Systemic analysis: How AI diffusion models in drug design reproduce extractive pharmaceutical paradigms and marginalise alternative innovation pathways
Original framing: “Q&A: What AI actually does in diffusion models for drug design” — Phys.org
Diffusion models in AI excel at generating novel molecular structures by learning from large chemical databases, but their 'success' is measured by metrics like docking scores and synthetic accessibility—not therapeutic accessibility or ecological sustainability. The models’ reliance on proprietary datasets (e.g., ChEMBL, PubChem) entrenches biases, as these databases underrepresent molecules from non-Western medicinal systems. Moreover, the energy-intensive nature of training diffusion models (e.g., 100s of GPU hours per model) raises ethical questions about their carbon footprint in the context of climate justice.
The AI diffusion models in drug design exemplify how technological 'innovation' can entrench historical inequities under the guise of progress.
By centring Western epistemologies, proprietary datasets, and profit-driven R&D, these systems reproduce the colonial logics that have long shaped global health—where diseases of the poor are ignored and traditional knowledge is commodified. Yet the solution lies not in rejecting AI but in reorienting it: through decolonial data governance, frugal innovation, and regulatory frameworks that prioritise public health over corporate monopolies. Indigenous and marginalised voices must lead this transformation, not as token participants but as architects of a new paradigm where medicine is co-created with communities, not extracted from them. The trickster’s role is to expose the absurdity of a system that claims to 'discover' cures while ignoring the wisdom that has sustained health for millennia—Hermes’ caduceus, not as a symbol of corporate power, but as a reminder that healing is a relational, not a transactional, act.