The rise of generative AI has led to an increasing awareness of the technologies that underpin it, including natural language processing, knowledge graphs, and large language models (LLMs), for example. With that new awareness comes an opportunity for important conversations about how those components determine the effectiveness and “fit-for-purpose-ness” of an AI solution, especially in mission-critical use cases.

Consider LLMs. A large language model is precisely what it sounds like: a neural network trained on very large volumes of text, images, and video so that it can recognize patterns in language, predict words, and generate sequences of characters that read like natural language. LLMs are how machines can answer questions, perform translations, and summarize or optimize human-generated text.

But just like doctors and attorneys have specialties, LLMs must be tailored for a specific purpose. You don’t want a podiatrist performing your heart surgery, and you don’t need a criminal defense attorney preparing your will. In the case of GDIT and our mission partners, you don’t need a geospatial model that can also detect tumors in X-rays. What you do need is domain expertise, because the risk tolerance in the mission space is fundamentally different from that of someone using ChatGPT casually. LLMs like ChatGPT have billions of parameters; if you can shrink things down, you can operate with fewer resources, extend access to your LLM to places like the edge, and get responses that are more accurate and actionable.

When using AI and LLMs for the mission, consider asking at least these five questions:

1. Will I Maintain Control of my Data?

Meaning, in-house data like emails, records, file shares, document repositories and internal documents. This is important because it’s all, ostensibly, trusted data that is chock full of language and lexicon unique to your specific mission space. Retaining full ownership lets you enforce governance, privacy and compliance policies.

2. Can I Securely Share That Data with the LLM?

Like any dataset, any organization’s own dataset is constantly growing. New context is introduced. New contributors and concepts come into play. It’s essential that teams can share that data with the LLM and ensure it’s protected at every stage.

3. What Does Continuous Improvement Look Like?

Can I continuously adapt my models? How does the LLM take in new information or changing context, and how are those changes persisted throughout the model? These are all key considerations to ensure the sustained accuracy and trustability of your LLM.

4. How Well Does It Understand the Mission Space?

Before I give it any new information, how well does the LLM understand the mission space and requirements? What data was it initially trained on and how did it perform? These answers will help determine if the LLM truly understands your mission, rather than some facsimile of it.

5. Is my Model Protected Against Data Poisoning?

It’s imperative that teams evaluate and select models that are free from manipulation, corruption, and bias as part of the supply chain process. They also need to know how data poisoning risks are detected, mitigated, and prevented through robust quality control processes. Starting this dialogue early helps avoid costly headaches later on.

Determining which LLM is right for you depends on its (and your) knowledge of your data, users and mission. Artificial intelligence, like ChatGPT, is a tool with a ton of utility, and it’s great when there is no need for a tailored LLM. It’s also a reflection of our current AI reality – where AI use is siloed and rarely coordinated across an enterprise. True mission AI demands tailored LLMs that can unify data sources, understand them well, and turn them into insights that are trustworthy and actionable.