The Office of Emergency Services must perform a risk analysis of potential GenAI threats to California's critical infrastructure, including mass-casualty scenarios, and provide the analysis and any recommendations to the Governor.

Operators must, on or before July 1, 2027, and annually thereafter, perform and document a comprehensive risk assessment to identify any child safety risk posed by the design, configuration, and operation of the companion chatbot. The assessment must evaluate: (1) the likelihood of a covered harm occurring to users; (2) differential risks across age groups and developmental stages; (3) known vulnerabilities of children; (4) empirical data from actual use; and (5) relevant academic research and regulatory guidance.

Healthcare insurers must establish an ongoing biannual quality assurance testing process for AI algorithms meeting Commissioner-specified safety and efficacy parameters aligned with national benchmarks, including validation for generalizability, risk-level-based evaluation, real-world evidence generation, benchmarking against Commissioner-approved standardized metrics in representative Arkansas populations, and multi-institutional, demographically representative testing datasets with documented provenance that are regularly updated.

Chatbot providers must, on a monthly basis, evaluate their chatbot for potential risk of harm to users and make information about the chatbot publicly available on their website. Chatbot providers must mitigate any identified risk of harm to users. These evaluations and mitigations must comply with rules adopted by the attorney general.

AI businesses must conduct a comprehensive pre-market risk assessment for each high-risk AI system, evaluating (1) the potential for discrimination or bias, (2) safety risks to children and vulnerable populations, (3) privacy and data protection risks, and (4) catastrophic risks.

Before deploying any new design or materially modifying an existing design, covered entities must assess the risk that the design will cause compulsive use in minors. Where a reasonable lower-risk alternative exists, the alternative must be set as the default until the consumer explicitly requests the original AND has been age-assured as not a minor. Designs whose minor-compulsive-use risk outweighs the benefit may not be deployed to any consumer absent that same gate. Existing designs must be retroactively assessed and mitigated.

Developers of AI products deemed by the office to have a higher potential risk must provide a correspondingly higher burden of affirmative proof demonstrating the safety of their product before deployment.

AAASPs must demonstrate at licensure that they meet Board-published safety and performance benchmarks, submit annual performance reports proving continued benchmark compliance, and report adverse events and reportable events as defined. The Board may suspend a license on evidence of model drift or degraded safety outcomes.

AAASPs must operate clinical AI services only within the model's validated technical specifications, training data, intended use case, and performance parameters as submitted to the Board, and must meet the human-equivalent standard of care for the clinical task.

Large chatbot providers must write, implement, comply with, and publicly publish a safety plan detailing how they assess child safety risks, apply mitigations to address those risks, and use third parties to evaluate the potential for and effectiveness of mitigations of child safety risks.

Academic medical centers seeking the research exemption must ensure all AI systems used in the research are HIPAA-compliant, undergo documented safety testing before use with participants, include real-time monitoring to detect crisis situations, harmful content, or therapeutic errors, and operate only as adjuncts to (not replacements for) licensed mental health professionals. The program must also maintain a data safety monitoring plan and procedures for reporting adverse events to the IRB.

Large chatbot providers must write, implement, comply with, and publicly publish a public safety and child protection plan that describes how they assess child safety risks, apply mitigations based on those assessments, and use third parties to evaluate child safety risks and mitigation effectiveness.

Developers must exercise reasonable care in the design of AI products, including accounting for intended and reasonably foreseeable unintended uses and adopting technologically feasible alternative designs that would reduce foreseeable risk of harm without significantly impairing product usefulness.

Developers earn a rebuttable presumption of non-defectiveness by (1) conducting documented NIST AI RMF-aligned testing, evaluation, verification, validation, and auditing; (2) mitigating foreseeable risks and considering alternatives; (3) disclosing foreseeable risks and mitigation tactics to deployers and consumers; (4) maintaining an AI data sheet — covering intended uses, training data, risk management, and red-teaming results — available to the Attorney General on request; (5) for products accessible to users under 17, documenting cognitive/emotional development impact assessments, implementing age-gating or content restrictions, and disclosing potential risks to deployers, consumers, and guardians; and (6) prominently including data-sheet information in product terms and conditions.

