Ethical Implications of Small Language Models (SLMs) in Healthcare Applications

Although Small Language Models (SLMs) show promise for healthcare applications, their use also introduces ethical risks that must be evaluated to develop appropriate mitigation strategies. Within a group activity during the SLM4Health workshop with 13 participants, we assessed ethical risks using the Digital Ethics Canvas along five lenses (welfare, autonomy, privacy, fairness, and sustainability). The participants considered two SLMs application use-cases: clinical documentation and clinical decision support. For SLMs in clinical documentation, identified risks included documentation errors from hallucinations or omissions leading to potential medical mistakes, social and data biases that can affect underrepresented groups, and a potential erosion of clinicians’ skills. For clinical decision support, challenges included biases harming vulnerable groups, unequal data quality exacerbating healthcare disparities, and privacy risks due to exposure of sensitive data. In both use cases, SLMs were noted to have a smaller environmental footprint than large models, though they still require energy and resources. Participants recognized that SLMs enhance efficiency in clinical documentation and decision support but introduce ethical challenges concerning patient welfare, fairness, autonomy, and privacy.