AI-driven drug discovery accelerates but risks reinforcing extractive pharmaceutical paradigms; systemic inequities in access and ecological costs remain unaddressed

Indigenous medicinal frameworks treat drug discovery as part of a broader ecological and spiritual system, where molecules are not 'designed' but co-created with nature. Systems like Ayurveda or Traditional Chinese Medicine (TCM) use pattern recognition and holistic diagnostics, which could inform AI models to prioritize multi-target therapies over single-molecule fixes. However, these traditions are systematically excluded from AI training datasets, which are dominated by Western biomedical literature and patented compounds, reinforcing epistemic injustice.

The history of drug development is rife with examples where speed and profit trumped safety, such as the thalidomide disaster or the opioid epidemic, both of which were enabled by regulatory failures and corporate malfeasance. AI diffusion models risk repeating these patterns by accelerating the production of molecules that may have unforeseen side effects or limited real-world utility. The current narrative echoes past techno-optimistic waves, like the genomics hype of the 1990s, which promised cures but delivered few for the masses.

Cross-culturally, drug discovery has thrived in decentralized, community-based systems where knowledge is shared and adapted locally, such as in the case of TCM or African traditional medicine. These systems often use iterative, low-tech methods to refine remedies, contrasting with the high-energy, high-cost AI approach. However, the global pharmaceutical industry has historically co-opted and patented these traditions, as seen with the neem tree or turmeric, turning communal knowledge into private property.

Diffusion models in AI generate molecules by iteratively refining structures, but their outputs are constrained by training data, which is skewed toward patented compounds and Western biomedical targets. The scientific literature rarely interrogates the ecological costs of training such models, including carbon emissions from data centers or the energy demands of molecular simulations. Additionally, the validation of AI-generated molecules often relies on in vitro assays, which may not predict real-world efficacy or toxicity in diverse human populations.

Artistic and spiritual traditions often frame healing as a creative act of balance, where molecules are seen as part of a larger cosmic or ecological harmony. For instance, the Japanese concept of *wa* (harmony) or the African principle of *ubuntu* (interconnectedness) could inspire AI models to prioritize therapies that restore systemic balance rather than target isolated proteins. However, these perspectives are rarely integrated into computational drug design, which treats the body as a machine to be 'fixed.'

Future scenarios must account for the unintended consequences of AI-driven drug discovery, such as the exacerbation of global health inequities if access to AI-designed drugs remains restricted by patents and pricing. Modeling should also explore alternative innovation models, like open-source drug development or AI tools trained on diverse, non-Western datasets. Additionally, the energy footprint of these models demands urgent attention, as computational drug design could become a major contributor to climate change if scaled unchecked.

Marginalized communities—particularly in the Global South—are often treated as test subjects for clinical trials rather than partners in drug development, a dynamic that AI diffusion models could exacerbate by further centralizing innovation. Patients in low-income settings, who bear the brunt of neglected diseases, are excluded from the narrative, as are traditional healers whose knowledge systems are sidelined. The lack of diversity in AI training data also risks producing molecules that are ineffective or harmful to non-Western populations, reflecting broader patterns of biomedical racism.