Open the Pod Bay Doors: Navigating the Intersection of Public Records and Generative AI

by Sarah Doar

 

Over the last few years, several major software companies introduced their generative artificial intelligence (AI) programs such as ChatGPT, Microsoft Copilot, and Google Bard. These programs represent the potential for significant upheaval in the world of content creation, policy analysis, and decision-making. We covered some issues in the blog Use of Generative AI by Local Governments. As the dust settles, we are seeing that this potential, however, comes with certain drawbacks that have unique implications for agencies subject to the Public Records Act (PRA), Chapter 42.56 RCW.

 

What Is Generative AI?

 

As defined by Dictionary.com, generative AI is artificial intelligence that is “designed to process prompts from users and respond with text, images, audio, or other output that is modeled on a training data set.” Artificial intelligence is, of course, the capacity of any technology to imitate human intelligence and has been a part of our lives since the first computers became possible in the 1950s and hinted at by science fiction for far longer.

 

If the resulting text or images or other output are then used in the conduct of government, they are public records. For folks that work with public records on a regular basis, this is likely not a big surprise.

 

But it is the prompts — or information put into the AI program — that have public record implications an average user may not be aware of.

 

Inputs and Prompts Are Records?

 

Most likely, yes. Any digital information captured by agency systems while a public employee or official is going about their duties is likely a public record because it meets the definitions in RCW 42.56.010 for a “writing” and a “record.” In other words, “public record” includes “existing data compilations” which “contain information relating to the conduct of government or the performance of any governmental or proprietary function.”

 

This conclusion is based on a Washington State Court of Appeals decision where automatically generated “internet access logs” (IAL), which contain a record of every time an agency computer accesses the internet, were found to be public records. The court in Belenski v. Jefferson County (2d Div. 2015) wrote:

 

County employees use the Internet to obtain information to perform their work. Therefore, there is no question that the IALs record work-related Internet use on a county-owned computer. Accordingly, we hold that the requested IALs fall squarely within the definition of public records.

 

MRSC has relied on this case to advise that internet browser histories and search histories on agency computers are likely public records if they “contain information relating to the conduct of government.”

 

It is a short step to extend this analysis to generative AI prompts that are saved in the system. Like search histories, prompts contain information related to an employee or official’s duties.

 

But Must Prompts and Inputs Be Retained?

 

Just the prompts? Most likely, no. There is not yet a retention schedule that specifically applies to generative AI prompts. However, there is a schedule that applies to a user’s internet browsing history. Local Government Common Records Retention Schedule (CORE) DAN GS2016-006 considers internet browser histories to be non-archival, non-essential records that do not have a required retention period (i.e., they are transitory and can be destroyed once no longer needed for agency business).

 

If we consider a history or log of generative AI prompts to be the functional equivalent of browser histories, then the same DAN should likely apply. The Washington State Archivist indicates that the use of generative AI as a sounding board for thoughts and ideas would be transitory under Brainstorming and Collaborating, DAN GS2016-002, but if the output is used to develop an ordinance or strategic plan, another retention period might apply.

 

Depending on the program used, a user may be able to specify that the prompt or input should not be saved by the program. For example, Google Bard allows users to turn off activity history and it appears that ChatGPT just recently added this feature as well. However, this function does not appear to be available for Copilot.

 

And if a record exists on the date a records request is received, an agency is required to produce that record, even if the record is past its retention period (RCW 42.56.100).

 

Given These Implications, Should Local Governments Be Using Generative AI?

 

I think the ship has already left the spaceport on this one. Odds are folks at your agency are already using it, so it is important to get policies in place. As Governor Inslee’s Jan. 30, 2024, executive order notes:

 

[G]enerative AI has the potential to catalyze innovation and the rapid development of a wide range of benefits for Washingtonians; and … generative AI can enhance human potential and creativity but must be deployed and regulated carefully to mitigate and guard against a new generation of risks, harms, and perpetuation of existing inequities. [emphasis added]

 

To that end, WaTech, a state agency created through the consolidation of information technology divisions at other state agencies and which also houses the state Office of Cybersecurity, has adopted Interim Guidelines for Purposeful and Responsible Use of Generative Artificial Intelligence (AI) in Washington State Government.

 

The Interim Guidelines highly recommend that local governments adopt their own policies and procedures around the procurement and use of generative AI technologies. In the City of Seattle’s Generative Artificial Intelligence (AI) policy, it specifically addresses public records and records management:

 

6.1. All records generated, used, or stored by Generative AI vendors or solutions may be considered public records and must be disclosed upon request.

 

6.2. All Generative AI solutions and/or vendors approved for City use shall be required to support retrieval and export of all prompts and outputs (either via exposed functionality or through vendor contract assurances).

 

6.3. City employees who use generative AI tools are required to maintain, or be able to retrieve upon request, records of inputs, prompts, and outputs in a manner consistent with the City’s records management and public disclosure policies and practices.

 

Analogue Tools on the Digital Frontier

The PRA was originally adopted by voter initiative in 1972 as part of a larger “government in the sunshine” movement. The legislature has updated the PRA periodically through the years, often in response to courts being forced to apply a decidedly 20th century law to 21st century issues. As records professionals, it often feels like we are struggling to apply a more than 50-year-old law to a digital workplace that seems to be evolving at near lightspeed.

 

For what it’s worth, the film 2001: A Space Odyssey was released in 1968. More than 50 years later, perhaps records officers can empathize with the HAL 9000, a central character in the film, and say:

 

I am putting myself to the fullest possible use, which is all I think that any conscious entity can ever hope to do.