Developers may earn a rebuttable presumption of non-defectiveness by: (1) conducting documented NIST AI RMF-aligned testing, evaluation, verification, validation, and auditing; (2) mitigating foreseeable risks and considering alternatives; (3) disclosing foreseeable risks and mitigation tactics to deployers and consumers; (4) maintaining an AI data sheet — covering intended uses, training data, foreseeable risks, and red-teaming results per NIST guidance — available to the Attorney General on request; (5) for products accessible to minors under 17, documenting cognitive/emotional impact assessments, implementing age-gating, and disclosing risks to consumers and guardians; and (6) prominently including AI data sheet information in product terms and conditions.

Manufacturers must conduct a child AI safety assessment for each artificial intelligence toy before marketing or selling it in Maryland, describing foreseeable risks and mitigation strategies. The assessment must be repeated annually and following any significant change to the toy's AI features. Toys already on the market before July 1, 2026, must have an initial assessment completed by January 1, 2027.

Developers must exercise reasonable care in the design of their AI product, including knowing or discovering design defects, accounting for intended and reasonably foreseeable unintended uses, and ensuring no technologically feasible and practical alternative design would have mitigated the foreseeable risk of harm without significantly impairing the product's intended use.

Developers may establish a rebuttable presumption of non-defectiveness by (1) conducting documented testing, evaluation, verification, validation, and auditing consistent with industry best practices, (2) mitigating foreseeable risks to the extent possible, (3) disclosing foreseeable risks and mitigation tactics directly to deployers and consumers, and (4) for products designed for or reasonably likely to be used by minors, considering and mitigating risks to minors including cognitive and emotional development impacts, implementing age and content restrictions, and providing clear disclosures to deployers, consumers, and guardians.

Developers must implement appropriate safeguards to prevent unreasonable risk of critical harm before deploying an AI model.

Developers must, before deploying any AI model, implement a written safety and security protocol that: (1) describes reasonable protections and procedures to reduce the risk of critical harm; (2) describes reasonable administrative, technical, and physical cybersecurity protections to reduce the risk of unauthorized access or misuse leading to critical harm; (3) describes in detail the testing procedure to evaluate whether the model poses an unreasonable risk of critical harm, could evade control, or could be misused or modified to increase the risk of critical harm; (4) enables the developer or third party to comply with the RAISE Act; and (5) designates senior personnel responsible for ensuring compliance. Developers must also implement appropriate safeguards to prevent unreasonable risk of critical harm.

Artificial intelligence companies must annually conduct safety tests on all AI technology they sell, develop, deploy, use, or offer for sale in New Jersey, adhering to the minimum requirements established by the Office of Information Technology. The safety tests must address potential biases, inaccuracies, and cybersecurity threats.

Artificial intelligence companies must annually conduct safety tests — covering cybersecurity threats, bias, inaccuracies, and legal compliance — on all AI technology they sell, develop, deploy, use, or offer for sale in New Jersey, adhering to OIT-established minimum requirements.

Employers, public entities, and vendors must not deploy an AEDS or EMT unless it is pretested and validated to serve one of six enumerated allowable purposes and is limited to the least-invasive means, smallest population, and least data and collection frequency necessary for those purposes.

Persons developing automated systems must conduct pre-deployment testing, risk identification, and mitigation before deployment, and must subject the system to ongoing monitoring demonstrating safety and effectiveness based on intended use, mitigation of unsafe outcomes beyond intended use, and adherence to domain-specific standards.

Persons developing automated systems must not deploy any system that fails to meet the safety and effectiveness requirements of § 504. If an already-deployed system fails to meet those requirements, it must be removed. No automated system may be designed with the intent or a reasonably foreseeable possibility of endangering the safety of any New York resident or community.

Persons developing automated systems must design those systems to proactively protect New York residents from harm stemming from unintended, yet foreseeable, uses or impacts.

Developers of AI technology intended for use in a professional domain regulated under Title VIII of the Education Law must demonstrate that at least one professional domain expert was directly and substantially involved in: (1) the technology design phase, (2) the data selection and training process, (3) validation and testing of system outputs, and (4) ongoing risk assessment and post-deployment evaluation. Covered application areas include healthcare diagnostics and treatment recommendations, legal decision-making or document generation, financial advising or lending tools, educational curriculum or assessment tools, construction and structural safety systems, and public safety or surveillance technologies.

