DIR Chatbot Notice Offers Look at AI Disclosure Expectations

A new disclosure for the Texas Department of Information Resources’ (DIR) website chatbot gives technology vendors a preview of the information state and local government customers may seek when procuring and deploying artificial intelligence systems.

The department published an AI System Facts notice for DIR’s Assistant for Search and Help, known as DASH. The chatbot was launched in July 2024 to help visitors find information about department programs and services.

The document identifies the system’s underlying model and developer, explains how government data supplements the model and addresses testing, human oversight, retention, feedback and encryption.

The notice is an example of a standardized template developed by DIR under state law. Texas agencies and local governments must use the notice for AI systems that are public-facing or serve as a controlling factor in a consequential decision. The notice must be included on associated applications, websites and public computer systems.

The form asks government entities to identify the baseline model and developer, disclose whether agency-specific information supplements the model and state whether user inputs are used for continued training. It also addresses external data sharing, pre-deployment testing, human review, automatic deletion, user feedback and protections for sensitive information.

Those questions could affect how vendors describe their products during procurement, security reviews and contract negotiations. Although the notice does not establish a solicitation requirement by itself, suppliers may need to provide agencies with enough information to complete the public disclosure accurately.

DIR’s completed notice says DASH uses Google’s Gemini-2.0-flash-001 model. The department website is the chatbot’s main source of information, supplemented by a list of frequently asked questions through retrieval-augmented generation. User inputs are not used to train the system following implementation and inputs and outputs are not shared outside the agency, according to the document.

The department also reported that subject-matter experts tested the chatbot for content, accuracy and accessibility before deployment. Staff periodically review conversation histories and update the supplemental information used to generate responses.

DASH retains inputs and outputs for two years before deleting them. Users can submit feedback through thumbs-up and thumbs-down controls. Conversations are encrypted while being transmitted and while stored within Google’s infrastructure.

The completed disclosure does not indicate that every vendor offering AI products will face identical requirements. It does, however, show the categories of information Texas government entities must be prepared to publish about certain systems. Vendors unable to explain model ownership, data use, retention, testing and oversight may have difficulty giving public-sector customers the documentation needed to meet those obligations.