If you're a veteran trying to use the GI Bill, retrain for a new career, find housing support, or access healthcare, the US Department of Veterans Affairs (VA) has an extensive body of information explaining what is available and how to apply. But knowing that a resource exists is different from knowing how it applies to your circumstances, which form you need and what step to take next.

AI offers an obvious way to navigate that complexity. But a general-purpose language model raises another question: can its guidance be trusted when the information is specific, consequential and personal? In healthcare, housing and benefits, useful guidance depends not simply on retrieving information, but on knowing its source and relevance - and recognizing when specialist judgment is required.

Sidekick is a case in point. Co-developed by Research Professor Skip Rizzo, director of medical virtual reality at USC Viterbi and USC Stevens School of Computing and AI's Institute for Creative Technologies (ICT), the conversational AI platform builds on nearly three decades of his virtual-human research at ICT, a US Army University Affiliated Research Center (UARC).

Recently approved by the VA after months of review for privacy, safety and accuracy, the first implementation of Sidekick uses thousands of pages of information selected by 10 VA subject-matter experts across areas including healthcare, employment, housing and the GI Bill. Veterans can ask questions in ordinary language and follow the conversation to relevant resources, applications and courses of action. When an issue calls for individual judgment, Sidekick can direct them to appropriate human support.

A division of labor

"The challenge of the next decade will not simply be providing more information," said Rizzo. "It will be helping people confidently navigate trusted information that specifically addresses their needs."

Sidekick is not intended to become an autonomous expert. Subject-matter experts determine the information, scope and boundaries; AI organizes and retrieves that knowledge, conducts the interaction and makes it available through conversation, with a route to human expertise when needed.

"This isn't about replacing jobs, but filling in gaps where there aren't people in those roles in sufficient numbers to provide all the support veterans need," Rizzo said.

Rizzo recently brought that position to Surrounded, a debate program in which 20 participants face a single celebrity opponent. Cast as an "AI optimist," Rizzo considers himself an "AI realist." His opponent turned out to be Andrew Yang, the 2020 Democratic presidential candidate known for his warnings about automation and AI-driven job loss.

"We can't put AI back in the box," Rizzo said. "It's changed how we interact with technology. We need to understand the risks and recognize the opportunities. When carefully combined with human judgment, AI can significantly expand access to knowledge and our capacity to make informed decisions."

Intelligence and ethics

Rizzo's approach was shaped by research that began well before contemporary generative AI. In 2009, SimCoach gave service members and veterans a private interface for discussing mental-health concerns and finding routes to care, while encouraging users to seek clinical care when needed rather than attempting diagnosis or therapy. In 2013, SimSensei added real-time analysis of facial, vocal and other behavioral signals associated with psychological distress. More recently, Battle Buddy brought the patient-facing work to mobile, helping veterans build coping skills, develop plans for staying safe during periods of crisis and remain connected to their support networks.

Across these projects, Rizzo and his collaborators repeatedly had to establish what virtual humans could responsibly do - and where responsibility should rest with a human agent. "We've had to deal with all the ethical challenges" of deploying virtual humans for patients and clinician training, Rizzo said. The team developed and documented approaches to privacy, safety, clinical scope and human oversight that, he argues, now provide a framework for using modern AI "on a larger scale, but with those ethical guardrails in place."

Scaling Sidekick

The veteran implementation is one application of a platform designed around different bodies of organizational expertise. Sidekick is being developed for uses including healthcare, education, nonprofits and rehabilitation. Rizzo, with long-term research partner Russ Shilling, is currently developing a conversational support app for parents navigating the first 100 days after a child's autism diagnosis, translating evidence-based guidance and practical resources into accessible, personalized support.

"Sidekick is all about making resources more easily known, available, and authorable," Rizzo said. "I could never have done that work anywhere else other than at USC, and within ICT," he said. "Our research has evolved in sync with technology, and now we have the tools with AI to really manifest the dreams that we had with the earlier virtual-human projects."

The technical possibilities have changed dramatically since those early projects, but the questions governing responsible use - what information deserves authority, what to automate and where a system's competence ends - were developed through research with real users. "Our research isn't ultimately about product development. It's primarily about idea development," he said. "It's the inspiration of academia and the testing of these ethical and practical principles that drives innovation in society on a large scale."

Published on October 9th, 2026

Last updated on October 9th, 2026