§ 1-509 Use of Generative Artificial Intelligence Technology and

Large Language Models

A. Definitions. In this section the following definitions apply: “Administrative Director” means the administrative director of the Administrative Office of the Courts (AOC). “AI-generated material” means materials generated, in whole or in part, by artificial intelligence (AI). “Artificial Intelligence Tool” or “AI Tool” means an AI product, solution, or application. “Criminal Justice Information Services (CJIS) data” means private or sensitive data gathered from local, state, or federal law enforcement agencies, including biometric data, such as fingerprints, and identity, person, organization, property, and case/incident history. It additionally includes criminal background information, copies of private documents, and anything else that could be classified as sensitive. It also includes CJIS-provided data necessary for civil agencies to perform their mission, including data used to make hiring decisions. “Clerk of superior court” means a person elected to the office provided by Article 6, Section 23 of the Arizona Constitution. “Content” means data, documents, information, or records that may be in the form of text, images, video, sounds, or any other medium. “Court personnel” means all judicial branch judicial officers, clerks of superior court, employees, law clerks, interns, externs, and volunteers. “Court proprietary content” means internal court content not meant for public release including but not limited to draft opinions, draft orders, and internal court manuals, as well as notes, drafts, work product, and memoranda prepared by judges, attorneys, and law clerks employed by the court or court personnel at a judge’s direction as provided by Supreme Court Rule 123(d)(4) and (e)(9). “Generative AI” means AI technology that can create new content through machine learning based on data input. Large Language Models (LLM’s) are a type of Generative AI that have been trained on vast amounts of data, which they use to understand prompts and produce content such as, but not limited to, text, images, videos, or sounds. “Judicial leadership” means, as applicable, the chief justice, the chief judge of each court of appeals division, and for the superior court, justice courts, and municipal courts in a county, the presiding judge of the superior court in that county. Judicial leadership may designate someone to oversee the use of AI tools. Such designee must be another judicial officer or a court administrator. “Non-Production Use” means use of an AI tool for testing and evaluation purposes only. “Non-public” means closed, confidential, sensitive content, or otherwise not subject to public disclosure as provided by law, court rule, this section, or court order. “Non-Sequestered AI System” means an AI system in which the vendor does not protect the confidentiality of user input or prompt data. “Sensitive content” means social security numbers, driver license numbers, bank account numbers, credit card numbers, any other financial account or personally identifying numbers, and any other content deemed sensitive by court rule or statute. “Sequestered AI System” means an AI system in which the vendor protects the confidentiality of user input or prompt data.

B. Applicability. This section applies to all court personnel.

C. Purpose. To promote the use of Generative AI tools when it is beneficial and appropriate, this section provides the administrative requirements, standards, and guidelines to ensure its appropriate use and safeguard controls.

D. Authorization. Court personnel are authorized to use approved Generative AI tools for work-related purposes as set forth in this section.

E. Considerations. When considering whether to permit the use of a Generative AI tool, judicial leadership must consider:

1. whether the tool is suitable for its intended purpose;

2. associated costs;

3. where content input by the user is processed or stored;

4. whether content input by the user is uploaded into the public domain;

5. how a vendor is permitted to use the content;

6. the security of the content transmission; and

7. whether use of the tool conforms to the requirements of this section and other judicial branch policies. Courts should review any relevant or governing document pertaining to the Generative AI tool, including the end user licensing agreement (EULA), to make these determinations.

F. Use of AI Tools.

1. Use of Work Email Accounts. Court personnel are permitted to use a judicial branch work email account to use or register for Generative AI tools or create AI-generated material for work-related purposes.

2. Use on Court-Owned and Personal Devices. Only approved Generative AI tools are permitted to be installed or used on court-owned devices, and on personal devices that are used to access court non-public content.

3. Compliance with Laws and Judicial Branch Policies. Any AI use must comply with all applicable laws and judicial branch policies.

4. Review of AI-Generated Material. Court personnel using AI tools are expected to understand the limitations of such tools. Court personnel must review their AI-generated material for accuracy, completeness, and potentially erroneous, incomplete, hallucinated, biased, or otherwise problematic output. Court personnel must use caution when relying on the output.