Deployers must have a diligent review and selection process for each AI device before deployment.

Suppliers must disclose in the written policy the procedures by which they: (i) conduct testing before making the chatbot publicly available and regularly thereafter to ensure output poses no greater risk than communicating with a human; (ii) identify reasonably foreseeable adverse outcomes and potentially harmful interactions; (iii) provide a mechanism for consumers to report potentially harmful interactions; (iv) implement protocols to assess and respond to risk of harm to consumers or others; (v) detail actions taken to prevent or mitigate adverse outcomes or harmful interactions; (vi) implement protocols to respond as soon as practicable to acute risks of physical harm; (vii) ensure regular, objective reviews of safety, accuracy, and efficacy, which may include internal or external audits; (viii) provide consumers with instructions on safe use of the chatbot; (ix) prioritize consumer mental health and safety over engagement metrics or profit; (x) implement measures to prevent discriminatory treatment of consumers; and (xi) ensure compliance with HIPAA security and privacy provisions (45 CFR Parts 160 and 164) as if the supplier were a covered entity.

Facilities must ensure that AI-based algorithms used for clinical decision making do not create foreseeable, material risks of harm to the patient.

Insurers must ensure that AI-based algorithms used in utilization review do not create foreseeable, material risks of harm to the covered person.

MA or CHIP managed care plans must ensure that AI-based algorithms used in utilization review do not create foreseeable, material risks of harm to the enrollee.

An AI mental health application may be considered to have successfully completed testing only after testing results demonstrate competency and safety in providing AI mental health services.

Covered entities must assess the need for guardrails or limitations on covered algorithm uses, including whether certain applications should be prohibited or restricted through terms of use, licensing agreements, or other legal agreements.

Covered entities must disclose benchmark evaluation results — whether self-driven or through audit — for each foundation model, including the precautions the model takes in high-risk domains such as healthcare, CBRN weapons, national security, cybersecurity, critical infrastructure, elections, law enforcement, financial and housing decisions, education, employment, public services, and information relating to vulnerable populations including minors and seniors.

Covered entities must assess the need for guardrails or limitations on certain uses or applications of their automated decision systems or augmented critical decision processes, including whether particular uses should be prohibited or restricted through terms of use, licensing agreements, or other legal agreements.

Developers must exercise reasonable care in the design of covered AI products; failure to do so creates liability for proximately caused harm. A product whose design is manifestly unreasonable triggers liability without the claimant needing to prove a reasonable alternative design. Noncompliance with an applicable product safety statute or regulation creates a presumption of defectiveness.

Covered advanced AI system developers must participate in the DOE Advanced Artificial Intelligence Evaluation Program, which conducts standardized safety testing, adversarial red-team evaluation, and classified assessments of advanced AI systems.

Developers and deployers must (1) take reasonable measures to prevent and mitigate harms identified in evaluations and assessments, (2) ensure independent auditors have all necessary information, (3) consult impacted stakeholders before deployment, (4) certify that the algorithm is not likely to cause harm, disparate impact, or deceptive practices, (5) ensure the algorithm performs consistent with its advertised purpose, (6) ensure data used is relevant and appropriate, and (7) ensure intended use will not violate this Act.

Developers and deployers of any inherently dangerous AI system that could reasonably be expected to impact consumers must exercise reasonable care to avoid any reasonably foreseeable risk arising out of the development, intentional and substantial modification, or deployment of the system that causes or is likely to cause: (1) crime or unlawful acts; (2) unfair or deceptive treatment; (3) physical, financial, relational, or reputational injury; (4) highly offensive psychological injuries; (5) offensive intrusion upon privacy or seclusion; (6) IP rights violations; (7) discrimination on the basis of race, color, ethnicity, sex, sexual orientation, gender identity, sex characteristics, religion, national origin, familial status, biometric information, or disability status; (8) behavioral distortion causing physical or psychological harm; or (9) exploitation of age- or disability-based vulnerabilities to distort behavior causing harm.

Developers must not place any inherently dangerous AI system in the stream of commerce unless the developer has conducted documented testing, evaluation, verification, and validation at least as stringent as the latest version of the NIST AI Risk Management Framework. For any AI system that creates reasonably foreseeable risks under the standard-of-care provision, developers must mitigate those risks to the extent possible, consider alternatives, and disclose vulnerabilities and mitigation tactics to deployers.

