This document, governing the ethical use of AI/artificial intelligence software, is offered for adoption by government agencies, including state/county agencies, municipal managers, police, fire, or emergency medical services agencies, emergency management agencies, public works and public schools as well as by public information officers.

Background:

The catch-all phrase “artificial intelligence” or “generative AI” may refer to any of a family of software and text-prompt products using generative pre-trained transformers (GPTs) or comprise the larger family of software known as “large language models” (LLMs).

In the context of LLMs, artificial intelligence does not exist.

Artificial intelligence, or the ability of machines to independently learn information, create new information and be perceptive and aware of their environment, does not exist in large language models. 

LLMs generate a response based on relationships between words, customized by both prompts and certain datasets that are installed (this process is called “training”) or prioritized by the LLM’s programmer. LLMs are subject to errors and so-called “hallucinations,” in which the software outputs data that is not only wrong but completely ungrounded in reality.

LLMs, including popular and emerging chatbots from ChatGPT, Grok, Copilot and Gemini, can process data and output text, sounds, videos, graphics and data visualizations. It can also analyze large datasets very quickly. 

The software has its uses, and it may become vitally important for government and public safety agencies, but it must never be allowed to be used as a substitute for or replacement of human interaction between constituents and public agencies. 

Statement of Principles

There is the potential for significant boosts in productivity and efficiency to the sector of government communications offered by large language models. There are also tremendous risks, including the risk of massive erosion in public trust in communications received from their government and public safety agencies, as well as the risk for delays in the delivery of government services or the arrival of first responders during an emergency.

Stipulation: “Artificial intelligence” and LLM software (the software) can not replace human communication and human intelligence. 

Therefore, as responsible public information officers, press secretaries and government communications directors, we adhere to the following MUST, MAY and MUST NEVER principles:

We Must:

• Utilize the software with transparency. We must declare to our constituents when it is being used, through a disclaimer statement, label or watermark
• Verify all claims, statistics and facts offered by LLMs.
• Be aware that the software is developed by private entities, including large for-profit corporations and foreign organizations. Utilizing one of these products comes with the understanding that it is programmed and trained by a third party who may not have the best interests of our constituents in mind. Data and responses given by the software will be affected by the programming and training given by its creator, and all data entered into the LLM will subsequently be available to the LLM. Do not enter confidential information into an LLM.
• Understand that the technique of reinforcement learning from human feedback (training the software) is important and useful to ensure more effective utilization of this suite of software, but no amount of training or feedback will make an AI/LLM product provide absolutely perfect outputs. 
• Understand that no software product may be “ready for live” right out of the box. We must commit to learning how to use the software – training it in the methods and priorities of our agency, not just those of a third-party corporation, before deploying it in a live environment.
• Consistently train ourselves and our staff on this ever-evolving technology. One-time training is not sufficient.
• Assign responsibility to a human for all errors made by LLM software. The human user must be designated as the responsible party. All errors, corrections or omissions are the responsibility of the human user.

We May:

• Utilize the software to create first draft content. A first draft, subject to editing, fact-checking and copyediting by a human employee makes LLM-generated content more acceptable for constituent receipt.
• Instruct the software to analyze data sets. Survey results, financial reports, and ballot reports create large datasets. The software may be used to give the government user analysis and to create data visualizations. We assume responsibility for the analysis, and it is incumbent upon us to verify and spot check all claims made.

We Must Never:

• Utilize the software to deliver communications directly to constituents. We must never post AI/LLM content directly onto websites, social media networks, newsletters or other platforms without human review, editing, fact-checking, and approval. Human employees of the government agency must always be the last eyes on content produced by the software, and human eyes must always review content produced by the software before it is published or offered to constituents.
• Utilize the software to communicate with 911 callers. Emergency communications centers are sacrosanct. The connection between a constituent and a 911 emergency telecommunicator allows for the rapid and efficient deployment of first responders and the delivery of vital instructions including emergency medical information. A 911 call taker using AI/LLM software to assist them in locating instructions or making recommendations is acceptable, but the software cannot speak to the constituent. 
• Utilize the software to communicate with unknowing constituents. The software may be offered as a fast “constituent services option” in lieu of waiting for a human response, but constituents must always be aware when they are communicating with software or a real government employee.
• Utilize the software as a replacement for human writing in police, fire and EMS incident reports. Our courts of law are run by human beings at all levels. Reports and items that may end up as narrative evidence in court should be written by human beings only. In cases where AI/LLM software generates reports automatically for things like surveillance cameras and body worn cameras, a human being must review and sign their name to – and assume responsibility for – the content.
• Violate the intellectual property of another party. Government communicators often create content that is automatically entered into public domain. We have an obligation to safeguard the works of others, to not violate their intellectual property rights, and to avoid taking actions or publishing LLM materials that inadvertently violates copyright or trademark. Government violations of copyright and trademark can result in cascading effects on the original party, causing great damage. Take great care to check LLM-provided materials for copyright or intellectual property violations before publishing.

Special thanks to Mark McClennan, APR, for his invaluable feedback.

John Guilfoil’s Statement of Principles Governing the Ethical Use of Large Language Models by Government and Public Safety Agencies by John Guilfoil is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International