5. Training and Education. Court personnel should receive appropriate training on the use of AI tools and the advantages and drawbacks of using AI technologies.

G. Protocols for Use of AI Systems. Content that is put into a Generative AI tool is used to train the system and, in some circumstances, may become part of the public domain. Therefore,

1. Non-Sequestered AI Systems.

a. Court personnel are not permitted to put non-public content into a non-sequestered AI system. Examples include, but are not limited to:

(1) CJIS data.

(2) Confidential and personal financial records. Ariz. Sup. Ct. Rule 123(c)(3)

(3) Court proprietary content. Ariz. Sup. Ct. Rule 123(d)(4) and (e)(9)

(4) Family court records that are closed or deemed confidential. Ariz. R. Fam. Law P. 13(e)

(5) Information or documents filed under seal or subject to a protective order.

(6) Juror records. Ariz. Sup. Ct. Rule 123(e)(10)

(7) Juvenile court records that are closed or deemed confidential. Ariz. R. Juv. Ct. 215(a), 313(a), and 403(a)

(8) Mental health case records. Ariz. Sup. Ct. Rule 123(b)(14) and (d)(6)

b. Court personnel are responsible for identifying the nature of content to ensure that non-public content is not put into a non-sequestered AI system. If the nature of the content is not clear or cannot be determined by court personnel, they must obtain approval from judicial leadership before putting the content into the system.

2. Sequestered AI Systems. Judicial leadership is responsible for determining the content that court personnel are permitted to put into a sequestered AI system, based on judicial leadership’s assessment of the risks and the benefits, including what content is collected and retained and how it is used.

H. Public Facing Generative AI Tools. All public facing Generative AI tools must be thoroughly tested before being deployed, regardless of whether the Generative AI tool is court-developed, to ensure that false, illegal, or misleading information is not provided to the public.

I. Approval of Generative AI Tools. Approval of any Generative AI tool must use the process adopted by the administrative director. The administrative director will maintain, regularly update, and make available to courts a list of Generative AI tools that are categorized as: (1) approved for all purposes, (2) approved for public content only, (3) approved for non- production use only, or (4) prohibited. Court personnel are not permitted to use any Generative AI tool not listed unless the tool has been approved by the relevant court’s judicial leadership as set forth in K.

1. Approved for All Purposes. Court personnel may use “approved for all purposes” Generative AI tools in accordance with this section. If any court personnel is preparing work or completing a task for a judicial officer, the person must obtain approval from the judicial officer before using any Generative AI material to complete the work or task.

2. Approved for Public Content Only. Court personnel may use any “approved for public content only” Generative AI tool except when working with any non-public content.

3. Approved for Non-Production Use Only. Court personnel may use any “approved for non-production use only” Generative AI tool for investigative, testing, prototyping, or pilot purposes as approved by judicial leadership. Judicial leadership must ensure adequate security controls are in place as a prerequisite to approving any Generative AI tool for non-production use. Court personnel are prohibited from using live (production) court data that has not been redacted or anonymized in any “approved for non-production use only” Generative AI tool.

4. Prohibited. Court personnel are not permitted to use any Generative AI Tool that is categorized as “prohibited.”

J. Court-Developed AI Tools. If a court develops an AI tool, it must document the source of the training information and generally how it works. For purposes of maintaining the list of AI tools under I, judicial leadership must provide to the AOC general information on any court-developed Generative AI tools and their level of approval.

K. Local Policies. Judicial leadership is permitted to establish local policies and approve additional AI tools for local use consistent with this section. For purposes of maintaining the list of AI tools under I, judicial leadership must provide to the AOC general information on any locally approved Generative AI tools and their level of approval.

L. Continued Review. This section and local court policies must be reviewed regularly and updated as necessary to account for changes in AI technologies and to ensure compliance with all applicable laws, rules, regulations, and other policies. Adopted by Administrative Order 2024-207, effective October 30, 2024. Amended by Administrative Order 2025-24, effective January 29, 2025.