Developers must exercise reasonable care in the design of AI products, accounting for both intended uses and reasonably foreseeable unintended uses, ensuring no defective design that could cause harm when a technologically feasible and practical alternative design exists.

Developers seeking the safe-harbor presumption of non-defectiveness must (1) conduct documented testing, evaluation, verification, validation, and auditing consistent with industry best practices, (2) mitigate foreseeable risks, (3) disclose foreseeable risks and mitigation tactics to deployers and consumers, (4) maintain an AI data sheet covering intended uses, training datasets, foreseeable risks, and red-teaming results available to the Attorney General on request, (5) for products designed for or likely used by minors, document impact assessments on cognitive and emotional development, implement age-gating or content restrictions, and provide clear risk disclosures to deployers, consumers, and guardians, and (6) prominently include the AI data sheet information in the product's terms and conditions.

Deployers who materially and substantially change an AI product, or who intentionally misuse it contrary to the developer's express warranty, are liable as developers — subject to the same design-defect, failure-to-warn, and express-warranty standards applicable to developers.

Developers earn a rebuttable presumption of non-defectiveness by (1) conducting documented NIST AI RMF-aligned testing, evaluation, verification, validation, and auditing; (2) mitigating foreseeable risks and considering alternatives; (3) disclosing foreseeable risks and mitigation tactics to deployers and consumers; (4) maintaining an AI data sheet — covering intended uses, training data, risk management, and red-teaming results — available to the Attorney General on request; (5) for products accessible to users under 17, documenting cognitive/emotional development impact assessments, implementing age-gating or content restrictions, and disclosing potential risks to deployers, consumers, and guardians; and (6) prominently including data-sheet information in product terms and conditions.

Developers may earn a rebuttable presumption of non-defectiveness by: (1) conducting documented NIST AI RMF-aligned testing, evaluation, verification, validation, and auditing; (2) mitigating foreseeable risks and considering alternatives; (3) disclosing foreseeable risks and mitigation tactics to deployers and consumers; (4) maintaining an AI data sheet — covering intended uses, training data, foreseeable risks, and red-teaming results per NIST guidance — available to the Attorney General on request; (5) for products accessible to minors under 17, documenting cognitive/emotional impact assessments, implementing age-gating, and disclosing risks to consumers and guardians; and (6) prominently including AI data sheet information in product terms and conditions.

Developers may establish a rebuttable presumption of non-defectiveness by (1) conducting documented testing, evaluation, verification, validation, and auditing consistent with industry best practices, (2) mitigating foreseeable risks to the extent possible, (3) disclosing foreseeable risks and mitigation tactics directly to deployers and consumers, and (4) for products designed for or reasonably likely to be used by minors, considering and mitigating risks to minors including cognitive and emotional development impacts, implementing age and content restrictions, and providing clear disclosures to deployers, consumers, and guardians.

Persons developing automated systems must design those systems to proactively protect New York residents from harm stemming from unintended, yet foreseeable, uses or impacts.

Covered advanced AI system developers must participate in the DOE Advanced Artificial Intelligence Evaluation Program, which conducts standardized safety testing, adversarial red-team evaluation, and classified assessments of advanced AI systems.

Developers seeking the safe-harbor presumption of non-defectiveness must (1) conduct documented testing, evaluation, verification, validation, and auditing consistent with industry best practices, (2) mitigate foreseeable risks, (3) disclose foreseeable risks and mitigation tactics to deployers and consumers, (4) maintain an AI data sheet covering intended uses, training datasets, foreseeable risks, and red-teaming results available to the Attorney General on request, (5) for products designed for or likely used by minors, document impact assessments on cognitive and emotional development, implement age-gating or content restrictions, and provide clear risk disclosures to deployers, consumers, and guardians, and (6) prominently include the AI data sheet information in the product's terms and conditions.

Persons developing automated systems must arrange for independent evaluation and reporting confirming the system is safe and effective, including reporting of steps taken to mitigate potential harms. Results must be made public whenever possible.

