Even with AI governance policies, employee training and technical safeguards in place, hospitals and health systems face limits on how effectively they can prevent staff from sharing patient information with unauthorized AI tools.
“No hospital can eliminate this risk entirely. An employee with legitimate access to a spreadsheet can still paste it into an unapproved tool,” Muhammad Siddiqui, chief digital and information officer of Richmond, Ind.-based Reid Health, told Becker’s. “What we can control is how easy we make the approved path and how quickly we detect and respond when something goes wrong.”
The concern has taken on new urgency after Baltimore-based Johns Hopkins Bayview Medical Center disclosed that an employee mistakenly uploaded a spreadsheet containing patient information to an unapproved public AI platform Sept. 21, potentially exposing information belonging to 3,182 patients.
The spreadsheet included names, admission and treatment dates, department information, account numbers, medical record numbers, medications and treating provider names for patients admitted and treated between May and September. It did not include Social Security numbers or financial account information. The hospital did not identify the AI tool involved.
The hospital discovered the incident the same day, had the employee delete the prompt, conversation and document from the tool’s history, and requested that the AI company delete the information. Johns Hopkins has offered affected patients 12 months of free credit monitoring and is reinforcing employee training and policies governing the use of unauthorized applications.
The incident highlights a broader problem facing healthcare organizations: Many lack visibility into how employees use AI.
An October report from cybersecurity company Proofpoint and Ponemon Institute, which surveyed 705 IT and IT security practitioners at U.S. healthcare organizations, found that only 32% reported full visibility into employees’ use of unauthorized AI tools. Sixty percent reported lacking controls governing sensitive data used in AI systems.
At Reid Health, Mr. Siddiqui said the health system has established an AI governance policy based on the National Institute of Standards and Technology’s AI Risk Management Framework. Every AI tool must undergo committee review before use, and departments cannot independently purchase or subscribe to tools outside the organization’s contracting process.
Reid Health also offers approved AI options, including ambient documentation technology for clinicians and Microsoft Copilot for certain employees. Mr. Siddiqui cautioned, however, that adoption of those tools does not show whether the underlying risk is shrinking.
“One thing we learned while building the program is that more use of approved tools doesn’t prove less use of unapproved ones. We track unapproved AI use and data events separately,” he said.
Reid Health’s AI literacy program requires employees to identify incorrect AI responses and demonstrate an understanding of appropriate data handling. Uploading protected health information to an unapproved tool results in an automatic failure of the program’s assessment, regardless of the employee’s performance on other questions.
Still, Mr. Siddiqui said these measures cannot close every gap.
“Blocking a site on our network won’t stop someone from using it on a personal phone over home Wi-Fi. That remains a gap for us, too,” he said.
He also identified another growing challenge: Vendors are increasingly embedding AI features into software hospitals already use. Reid Health now requires those new capabilities to undergo separate review, even when the underlying product has already been approved.
For Raymond Lowe, senior vice president and CIO of Los Angeles-based AltaMed, addressing these risks starts with understanding where AI is already being used across an organization.
“The foundation starts with visibility,” Mr. Lowe told Becker’s. “Organizations need to understand how AI is already being used, identify higher-risk applications and use cases, and prioritize protection of sensitive information.”
Mr. Lowe said health systems should establish cross-functional AI governance involving privacy, legal, compliance, clinical, technology and business leaders rather than leaving oversight to IT and cybersecurity teams alone. That process should include evaluating AI applications before adoption, maintaining inventories of approved tools and continuously reassessing risks as new capabilities emerge, he said.
But both CIOs emphasized that restrictions alone will not solve the problem.
At Reid Health, Mr. Siddiqui compared making approved AI tools readily accessible to placing hand sanitizer where healthcare workers need it.
“If the only message is ‘don’t,’ some will go around you,” he said.
Mr. Lowe expressed a similar view, arguing that organizations should provide trusted AI options alongside safeguards and employee education.
“Ultimately, responsible AI requires organizations to make the approved path the easiest path,” he said.
The post Can hospitals stop employees from uploading patient data to unauthorized AI? appeared first on Becker's Hospital Review | Healthcare News & Analysis.