AI roundup: Power redistribution | Governance playbooks | LLMs vs. search engines

News and views that high-level healthcare AI watchers ought to know about. 

1. That’s a lot of co-authors for a single set of AI implementation aids. Then again, there’s a lot of AI to implement in healthcare.  

More than 150 health AI thought leaders from more than 100 healthcare organizations contributed—in one way or another—to a new set of AI-adoption “governance playbooks” for hospitals and other provider entities. The yield is the fruit of an effort directed by CHAI, the Coalition for Health AI.  

CHAI sought and received expert input from far and wide to make sure the playbooks reflect the needs of diverse care settings, according to a May 27 announcement. The organization also expects the little library to help provider orgs earn certification from the Joint Commission once that group, a CHAI collaborator, comes out with a voluntary program it’s soon to release. 

CHAI says the guides, which cover eight critical components of responsible AI use, are readily adaptable into existing processes and contexts. The content is ideally suited for AI governance committees but is also useful for stakeholders outside of those specialized teams. 

CHAI chief executive Brian Anderson, MD, says the playbooks will help U.S. healthcare “define responsible AI” while also making the technology accessible to healthcare delivery organizations across the country “regardless of resource level and [united by] the goal of translating AI innovation into high-quality care for every patient.”

Announcement here, playbooks here.

2. More people are turning to LLMs for helpful answers to subjective questions. Might they be better off sticking with web search engines?

It depends, according to researchers at the University of California, Riverside. Going straight to standalone AI will do better at giving you hard facts and logical reasoning—pretty much what one would expect of a robot, the researchers found. Web searches, including their instant AI overviews, are more likely to sprinkle in humanlike touches of moral, practical and emotional considerations.  

Computer science PhD student Taukir Chowdhury and co-authors presented their findings in May at the ACM Web Science conference in Germany. (ACM = Association for Computing Machinery.) The subjective questions the team used for the experiment included, to name two examples, Should vaccines be mandatory? and Is remote work better for productivity?

“LLM responses, by default, do not entirely reflect the diversity of opinions present in online sources and more frequently rely on epistemic or predominantly logical forms of justification,” Chowdhury and colleagues report in the conference’s published proceedings. 

“Since LLMs are becoming key tools for accessing information,” the authors add, “it is important to understand where their reasoning differs from the Web and how this affects the diversity of viewpoints they present.”

In coverage of the work by UC-Riverside’s news operation, co-author Vagelis Hristidis, PhD, warns: “As people increasingly rely on AI systems for information discovery at the expense of traditional web searches, the web may gradually lose its soul and cease to reflect the human nature that has shaped it over the past 25 years.” 

The bright side of his dark prediction: “This may give rise to new information dissemination platforms in the future.”

Study report here, university news item here.

3. In healthcare as elsewhere, AI has changed the balance of power.

The experts—mainly physicians—no longer have a monopoly on authoritative information. Patients can compare and contrast what their doctor tells them in a visit with what AI told them before they came in. To be sure, the physician’s advice still has to trump the machine’s. There’s more to good care than raw information. (See item above.) 

But the patient-doctor exchange is a lot less one-sided than it was for many generations. In an opinion piece published May 28 by Forbes, the chief innovation officer of the American College of Cardiology gives a concise overview of what AI has changed, what it hasn’t and what’s to come. 

“The end of institutional gatekeeping does not mean the end of institutions,” writes the CINO, Ami Bhatt, MD. “It means their role has to mature.”

More from the piece:  

• ‘Today the ability to interpret information is available outside of institutions. This does not mean expertise is going away. Instead, the power to understand information is being shared more widely.’
   
• ‘Can institutions evolve fast enough to shape how distributed intelligence operates within clinical care and avoid fragmentation? If we build in validation, transparency and monitoring from the start, sharing power can build trustworthy healthcare delivery instead of weakening it.’
   
• ‘AI has already moved interpretive power beyond traditional walls. The question now is not whether patients and clinicians will receive guidance from new sources. They will.’

Get the rest.