How feds can create their own mini Chat GPT AI assistants

Parradee Kietsirikul/Getty Images

COMMENTARY | Configuring a mini-GPT can be a highly effective way to harness the awesome power of this new technology while mitigating many of the risks associated with generative AI.

Although artificial intelligence is extremely popular with the general public, government has been more hesitant to embrace the new technology. There are ethical concerns, which led to the White House releasing an executive order stipulating how AI should be used in federal service. And officials are also rightly concerned about issues with what would happen if an AI acting in an official capacity started to hallucinate or make up information.

On the ethical side, federal officials that I talked with following the release of the executive order were still excited about the potential for it to enhance services and assist with agency missions. They were also confident that AI could be employed by the federal government in ways that complied with the executive order.

That leaves the issue of AI accuracy, and its tendency to hallucinate or guess at answers that it does not know, as a major concern. One of the biggest reasons for the hallucinations is the fact that generative AI got its start as a tool that tries to guess what a user wants to hear, and the fact that it needs to know something about everything. Most generative AIs need to be ready to explain everything from why the sky is blue to what the secret ingredient might be in grandmother’s cookie recipe. And it has to be ready to complete tasks as varied as designing webpages to helping to write a business plan. The information that it has to draw upon is also massive, especially for AIs like Google Bard which are connected to the internet and able to download tons of information, and misinformation, from there.

But what if an AI could specialize in its required tasks while also being restricted in the information it draws from to perform its duties? Theoretically, a specialized AI with the same power as the more general-purpose generative AIs, but tasked with a very specific job, could be both extremely accurate and helpful in federal service.

For example, one of the often cited early missions for AI in government is streamlining procurement, which is extremely complex with hundreds of pages of guidelines, rules and regulations for overseeing the process. If a generative AI could be loaded up with all of that information, plus whatever other relevant data an agency uses to do things like issue requests for purchase orders and other documents as part of that process, then that specialized AI could potentially eliminate hundreds of hours of repetitive, labor-intensive work from humans while still complying with all relevant federal regulations.

That is the idea behind the new customized or mini-GPT feature added to ChatGPT. I’ve spent the past few weeks experimenting with how to create efficient mini-GPTs that could be tasked for federal service, as well as reviewing their efficiency. What I have found is that when this powerful technology is relieved of things like needing to know what soap the Kardashians like to use, it can instead focus on specific tasks within narrowly defined fields, and do so with almost no errors or hallucinations. 

Creating your own GPT

The following guide will explain how feds can create their own mini-GPT AI and task it for federal service. For this task, users must have access to the paid version of ChatGPT as the free one does not support this feature.

1. Getting started

Users should head over to the main ChatGPT page and click on the “Create a GPT” link to get the process started. The site will prompt you to name your specialized GPT and set up basic information about what you expect the new AI to do.

The new AI will have the same backend programming of the main generative AI, including all of the overall ethical guidelines and usage restrictions, but will otherwise be pretty much a blank slate ready for input.

For a procurement-focused AI, users might want to describe it as something similar to the following. “I want to create a GPT that is an expert in government procurement for my agency. I want to task the AI with issuing RFPs and other official documents that fully comply with all federal regulations and which target products and services that will best suit my agency’s needs for the best possible price.”

Note that this is just the first step in the creation process, but it’s important to set the overall goals here so the AI knows what you are attempting to achieve. After this process, you will be prompted to name the GPT and select an image to represent it. If you don’t have an idea for an image, you can have DALL-E, Chat-GPT’s graphical-generating component, automatically make one for you.

2. Configuring your unique GPT

Next you will want to click on the Configure tab to really begin to set how the mini-GPT will perform. You will see some sample questions that will be used as part of the GPT interface which you can configure if you want, but the most important step is building up the knowledge base that the new GPT will work from. This is all done from the Knowledge tab. You will first explain what you are uploading and then give it to the AI for analysis. It can read documents in a variety of formats including text files, word processing documents and PDFs.

For a procurement AI, you will certainly want to upload all of the relevant federal laws and regulations. In my testing, the size of the document does not seem to matter too much. Many of the PDFs I uploaded were hundreds of pages long. You might also want to provide the AI with good examples of the kinds of documents that you want it to generate. Because generative AIs are used to working with large language models, you won’t be able to overload it with information. In fact, the more you can share with it, the more accurate it will become.

Once you are finished, you click on the Publish GPT button to bring your new AI to life. You can always go back and reconfigure it later, so don’t worry if you forgot something or if the AI is not quite acting like you wanted. Which brings up the last phase in the process.

3. Testing, retesting and adding advanced features

The next step is to test the newly created AI to see how it performs. You can start by asking it a specific question from the materials that you shared with it, like requesting that it generate a list of all relevant Section 508 requirements that apply to IT purchases your agency wants to make. Most likely, the mini-GPT will perform flawlessly in this phase if you gave it enough information in the previous step. If not, you may need to go back and share more data with it, modify the documents it’s currently using, or tweak the goal statement that governs its overall behavior.

Once you are happy with the baseline functionality of the newly-formed GPT, you can multiply its efficiency by integrating it with relevant databases and other external sources of information using application programming interfaces or APIs. This is done under the Actions tab and requires some level of skill, as well as access to the requested resources, although the ChatGPT interface does a great job of streamlining this process.

As an example, you might want to link your procurement GPT with a live database of procurements that your agency is currently working on. That way, the GPT AI can see in real-time when new opportunities are made available and further learn about the process and the rules governing it. Linking to a live database using an API like that would also enable users to ask the GPT questions about those ongoing procurements such as “Does a specific RFP comply with all federal guidelines?” You might be surprised at how effective the new AI is at finding tiny problems hidden within the hundreds of pages of documents relating to some procurements.

Think of your new specialized GPT AI as an evolving tool that will need to grow over time. If you have tightly configured it, then it will be pretty impressive right from its creation. But even then, it will probably make small stumbles. Having humans correct its mistakes, feeding it more information or live data via an API connection and generally training it almost like you would a new human employee will help it to gradually become even more self-sufficient as it increasingly masters its assigned tasks.

Configuring a mini-GPT can be a highly effective way to harness the awesome power of this new technology while mitigating many of the risks associated with generalized, generative AI. It might be a perfect way to start bringing AI into federal service even while the overall technology, and the ethical and technical discussions surrounding it, continues to evolve.

John Breeden II is an award-winning journalist and reviewer with over 20 years of experience covering technology. He is the CEO of the Tech Writers Bureau, a group that creates technological thought leadership content for organizations of all sizes. Twitter: @LabGuys

NEXT STORY: Taking open source risks seriously