5 General Principles on AI Development

For organizations that are not specifically designed to produce General Use AI, it is best practice to ensure that all AI development projects are directed to specific business objectives. The 5 general principles on AI development listed below are designed to contain cost and risk to the organization. Again, this is not applicable to companies that intend to create General Use AI systems, but serves as a foundation for organizations that are developing internal AI based solutions.

1.   AI based projects should minimize overlapping functions, and any gain of function experimentation must be assessed for risk and authorized.

2.   All AI development projects must adhere to the Containment Principle. This means that AI systems must not be capable of self-replication (in total or in part) outside of the intended development or production environment.

3.   AI systems under development and in production shall only have access to training data that is required for the target objective. No AI system shall be given unrestricted access to the open Internet and all content therein, for training or other purposes. Further, all training data access must be evaluated as to the classification of the data and any associated privacy and/or security controls.

4.   While it is expected that AI systems will gain knowledge over time, gain of capability through self-directed improvement and self-directed code modification must not be supported. All system changes must be reviewed, understood, and approved the project leadership.

5. All AI generated actions or content which will be relied upon for critical business decisions or leveraged in a way as to have significant impact on customer security shall be explainable or validated for reasonableness by human subject matter experts.