Chatbot providers must, on a monthly basis, evaluate their chatbot for potential risk of harm to users and make information about the chatbot publicly available on their website. Chatbot providers must mitigate any identified risk of harm to users. These evaluations and mitigations must comply with rules adopted by the attorney general.

Health care providers that use AI to make or substantially factor into consequential decisions must (1) monitor the usage of AI systems in consequential decisions on an ongoing basis; (2) conduct regular performance evaluations of the AI systems, including assessment of potential biases, risks to patient safety, rights, and personal data confidentiality, and mitigation strategies for identified risks; and (3) implement procedures to address any deficiencies identified through monitoring or performance evaluations, including suspension or recalibration of the AI system.

Deployers must implement and maintain protocols designed to detect, respond to, report, and mitigate harm their chatbot may cause users. These protocols must prioritize user safety and well-being over the deployer's commercial interests.

Deployers must implement and maintain protocols to detect, respond to, report, and mitigate harm that the public-facing chatbot may cause users, taking commercially reasonable steps to protect the safety and well-being of users.

Deployers must implement and maintain protocols to detect, respond to, report, and mitigate harm the chatbot may cause a user. These protocols must prioritize the safety and well-being of users over the deployer's interests.

AAASPs must demonstrate at licensure that they meet Board-published safety and performance benchmarks, submit annual performance reports proving continued benchmark compliance, and report adverse events and reportable events as defined. The Board may suspend a license on evidence of model drift or degraded safety outcomes.

Academic medical centers seeking the research exemption must ensure all AI systems used in the research are HIPAA-compliant, undergo documented safety testing before use with participants, include real-time monitoring to detect crisis situations, harmful content, or therapeutic errors, and operate only as adjuncts to (not replacements for) licensed mental health professionals. The program must also maintain a data safety monitoring plan and procedures for reporting adverse events to the IRB.

Covered online platforms must maintain at least one holdout group and subject all changes to the design of an algorithmic recommender system to a long-term holdout assessment (a minimum 12-month controlled assessment).

Chatbot providers must assess their chatbot for risks of harm to users on a monthly basis using metrics to be established by attorney general rulemaking, and must mitigate identified risks per those rules.

Manufacturers must conduct a child AI safety assessment for each artificial intelligence toy before marketing or selling it in Maryland, describing foreseeable risks and mitigation strategies. The assessment must be repeated annually and following any significant change to the toy's AI features. Toys already on the market before July 1, 2026, must have an initial assessment completed by January 1, 2027.

Carriers must conduct a model review of any AI, algorithm, or software tool used in utilization review if, within a six-month period, more than a Commissioner-specified percentage of adverse decisions made using that tool result in grievances. The carrier must submit the findings of the model review in its quarterly report to the Commissioner.

After a security breach, employers must contract with a third party to audit the affected monitoring or automated decisions tool to confirm that vulnerabilities have been fixed.

Covered businesses must maintain at least one holdout group and subject all algorithmic recommender system design changes to a long-term holdout assessment of at least 12 months, under rules to be adopted by the Commissioner of Commerce.

Persons developing automated systems must conduct pre-deployment testing, risk identification, and mitigation before deployment, and must subject the system to ongoing monitoring demonstrating safety and effectiveness based on intended use, mitigation of unsafe outcomes beyond intended use, and adherence to domain-specific standards.

Developers and deployers must: (1) take reasonable measures to prevent and mitigate any harm identified by pre-deployment evaluations or impact assessments; (2) take reasonable measures to ensure independent auditors have all necessary information to complete accurate evaluations and assessments; (3) consult impacted stakeholders and communities before deploying, licensing, or offering a covered algorithm; (4) certify, based on evaluation/assessment results, that (a) use of the algorithm is not likely to result in harm or disparate impact, (b) benefits to affected individuals outweigh harms, and (c) use is not likely to result in a deceptive act or practice; (5) ensure the algorithm functions at a level of reasonable performance consistent with its expected and publicly-advertised performance, purpose, or use; (6) ensure all data used in design, development, deployment, or use is relevant and appropriate to the deployment context and publicly-advertised purpose; and (7) ensure intended use is not likely to result in a violation of this article.

Licensed independent verification organizations must implement their approved plan to verify AI models' and applications' ongoing mitigation of risks for which they are licensed to verify.

