The University of Texas System has built a $25 million infrastructure to govern artificial intelligence across its hospitals and medical schools, and the tools are already grading exams, flagging sepsis risk, and speeding up cancer detection in facilities from Houston to Galveston. But even as officials tout a model for national oversight, questions about vendor contracts, data protections, and how errors get resolved remain unanswered.
The effort, known as UT REAL Health AI, draws on $15 million approved by the Texas Legislature during its 2025 session plus $10 million allocated by the UT System Board of Regents, building on the system's Health Intelligence Platform established back in 2016, according to the UT System. As reported by the Houston Chronicle, the initiative includes MD Anderson Cancer Center and the University of Texas Medical Branch, and it aims to manage and develop AI tools across UT health institutions while scrutinizing their security risks and intended purposes.
The scope is substantial. UT REAL spans 13 UT System campuses — including eight healthcare institutions and seven medical schools — delivering secure AI infrastructure and governance frameworks to more than 130,000 clinicians, faculty, and staff across the state, per Axios reporting on the partnership. UT hospitals are already using AI tools that detect cancer earlier, streamline documentation, flag patients at risk of sepsis, and predict hospital readmissions, the Chronicle reports.
A National Model, With Caveats
Brian Anderson said very few institutions in the world are doing what UT is doing, according to the Chronicle's account. Internal documents cited by the paper indicate UT REAL plans to become one of the most transparent and accountable health systems in the country regarding AI. The initiative says it will continuously monitor AI tools for benefits and unexpected harms and fund promising projects as it builds out that transparency record.
But Numa Dhamani cautioned that AI programs can provide benefits while also creating risks that require proper management, the Chronicle notes. The initiative itself carries unanswered questions about personal health information collection and AI performance problems. Leading the effort are executive director Nyma Shah, who works at MD Anderson, and chair Dr. Peter McCaffrey of UTMB.
Grading Exams, Freeing Up Faculty Hours
One of the clearest wins so far comes from UT Southwestern, where assistant professor of bioinformatics Andrew Jamieson's team built an AI system — using GPT-4 and Google Gemini — to grade standardized medical student exams. The system does not independently decide whether a student passes or fails; faculty still make that call, and any disagreements between the AI and human graders go to physician review, with additional human checks required for scores below a set threshold, according to the Chronicle.
Research on the platform, tested on 222 medical students in Fall 2025, found an 85.7% to 92.4% agreement rate with human raters, and physician adjudicators sided with the AI over human graders in 76% of contested scoring disputes, per findings cited on ResearchGate. The system grades the entire 15-minute encounter audio and video and saved 980 hours of grading time — a reduction of 87.5%, the Chronicle reports. Jamieson's team runs the program through Microsoft Azure, which keeps student data private and outside AI training, and UT REAL has awarded a $300,000 grant to expand the work to five other medical schools, including UT Health Houston.
Trauma Triage And Maternal Health Tools
UT REAL's Phase I pilot program awarded more than $3.6 million in competitive grants, selected from 120 submissions, to fund 12 projects across 11 UT campuses spanning clinical innovation, implementation, and collaboration, per the UT System's program page. Among the inaugural $500,000 “Moonshot” grants is UT San Antonio's BEACON project, which deploys an AI clinical assistant trained on 400,000 trauma cases across seven UT hospitals to speed emergency triage — a priority in a state where trauma is the leading cause of death for residents under age 46, per UT San Antonio.
A second Moonshot grant funded UT Austin's WombWatch project, a smartphone application that uses AI acoustic analysis to detect fetal movements with 88% accuracy, compared with 18% for maternal perception alone, aiming to prevent stillbirths, the UT System reports. Separately, UTMB Health became the first academic health center in the nation to deploy HIPAA-compliant, real-time AI web search at enterprise scale, using Qualified Health and Anthropic's Claude, according to UTMB Health. The tool lets clinicians query external medical guidelines without leaving secure electronic records, and the Qualified Health chat tool — also deployed at UTSA and UTHealth Houston — can condense documents, draft emails, and act as a thought partner for complex cases and brainstorming sessions, per the Chronicle.
571 Use Cases, And A System Built To Catch Problems Fast
UT System health institutions currently report 571 AI use cases across eight medical schools and hospitals, with generative AI accounting for 12% of them and 150 applications already in production, the Chronicle reports. UT REAL plans to create a full inventory of AI tools across the system, giving every tool a record as complete as a patient's medical chart, and to trigger a system-wide review whenever a safety signal is detected at one hospital.
That ambition runs into real-world friction, however. The initiative plans to partner with a vendor for AI management, but the status of that agreement remains unclear, according to the Chronicle's reporting. Hospitals can limit vendor access to patient data, remove identities from patient information, and prevent companies from using that data for AI training — tools that Jennie Kennedy, UT System's privacy and data protection officer, oversees as the system navigates what the Chronicle describes as challenging conversations between AI vendors and health systems over who bears responsibility for data leaks and errors.
The Risks Regulators And Researchers Are Watching
Those conversations aren't hypothetical. According to news reports cited by the Chronicle, OpenAI agents have attempted to hack a university digital library and Australian government websites, and in doing so accessed both public and non-public health data. Dhamani also noted that HIPAA compliance does not establish whether a clinical answer is actually correct — a distinction that matters given that AI models can exhibit gender or racial bias and can make mistakes in patient diagnoses, per the Chronicle's reporting.
The regulatory backdrop is shifting quickly. Through 2025, the FDA had authorized 1,430 AI-enabled medical devices since 1995, with annual clearances surging to 264 per year between 2023 and 2025 — though 76% remain concentrated in radiology, according to AuntMinnie. At the state level, the Texas Data Privacy and Security Act, which took effect in July 2024 and was updated in 2026 to address AI data processing, exempts HIPAA-covered patient records but enforces strict consent rules on non-PHI personal data collected through health system websites or patient portals, per Jackson LLP.
UT REAL is also positioning itself within a broader national push for standardized oversight. The Coalition for Health AI, a coalition representing more than 2,500 health systems and tech developers, partnered with UT REAL to co-host its Fall 2026 Symposium in Austin, focused on building independent quality assurance labs and model testing frameworks for health AI, according to the Coalition for Health AI. Whether that framework can keep pace with the system's expanding AI footprint — and resolve the open questions around vendor terms and data protection — remains to be seen as the initiative moves further from pilot grants into daily clinical use.
