In AI we trust: how DoD’s Task Force Lima can safeguard generative AI for warfighters

This week is shaping up to be a milestone in how the US government approaches artificial intelligence, following an Executive Order from President Joe Biden. In this new op-ed, former Pentagon official Tony DeMartino argues that the Pentagon needs to make sure it doesn’t fall behind on the AI race, with an emphasis on using the newly-formed Task Force Lima correctly.

President Biden’s Executive Order issued this week orders the development of a National Security Memorandum to ensure the military and the intelligence community use AI safely and effectively. I recently argued that Pentagon leaders should create shared acquisition guidelines across the defense enterprise to capture universal standards when adopting commercial AI software. Nowhere is this more pressing than in the development and application of generative AI.

RELATED: White House AI exec order raises questions on future of DoD innovation

Some have argued for a temporary cessation on the development and public acquisition of generative AI models. But to the Department of Defense’s credit, two senior leaders pushed back for one main reason: potential adversaries won’t pause, and the US can’t risk falling behind. China is already experimenting with generative AI to spread disinformation; we need to keep pace and secure our competitive advantage.

The newly formed Task Force Lima at the Chief Digital and Artificial Intelligence Office should learn from prior Pentagon successes, like the “own the night” revolution in night vision optics, to ensure we own the ongoing revolution in data. The Task Force will make informed decisions and shape the implementation of generative AI models across mission sets from warfighting and readiness to health and business affairs. Part of their mandate includes reaching out to most innovative in industry to answer pressing questions around risk mitigation and safeguards.

As a founding partner at an innovation-forward national security advisory firm, I hear from AI companies daily. There are a number of industry-proven standards and safeguards Task Force Lima could establish now when looking to acquire lower-risk commercial software. I’ve captured some direct industry feedback so we can continue to push the limits on developing generative AI for the frontlines:

Not all generative AI is high risk. Generative AI means more than just hallucination-prone chatbots like ChatGPT. It also encompasses tools that can automate article summaries, synthesize intelligence reports or decipher if an image is artificially generated, saving analysts time and reducing risk. The risks that these secondary models pose, relative to the incident rate of inaccuracies, the spread of misinformation, and the presence of biases, is inherently much less as  compared to chatbots, and that risk continues to decrease when additional safeguards are implemented.

So, to avoid over-regulating some forms of generative AI and ensure our warfighters, analysts, and officers have access to indispensable tools, Task Force Lima should differentiate between different versions of generative AI to ensure risk is mitigated proportionally. Concerns about higher risk models shouldn’t slow down adoption of lower risk ones.

To speed the adoption of those lower risk models, Task Force Lima should quickly codify industry best practices and proven standards. Some technology companies are already safely deploying generative AI for specialized, controlled tasks and have implemented safeguards to ensure their model performs as intended. Creating and implementing these as interim standards drives responsible deployment and creates space to refine standards as the technology matures.

Start with Human-in-the-Loop. The most immediate  safeguard is standardizing human intervention in the model from development to deployment, in training, testing and evaluation, and refinement. In the training phase, humans can ensure the dataset is varied and intervene when data is labeled incorrectly. This intervention, in tandem with the fine-tuning process in the test and evaluation phase, is vital in ensuring the model performs as intended when presented with new data.

Build user trust by citing original sources. To increase trust in the model, users should understand the original sources the model used to generate an output, which requires the AI to list those sources as a standard part of every answer. This enables the user to refer back to the primary source for collection, verification, and additional context if necessary. Essentially, allowing users to “follow the breadcrumbs.”

Work with companies to gain access to large and varied data sources. As the number of quality sources increases so does the accuracy of the model. This means that the inverse is also true: If a model does not have access to a variety of data sources, the accuracy of the model decreases and can increase the risk of inaccuracy and bias. To take it to the next level, the model should be multi-modal, or capable of processing different modes of sources beyond text.

Bring in human subject matter experts. Not only is it important for companies deploying AI to prove a positive track record of trustworthy generative AI models, but they must also have access to subject matter experts – actual human beings with substantive knowledge of a given field. For example, a company focused on generating real-world crisis summaries should employ a number of former journalists, regional and functional specific analysts, and linguists, in addition to a team of data scientists and engineers. Leveraging expertise allows engineers to address inaccuracies before, during, and after user deployment.

The Department of Defense is a perfect test bed for the federal government to advance and realistically train on generative AI technology in a closed environment and a responsible manner. Whatever guidelines, best practices, or standards that Task Force Lima codifies must be shared across the federal government to inform the efforts of other departments and agencies.

To be ready to fight tonight, the national security community must rely on the strengths of the US industrial base and match the speed of innovation. Let’s get the safeguards and generative AI standards in place —and widely shared across the US government — fast. We’ll need every advantage we can get if we want to win the data revolution and the next war.

Tony DeMartino is a founding partner of Pallas Advisors, a strategic advisory firm specializing at the intersection of technology and national security. He served 25 years as an Army officer, with over five years of combat deployments in both Iraq and Afghanistan. He was most recently the Deputy Chief of Staff to the Secretary of Defense and the Chief of Staff to the Deputy Secretary of Defense, where his portfolio focused on technology and modernization.