Deployers must conduct and document regular performance evaluations and risk assessments of each deployed AI device, informed by invited feedback from qualified end-users and, when applicable, participation in national specialty society-administered AI assessment registries. Whenever performance concerns are identified, deployers must implement appropriate corrective actions to mitigate risk to patients.

Deployers must continuously monitor the performance of all deployed AI devices, including assessing any impact on patient safety or the quality of patient care. In conducting this performance monitoring, deployers must participate in national specialty society-administered AI assessment registries when feasible.

Deployers must implement reasonable, ongoing oversight, safeguards, and fail-safe mechanisms designed to prevent unlawful, negligent, or harmful conduct and must take timely corrective action upon becoming aware of a risk or failure relating to the AI system, in order to preserve the affirmative defense to deployment liability.

Suppliers must disclose in the written policy the procedures by which they: (i) conduct testing before making the chatbot publicly available and regularly thereafter to ensure output poses no greater risk than communicating with a human; (ii) identify reasonably foreseeable adverse outcomes and potentially harmful interactions; (iii) provide a mechanism for consumers to report potentially harmful interactions; (iv) implement protocols to assess and respond to risk of harm to consumers or others; (v) detail actions taken to prevent or mitigate adverse outcomes or harmful interactions; (vi) implement protocols to respond as soon as practicable to acute risks of physical harm; (vii) ensure regular, objective reviews of safety, accuracy, and efficacy, which may include internal or external audits; (viii) provide consumers with instructions on safe use of the chatbot; (ix) prioritize consumer mental health and safety over engagement metrics or profit; (x) implement measures to prevent discriminatory treatment of consumers; and (xi) ensure compliance with HIPAA security and privacy provisions (45 CFR Parts 160 and 164) as if the supplier were a covered entity.

Chatbot providers must, on a monthly basis, (1) evaluate their chatbot for potential risk of harm to users and (2) make information about the chatbot publicly available on the provider's website. Chatbot providers must also mitigate any risk of harm to users identified through this evaluation. Evaluations and mitigations must comply with rules and regulations promulgated by the Attorney General.

Chatbot providers must, on a monthly basis, (1) evaluate their chatbot for potential risk of harm to users and (2) make information about the chatbot publicly available on the provider's website. Chatbot providers must also mitigate any identified risk of harm to users. All evaluations and mitigation must comply with rules and regulations promulgated by the Attorney General.

Covered entities must attempt to eliminate or mitigate in a timely manner any impact of a covered algorithm that demonstrates a likely material negative impact with legal or similarly significant effects on a consumer's life.

Chatbot providers must, on a monthly basis, assess their chatbot for risks of harm to users using metrics prescribed by Attorney General rules, and must actively mitigate any identified risks of harm.

Operators must take and document measures that reasonably mitigate any child safety risk identified in the comprehensive risk assessment required under subdivision (a).

The commission's standards must require electrical and gas corporations to identify and take steps to ensure that AI models do not impact utility safety, affordability, and reliability.

Deployers and vendors must use reasonable care to protect consumers from known or reasonably foreseeable material risks of algorithmic discrimination, material errors (including systematic reliability failures), and cybersecurity and data integrity failures that materially affect output reliability or security.

Deployers must implement and maintain protocols designed to detect, respond to, report, and mitigate harm their chatbot may cause users. These protocols must prioritize user safety and well-being over the deployer's commercial interests.

Deployers must implement and maintain protocols to detect, respond to, report, and mitigate harm that the public-facing chatbot may cause users, taking commercially reasonable steps to protect the safety and well-being of users.

Deployers must implement and maintain protocols to detect, respond to, report, and mitigate harm the chatbot may cause a user. These protocols must prioritize the safety and well-being of users over the deployer's interests.

Large chatbot providers must write, implement, comply with, and publicly publish a safety plan detailing how they assess child safety risks, apply mitigations to address those risks, and use third parties to evaluate the potential for and effectiveness of mitigations of child safety risks.

Large chatbot providers must write, implement, comply with, and publicly publish a public safety and child protection plan that describes how they assess child safety risks, apply mitigations based on those assessments, and use third parties to evaluate child safety risks and mitigation effectiveness.

