authored elements such as selection, arrangement, editing, or other creative contributions. That is now a settled baseline in US copyright administration.

As a result, companies creating commercial content with AI should think carefully about both ownership and reuse. Even if copyright is limited, contract, trade mark, trade-secret, and publicity issues may still be significant.

Foundation models and open-source AI create a distinct IP profile. Open models can lower cost and concentration, but they also raise questions about license scope, attribution, redistribution, derivative works, training provenance, security updates, and support. Open weights are not legally simple just because access is broad.

Companies adopting these tools should review upstream licenses, downstream restrictions, acceptable-use terms, and whether the provider offers meaningful commitments on provenance, security, or indemnity. Open-source governance now belongs inside AI governance.

AI training and fine-tuning can create acute data-protection issues when personal data, sensitive information, or confidential enterprise data is included in model inputs or corpora. The United States still lacks a single comprehensive federal privacy law, so the governing rules come from sector statutes, state privacy laws, unfairness doctrines, contracts, and security obligations. That makes data mapping especially important.

Businesses should not assume that data collected for one purpose can automatically be repurposed for training or model improvement. Key questions include notice, consent where required, purpose limitation, minimisation, retention, deletion, and whether sensitive or regulated data has been adequately segregated. Vendor terms on training, retention, and secondary use should be reviewed closely.

Deployment creates its own privacy issues, separate from training. A growing body of state privacy law gives individuals rights of access, deletion, correction, and opt-out, and in some jurisdictions also imposes duties relating to profiling or automated decision-making. Sector laws can add notice, adverse-action, or contestability requirements in specific contexts such as credit, employment, or insurance.

Organisations using AI in consumer- or employee-facing workflows should therefore build processes to explain decisions where required, honour applicable privacy requests, and escalate complaints about inaccurate or unfair outputs.

AI data governance requires clarity about roles across the stack: developer, deployer, customer, cloud provider, data source, and downstream recipient. Strong governance addresses provenance, retention, data-lineage tracking, localisation where relevant, security classification, and contract restrictions.

The United States generally does not impose broad outbound-transfer restrictions for ordinary commercial data, but cross-border AI activity can still be constrained by contracts, sector rules, export controls, sanctions, foreign privacy law, and government-access concerns. Global AI deployments therefore need both privacy review and geopolitical review.

Antitrust law applies to AI markets through familiar theories rather than through an AI-specific test. Current concerns include concentration in chips, cloud infrastructure, model access, and data; exclusive or preferential arrangements between model developers and infrastructure providers; self-preferencing by platforms; access restrictions on APIs or interoperability; and the possibility that algorithmic systems could facilitate collusion or unfair exclusion.

The main practical point is that partnerships, investments, data-sharing arrangements, and distribution agreements in AI should be reviewed with the same care given to other strategically sensitive technology markets. The fact that a market is innovative does not reduce antitrust scrutiny.

Cybersecurity law applies to AI systems in two directions. First, AI systems themselves must be secured under existing cybersecurity duties. Second, AI changes the threat landscape by enabling faster fraud, better phishing, code generation, model manipulation, data extraction, and other attacks. Existing laws and rules therefore remain highly relevant, including the SHIELD Act, NYDFS cybersecurity obligations, sector rules, and incident-response requirements.

From a controls perspective, AI systems need the same security basics as other critical systems, plus AI-specific measures such as prompt-injection defences, access segmentation, secure model and key management, dependency monitoring, output filtering where appropriate, and logging capable of supporting forensic review. Agentic systems deserve heightened scrutiny because they can convert a prompt compromise into downstream system action.

AI has meaningful ESG implications even though the United States does not have a general AI-ESG disclosure regime. Environmental concerns include compute intensity, data-centre power and water consumption, hardware supply chains, and e-waste. Social concerns include bias, labour displacement, surveillance, misinformation, and unequal access. Governance concerns include board oversight, accountability, incident escalation, and the risk of overstating responsible-AI commitments.

For most companies, the right approach is to integrate AI into existing ESG, risk, and disclosure processes rather than to treat AI as a separate marketing narrative. AI-related claims can become material under securities, consumer-protection, or employment law if they are overstated or unsupported.

An effective AI governance programme should begin with an inventory of use cases and a risk-based classification process. From there, organisations should establish approval pathways, written policies, role-based accountability, training, vendor diligence, testing and red-teaming where appropriate, deployment gates, monitoring, incident response, change management, and board or senior-management reporting. The best programmes are living systems, not one-time policy documents.

A practical AI governance programme is usually simpler than an attempt to implement every available framework line by line. Most enterprises are better served by a tailored, risk-based programme with sector-specific overlays where needed. For New York-facing programmes, that means layering in NYDFS expectations, Local Law 144, privacy and cybersecurity rules, employment law, the RAISE Act where applicable, and the state’s other newer AI-specific statutes. The overall objective is not to eliminate all AI risk. It is to know where the systems are, understand what they do, decide where they may be used, and preserve enough governance and evidence to defend those decisions.