Building AI that Louisiana Can Trust

This is the last chapter in the series. The first three covered specific research programs. This one steps back and asks a harder question: how do you know whether to trust any of it?

In high-stakes settings, a powerful AI system that cannot be trusted may create more risk than value. This chapter covers what we have done to earn trust, what makes our program different from peer institutions, and what I think all of this adds up to for Louisiana and beyond.

Detecting When an AI Encounters Unfamiliar Data
Every AI system is trained on a limited range of examples. When it encounters data that differ substantially from that range, its predictions may become unreliable even if its confidence remains high. We built COMBOOD to detect that condition before it causes harm.

Think about what we are deploying. A wildfire detection system. A drug-reaction monitor in a hospital. An antibiotic resistance classifier for a patient with a serious infection. These are not low-stakes applications. A confident wrong answer can cost a life.

All AI systems have limits. They are trained on specific data. When they encounter something genuinely outside that experience, a malfunctioning satellite sensor, a hospital record formatted in an unusual way, a bacterial genome combination never seen before, a poorly designed system does not pause or flag uncertainty. It produces an answer that looks just as confident as it does when it is right.

Our COMBOOD system addresses this directly. It adds a monitoring layer to any AI that continuously asks: does this new input actually resemble what the system was trained on, or does it differ substantially from that training distribution?

COMBOOD combines two detection signals. One measures how far a new input sits from the nearest familiar examples in the AI's training data. The other measures how far it drifts from the overall statistical pattern of the training set. Each signal catches what the other misses. Together they flag anomalies reliably, whether the departure from training data is subtle or obvious.

The results on the OpenOOD international benchmark, which tracks the state of the art in out-of-distribution detection, are concrete: COMBOOD achieves 95.22 percent AUROC on near-OOD detection and 90.24 percent AUROC on far-OOD detection for the ImageNet-1K dataset. On the benchmark and metrics reported in our study, COMBOOD outperformed the comparison methods evaluated. We also tested it on a practical high-stakes case: detecting unusual or damaged documents fed to an AI reading handwritten forms such as tax records and medical prescriptions, where a confident wrong reading causes direct, traceable harm.

For high-stakes applications, out-of-distribution monitoring should be treated as a core safety requirement rather than an optional feature. Before a fire manager, physician, or infection-control nurse relies on one of our systems, the system should be able to flag inputs that differ substantially from those used during training. COMBOOD provides that additional safeguard.

The research community recognized this work. At SDM 2024, the SIAM International Conference on Data Mining, COMBOOD received the runner-up Best Poster Award.

 

What Sets This Program Apart
Many leading universities organize AI research around individual application areas such as medicine, remote sensing, or climate modeling. CC&E's program is unusual in bringing environmental intelligence and human health together within a single Louisiana-centered research agenda.

Most peer programs are organized around a single domain. A drug discovery center works on drugs. A remote sensing lab works on satellites. An environmental AI group works on climate. That is a sensible structure for a focused research lab. It is not well suited to Louisiana's problems, which cut across those boundaries.

The same coastal community that faces hypoxia from agricultural runoff also faces above-average cancer rates and rural pharmacies with limited hours. The same warming climate that dries out Louisiana's forests and accelerates permafrost thaw in the Arctic also creates conditions for new infectious diseases to reach here first. These problems are intertwined. Treating them one domain at a time misses the connections.

CC&E is unusual in combining environmental intelligence, human health, and AI safety within one program. That combination exists here because Louisiana's challenges demanded it.

The program's validation record extends beyond benchmark performance. DeepFire was tested in a live operational environment against the world's best wildfire detection systems over New South Wales during the XPRIZE Wildfire Competition and outperformed NASA's FIRMS monitoring service on early detection timing. DeepDrug advanced to the semifinals of the $5 million IBM Watson AI XPRIZE among 147 teams worldwide, and a combination therapy it identified for COVID-19 progressed to human trials. DeepSAT shaped a decade of satellite AI research globally and received the runner-up ACM SIGSPATIAL 10-Year Impact Award. COMBOOD received the runner-up Best Poster Award at SDM 2024.

Our program places particular emphasis on carrying computational results toward operational or clinical use. A drug candidate that never reaches a patient is incomplete research. A wildfire model that never informs an operational decision is incomplete research. Our drug discovery work feeds directly into LSU Health clinical researchers working with Louisiana patients. Our carbon capture siting framework is informing Louisiana's state-level carbon capture and storage program. The program is designed from the outset to connect computational research with clinical, environmental, or operational partners.

 

Winning in Louisiana for the World
Louisiana is a living laboratory. The Dead Zone, drug-resistant infections in underfunded hospitals, rural pharmacies with limited hours, infants on breathing tubes without overnight nursing coverage, a coastline retreating faster than almost anywhere on Earth. These are not abstract research scenarios. They are documented, measurable conditions affecting real people, and they demand solutions that work under real constraints.

The work we are doing here is not Louisiana-scale in its eventual reach. It starts here because this is where the problems are most acute, where we have the data and the partners, and where the design constraints are honest enough to make the solutions robust.

The Dead Zone off our coast is the most studied hypoxic zone in the world. The AI we built to forecast it can be adapted to dead zones in the Chesapeake Bay, the Baltic Sea, and the coast of China. The antibiotics we identified that work against drug resistant pathogens will not only protect Louisiana patients. These superbugs kill people in hospitals on every continent.

The drug-reaction monitoring system built for rural Louisiana hospitals with limited overnight pharmacy staffing is equally relevant for rural hospitals in India, Brazil, and sub-Saharan Africa facing the same constraint. The infant monitoring system was designed for a rural Louisiana facility with limited staffing and limited data infrastructure. Designing for those constraints may improve the system's portability to other settings, although deployment elsewhere would still require local validation. That was the design logic from the start: build for the hardest case.

This is what I mean by winning in Louisiana for the world. Not that we build something here and export it as an afterthought. The Louisiana constraint is what makes the solution specific enough to matter and robust enough to work in analogous circumstances elsewhere.

 

The Obligation to Get This Right
AI is being deployed at scale in high-stakes settings right now, in medical diagnosis, emergency management, criminal justice, and financial decisions. In most of those applications, the system does not come with a reliable way to flag when it is operating outside its competence. That should worry everyone. It worries me.

The research we have done on AI safety, the COMBOOD work, is partly a technical contribution. It is also a statement of principle. We will not deploy a system that cannot acknowledge its own limits. A physician should not have to take an AI's word for its own reliability. An emergency manager should not have to trust a fire risk alert without knowing how confident the system is and why.

Building capability and building safety are not separate tracks. They run together. Every tool in this program has both. That is a harder and slower way to work. It is the right way.

 

Where This Ends Up
Louisiana is a complicated place. The coast is retreating. The fisheries are under pressure. Cancer rates are elevated. Rural communities are underserved in ways that do not get fixed easily. The state faces all of this with a directness that I have come to respect.

The research program this series has described is our response. Not a complete one. There is more to do, and some of what we are working on now will take years to reach the people who need it. But the direction is clear. Build teams that span disciplines. Root the work in real Louisiana problems. Validate against the hardest available standards. Make sure the system can flag its own limits. Get the results to the people who can use them.

That is what building teams that win in Louisiana for the world looks like in practice. The work is underway.