Developers earn a rebuttable presumption of non-defectiveness by (1) conducting documented NIST AI RMF-aligned testing, evaluation, verification, validation, and auditing; (2) mitigating foreseeable risks and considering alternatives; (3) disclosing foreseeable risks and mitigation tactics to deployers and consumers; (4) maintaining an AI data sheet — covering intended uses, training data, risk management, and red-teaming results — available to the Attorney General on request; (5) for products accessible to users under 17, documenting cognitive/emotional development impact assessments, implementing age-gating or content restrictions, and disclosing potential risks to deployers, consumers, and guardians; and (6) prominently including data-sheet information in product terms and conditions.

Developers may earn a rebuttable presumption of non-defectiveness by: (1) conducting documented NIST AI RMF-aligned testing, evaluation, verification, validation, and auditing; (2) mitigating foreseeable risks and considering alternatives; (3) disclosing foreseeable risks and mitigation tactics to deployers and consumers; (4) maintaining an AI data sheet — covering intended uses, training data, foreseeable risks, and red-teaming results per NIST guidance — available to the Attorney General on request; (5) for products accessible to minors under 17, documenting cognitive/emotional impact assessments, implementing age-gating, and disclosing risks to consumers and guardians; and (6) prominently including AI data sheet information in product terms and conditions.

Consumer reporting agencies must maintain a contingency plan for system failures or data breaches that could compromise algorithmic integrity.

Developers must, before deploying any AI model, implement a written safety and security protocol that: (1) describes reasonable protections and procedures to reduce the risk of critical harm; (2) describes reasonable administrative, technical, and physical cybersecurity protections to reduce the risk of unauthorized access or misuse leading to critical harm; (3) describes in detail the testing procedure to evaluate whether the model poses an unreasonable risk of critical harm, could evade control, or could be misused or modified to increase the risk of critical harm; (4) enables the developer or third party to comply with the RAISE Act; and (5) designates senior personnel responsible for ensuring compliance. Developers must also implement appropriate safeguards to prevent unreasonable risk of critical harm.

Persons developing automated systems must develop those systems in collaboration with diverse communities, stakeholders, and domain experts to identify and address potential concerns, risks, or impacts.

Every person, firm, partnership, association, or corporation doing business or offering products to consumers in New York must develop a responsible capability scaling policy for the use and development of artificial intelligence by that entity. The policy must be a set of best practices that identify, monitor, and rectify or mitigate risk of harm.

Developers and deployers must: (1) take reasonable measures to prevent and mitigate any harm identified by pre-deployment evaluations or impact assessments; (2) take reasonable measures to ensure independent auditors have all necessary information to complete accurate evaluations and assessments; (3) consult impacted stakeholders and communities before deploying, licensing, or offering a covered algorithm; (4) certify, based on evaluation/assessment results, that (a) use of the algorithm is not likely to result in harm or disparate impact, (b) benefits to affected individuals outweigh harms, and (c) use is not likely to result in a deceptive act or practice; (5) ensure the algorithm functions at a level of reasonable performance consistent with its expected and publicly-advertised performance, purpose, or use; (6) ensure all data used in design, development, deployment, or use is relevant and appropriate to the deployment context and publicly-advertised purpose; and (7) ensure intended use is not likely to result in a violation of this article.

Deployers must conduct and document regular performance evaluations and risk assessments of each deployed AI device, informed by invited feedback from qualified end-users and, when applicable, participation in national specialty society-administered AI assessment registries. Whenever performance concerns are identified, deployers must implement appropriate corrective actions to mitigate risk to patients.

Deployers must implement reasonable, ongoing oversight, safeguards, and fail-safe mechanisms designed to prevent unlawful, negligent, or harmful conduct and must take timely corrective action upon becoming aware of a risk or failure relating to the AI system, in order to preserve the affirmative defense to deployment liability.

Suppliers must disclose in the written policy the procedures by which they: (i) conduct testing before making the chatbot publicly available and regularly thereafter to ensure output poses no greater risk than communicating with a human; (ii) identify reasonably foreseeable adverse outcomes and potentially harmful interactions; (iii) provide a mechanism for consumers to report potentially harmful interactions; (iv) implement protocols to assess and respond to risk of harm to consumers or others; (v) detail actions taken to prevent or mitigate adverse outcomes or harmful interactions; (vi) implement protocols to respond as soon as practicable to acute risks of physical harm; (vii) ensure regular, objective reviews of safety, accuracy, and efficacy, which may include internal or external audits; (viii) provide consumers with instructions on safe use of the chatbot; (ix) prioritize consumer mental health and safety over engagement metrics or profit; (x) implement measures to prevent discriminatory treatment of consumers; and (xi) ensure compliance with HIPAA security and privacy provisions (45 CFR Parts 160 and 164) as if the supplier were a covered entity.

