Section 2.1 - Assessment requirements, general

(a) Any first-party developer any legal person with direct involvement in the design, coding, training, or other development of an AI system
of an AI system shall be required to do the following¹:
- (1) Internally document the training the process of using a dataset to learn the parameters or otherwise build a model
  and evaluation process, specifically:
  - (A) The provenance of the training dataset any collection of data that is used to train a model
    and evaluation dataset,
  - (B) The person(s) responsible for maintenance of the system, and
  - (C) Any risk assessments that have been completed for the system.
- (2) Conduct rigorous testing of the AI system, specifically:
  - (A) Testing for disparities in performance,
  - (B) Evaluating the system’s performance for inputs that do not match the properties of any data on which the system was trained or evaluated (“out of distribution” data), and
(b) The Commission on AI Technology, as defined in Section 4, shall have the right to examine a first-party developer’s compliance with the requirements of Section 2.1(a) as they see fit. If the developer is found to be non-compliant, penalties as defined in Section 6 can apply.

These requirements are purposely written to be construed in a flexible way depending on the specific application. ↩