Reducing Risk for Government Agencies Implementing Generative AI
Personal experiences with ChatGPT or other AI tools may leave government agencies feeling like AI technology isn’t ready. They may have returned incorrect, biased, or untrustworthy information and that creates concern.
However, these tools aren’t an accurate representation of AI as a whole. It can be trustworthy. Agencies have the power to control the results. And they can return information that is safe and predictable, aligning with the White House’s recent Fact Sheet on AI.
Web and in-site searches may return tons of results but oftentimes they don’t answer the actual query. A well-trained AI engine with a curated content dataset is more likely to return the correct information the first time.
With so much at stake, how do we harness this emerging technology while reducing risk?
Curate your content
AI engines are only as smart, consistent, and accurate as the information you give it. Your AI engine needs good, high quality content…and a lot of it.
Through user research and discovery, defining who’s going to be using your AI engine and why is crucial to the success of your project. This up front work will determine what content supports their questions.
Some examples of content include:
- All relevant documentation, reports, test or research data, etc.
- Knowledge base articles
- Marketing content
- Relevant databases of information
Through content curation, you’ll remove outdated, conflicting, or inaccurate content. If conflicting information ends up in the dataset, it may lead to confabulation, also known as hallucinations. In other words, the AI engine will make something up to the best of its ability.
Government agencies should be aware of bias in content. Biased terms and phrases may still linger in legacy content and could show up in results.
Here are some examples of biased terms and phrases that are common:
- Blacklist/whitelist
- Handicapped
- Illegal aliens
- Mankind
Imagery can also train your AI engine to be biased. For example, if all your images of scientists are male, the AI engine may make the assumption that all scientists are male. It could then respond to queries with that understanding.
Content curation is a big job, but it is one of the most important steps to getting your AI engine to return trustworthy and accurate results. The more accurate information you can upload to the engine, the easier training it will be, and the better answers it will provide.
Train your AI engine
During the training process pairs of questions and answers teach the engine how to process connections between information. It’s common to have three sets of content for this process: a training set, a validation set, and a testing set.
- The training set makes up the bulk of your content dataset. This is the content that’s used when you’re getting your AI engine up and running.
- The validation set makes up a smaller amount of content. This set helps to ensure that the AI engine does not overfit, or learn the training content so well that it performs poorly on new information.
- The testing set determines how well the AI engine performs on new, unseen content.
Throughout training, you’ll be teaching the AI engine about linguistics and understanding meaning and intent. Once the AI engine is trained, it can be used for its intended purpose.
Gather metrics
As you would for any good digital product, you’ll want to track metrics. A few things you could track include:
- What questions are people asking?
- What was the quality of the answer?
- Did it solve for their query?
When you find opportunities for fine tuning, it’s time to adjust your content and retrain your AI engine to return updated results. Consistent, rolling content updates are critical to the longevity of your AI engine investment.
Implementing AI
The concept of implementing AI in the government space can be a scary one. In reality, AI has been around for nearly as long as computers. Processing speeds and other supportive technology has slowed its progress.
Having strong ownership and maintenance of your AI engine’s content will not only ensure your project gets off the ground, but that it can be trusted to return predictable, accurate results.