Developers and deployers must (1) take reasonable measures to prevent and mitigate identified harms, (2) ensure auditors have all necessary information, (3) consult affected communities before deployment, (4) certify that the algorithm is not likely to cause harm or disparate impact, that benefits outweigh harms, and that use is not deceptive, (5) ensure the algorithm performs at a reasonable professional level consistent with its advertised purpose, (6) ensure data used is relevant and appropriate, and (7) ensure intended use is not likely to violate the Act.

Providers of covered platforms must establish, implement, maintain, and enforce reasonable policies, practices, and procedures to address harms to minors including severe physical-violence threats, sexual exploitation, drug/tobacco/cannabis/gambling/alcohol distribution, and deceptive financial harm, proportionate to platform size and technical feasibility.

Developers and deployers must (1) take reasonable measures to prevent and mitigate harms identified in evaluations and assessments, (2) ensure independent auditors have all necessary information, (3) consult impacted stakeholders before deployment, (4) certify that the algorithm is not likely to cause harm, disparate impact, or deceptive practices, (5) ensure the algorithm performs consistent with its advertised purpose, (6) ensure data used is relevant and appropriate, and (7) ensure intended use will not violate this Act.

Developers and deployers of any inherently dangerous AI system that could reasonably be expected to impact consumers must exercise reasonable care to avoid any reasonably foreseeable risk arising out of the development, intentional and substantial modification, or deployment of the system that causes or is likely to cause: (1) crime or unlawful acts; (2) unfair or deceptive treatment; (3) physical, financial, relational, or reputational injury; (4) highly offensive psychological injuries; (5) offensive intrusion upon privacy or seclusion; (6) IP rights violations; (7) discrimination on the basis of race, color, ethnicity, sex, sexual orientation, gender identity, sex characteristics, religion, national origin, familial status, biometric information, or disability status; (8) behavioral distortion causing physical or psychological harm; or (9) exploitation of age- or disability-based vulnerabilities to distort behavior causing harm.

Developers seeking the safe-harbor presumption of non-defectiveness must (1) conduct documented testing, evaluation, verification, validation, and auditing consistent with industry best practices, (2) mitigate foreseeable risks, (3) disclose foreseeable risks and mitigation tactics to deployers and consumers, (4) maintain an AI data sheet covering intended uses, training datasets, foreseeable risks, and red-teaming results available to the Attorney General on request, (5) for products designed for or likely used by minors, document impact assessments on cognitive and emotional development, implement age-gating or content restrictions, and provide clear risk disclosures to deployers, consumers, and guardians, and (6) prominently include the AI data sheet information in the product's terms and conditions.

Healthcare insurers must establish an ongoing biannual quality assurance testing process for AI algorithms meeting Commissioner-specified safety and efficacy parameters aligned with national benchmarks, including validation for generalizability, risk-level-based evaluation, real-world evidence generation, benchmarking against Commissioner-approved standardized metrics in representative Arkansas populations, and multi-institutional, demographically representative testing datasets with documented provenance that are regularly updated.

Chatbot providers must assess their chatbot for risks of harm to users on a monthly basis using metrics to be established by attorney general rulemaking, and must mitigate identified risks per those rules.

Consumer reporting agencies must implement a system to continuously improve algorithmic performance based on review outcomes and human expert input.

Developers and deployers must (1) take reasonable measures to prevent and mitigate identified harms, (2) ensure auditors have all necessary information, (3) consult affected communities before deployment, (4) certify that the algorithm is not likely to cause harm or disparate impact, that benefits outweigh harms, and that use is not deceptive, (5) ensure the algorithm performs at a reasonable professional level consistent with its advertised purpose, (6) ensure data used is relevant and appropriate, and (7) ensure intended use is not likely to violate the Act.