Ethical considerations for multimodal artificial intelligence in healthcare

Area of Science:

• Biomedical informatics
• Artificial intelligence in healthcare
• Medical ethics

Background:

• Multimodal artificial intelligence (MMAI) integrates diverse data types (images, speech, text, etc.) into unified spaces.
• This enables advanced cross-modal inference and data synthesis for improved diagnostics and patient support.
• Current AI governance frameworks are insufficient for MMAI's unique ethical challenges.

Purpose of the Study:

• To characterize the ethical novelty of MMAI-generated data.
• To address concerns regarding the integration of inference-based data into clinical and research records.
• To propose a new governance framework for MMAI in biomedicine.

Main Methods:

• Analysis of MMAI's capability to infer sensitive information and generate new data objects.
• Examination of how MMAI blurs the line between observation and generation.
• Development of a four-part governance agenda for MMAI.

Main Results:

• MMAI ethically differs by rendering cross-modal inferences as recordable data, challenging data provenance.
• Inference-based data objects can be integrated into medical records without clear origin or consent.
• Existing governance models fail to adequately address the infrastructural embedding of such data.

Conclusions:

• A shift from data-centric protection to governance of inference and infrastructuring is required.
• Proposed agenda includes provenance labeling, evidence-building for inference capacities, dynamic consent, and privacy-preserving techniques.
• These measures aim to balance innovation with safeguarding individual rights and privacy in MMAI applications.