Last week Kyle and I attended a virtual meeting of the Oregon Artificial Intelligence Advisory Council. The featured speaker was the Chief Data and AI Officer for the State of Vermont. Vermont has had an AI advisory council since 2020, a bit before the rise of ChatGPT and the like, so they had a head start developing guidelines on the ethical use of AI in the work place. His presentation was quite thought provoking and I thought I would share with you Vermont’s AI use guidelines as food for thought. The entire document is an interesting read and I’ve provided a link to it at the end of this article.
From the Vermont Guidelines:
- LLM Usage Guidelines: Employees must use LLMs like ChatGPT, Bard, Bing Chat, and LLaMA in accordance with State Personnel Policy 11.7, the AI Code of Ethics, and these guidelines.
- Employees must get supervisory approval before using LLMs for official state business. Supervisors may consult with the Director of AI or their ADS IT Lead to help them decide whether a use is acceptable.
- Employees may not input non-public information into LLMs. All information put into an LLM becomes public. The following is a non-exhaustive list of information that may not be used in LLMs:
- Confidential or privileged information or communications
- Personally-identifying information (PII)
- Protected health information (PHI)
- Code containing passwords/secrets
- Information that has the potential to erode trust in Vermont’s institutions
- This policy only enables employees to use LLMs that are available free of charge. Employees may not pay for a LLM services or sign up for services requiring payment. Any purchase of such products must go through the mandated state procurement processes.
- Employees must ensure the accuracy of the information obtained from LLMs. LLMs may generate content that is incorrect or fictitious. This content may seem reasonable and not be readily distinguishable from factual information. Employees using an LLM must review all information obtained from the LLM for accuracy, veracity and completeness, as with any other source of information.
- Employees using LLMs are responsible for their work product, regardless of what portion of it is produced by the LLM, and must be willing to sign their name to it. LLMs have demonstrated biases which can negatively impact groups or individuals and are not in keeping with Vermont’s values. The user must review all content for accuracy, completeness, and alignment with Vermont’s values. It is not acceptable to blame any deficiencies on the AI.
- Employees using an LLM for official state business must log in and create an account using their state email address only. Official business may not be conducted using an account established with a personal email address.
- Do not use LLMs in a way that could cause reputational harm to the state if it were known.
- While it is acceptable to use LLMs to perform official job duties, it is not acceptable to rely on them to perform those duties. If there is an opportunity to make Generative AI or LLMs a part of a standard work process ADS and the AI Director will assist with procuring and implementing a robust AI-powered solution. Employees should not develop reliance on free LLMs to accomplish their official duties, as they have significant down times. Treat it as an optional enhancement.
- Employees may not use LLMs in any way that infringes on the intellectual property rights of others.
The Vermont guidelines also advocate users cite when and how generative AI was used in the creation of content:
“To maintain trust in state government and meet the requirements of transparency set forth in the code of ethics, it is important to appropriately cite the use of AI where required:
Standard Citation Format: This content was [drafted, edited, translated] with the assistance of a generative artificial intelligence, [Bard, ChatGPT, LLaMA]. The content has been reviewed and verified to be accurate and complete, and represents the intent of [office, department, the State, or a person’s name].”
He suggested that AI should be viewed as a “power tool”; a screw gun isn’t going to build the whole house, but it can certainly be a big help. Again, food for thought as we wade into how best to make use of this powerful new power tool.
Source:
“Guidelines for Use of Content Generating AI”; State of Vermont Digital Services;
November 29, 2023: https://digitalservices.vermont.gov/sites/digitalservices/files/documents/Guidelines%20for%20use%20of%20Generative%20AI.pdf
