Implementing LLMs in healthcare? First, do some harm reduction

By viewing large-language AI models through the lens of harm-reduction thinking, healthcare adopters can cultivate a responsible, ethical and optimal integration of the technology. And they can do so while actively mitigating the inherent risks for all stakeholders involved.

Two researchers in Sweden lay out the how’s and why’s in a paper published July 25 in the Journal of Medical Internet Research. 

“Harm reduction, traditionally applied in public health contexts such as substance use management, focuses on minimizing the negative consequences associated with certain behaviors rather than seeking solely to eliminate the behavior itself,” explain Birger Moëll, MSc, of KTH Royal Institute of Technology in Stockholm and Fredrik Sand Aronsson, MSc, of the Karolinska Institute. 

“Applied to the use of LLMs in medicine,” they add, “this means acknowledging their inevitable use by both patients and professionals and proactively developing strategies to make that use as safe, ethical and beneficial as possible.”

Here are excerpts from the case they make. 

1. A harm-reduction lens, adapted from public health practice, offers a pragmatic vision between prohibition and uncritical adoption. 

“For patients, the priority is to transform passive consumption of model output into a critically verified information-seeking process that preserves privacy and promotes timely care-seeking,” the authors write. “For clinicians, the core insight is that LLMs can safely augment but never replace clinical judgment when they are used within secure, governed environments that mandate human verification, bias checks and transparent disclosure.” More:

‘Implementing these measures demands institution-wide policies, continuous training and interdisciplinary oversight but can preserve patient safety, equity and trust while unlocking administrative efficiencies and decision-support benefits.’

2. By advocating for a human-in-the-loop baseline, we can guarantee safe deployments of AI systems in the medical domain.

Once these systems are deployed and work well, evaluation can become more automated, Moëll and Aronsson point out. “This approach is also in line with how work might change through AI in other sectors,” they add. “Many occupations might have tasks involving the orchestration and evaluation of LLMs and LLM agents.”

‘As such, human-in-the-loop assurance seems like a sensible, safe way forward.’

3. Adopting a harm-reduction framework acknowledges the inevitability of LLM use by patients and clinicians.

It also prioritizes strategies to mitigate harm rather than promoting “futile attempts at prohibition,” the researchers state before adding: 

‘Ultimately, harm reduction is not a barrier to progress but a necessary, pragmatic pathway to ensure LLMs enhance, rather than undermine, the core values of equitable, evidence-based and patient-centered medicine.’

Read the whole thing.

• In other research news:
   
  • University of Texas at Austin: New AI tool accelerates mRNA-based treatments for viruses, cancers, genetic disorders 
     
  • American Academy of Family Physicians: Machine learning model predicts missed appointments in primary care clinics
     
  • Technical University of Denmark: AI turns immune cells into precision cancer killers—in just weeks
     
• Funding: 
   
  • Digital health faces Q2 ’25 pullback: Funding falls to 5-year low