Show summary
With rapid advances in AI technology, there is an increasing demand for data to train AI models. These models are used in medicine and health care for clinical research, diagnosis, drug development, and prevention. However, while significant amounts of health-related data are generated, large datasets may not be available due to privacy concerns and regulatory frameworks (e.g., GDPR). To obtain patients’ data, informed consent is required, therefore creating a high threshold for data sharing. In addition, for rare diseases or cases of “no diagnosis”, large data sets are not available.
Synthetic medical data (SMD) has been introduced to address data-acquisition challenges; It has been described as a solution that enables model training without compromising patients’ privacy and is therefore considered "ethically desirable" (Bellovin et al. 2019, Savage, 2023).
In this talk, we will introduce three main models of SMD: Generative models, Differential Privacy (DP) models, and the C-vine copula-based synthetic data generator. As we will show, these models are not value-neutral, and they pose both epistemic and ethical trade-offs. We will illustrate how SMD use creates new value hierarchies: Privacy over transparency: SMD can obscure the specific characteristics and nuances of real patient populations, which can create explanation gaps that undermine clinicians’ ability to interpret model outputs, erode trust, hinder bias mitigation (Susser et al., 2024), and knowledge development, and risk long-term de-skilling. These risks are especially relevant in the case of rare diseases, where small original datasets and oversimplified synthetic approaches can lose critical dependencies and idiosyncratic features, leading to misleading clinical insights. Scalability over fidelity: developers of SMD aim for low individual fidelity (to protect privacy) but high distributional fidelity, yet, as we will show, noise and selective-distortion
methods can break crucial dependencies or entrench biased dependencies from the original data (ibid.), perpetuating the “garbage in, garbage out” paradox (in the process of keeping what seems as important dependencies in C vein copula models), or even introduce new biases (when adding noise in DP models). The preference for scalable, quantified data over individual experience holds risks of alienating patients, erasing contextual and minority experiences (creating what we call “zombie data”), expanding diagnostic disparities, and institutionalizing these trade-offs as a new normal.
We will discuss how this prioritization fosters fungibility in medical practice, which we conceptualize as an advanced stage of the medical reification process. We will distinguish
between the stage of “datafication”, which is characterize by decontextualization and translation of messy human experiences into transportable data points that creates a partial, curated technical form of objectivity (Leonelli,2016; Nguyen,2024), and SMD, which goes a step further, generating “zombie data” that deepens reification by stripping lived experience, context and agency from those experiences. Finally, we will argue that this new level of reification is a step forward in the process of commodification and, hence, in capitalism.
References
Bellovin, S. M., Dutta, P. K., & Reitinger, N. (2019). Privacy and synthetic datasets. Stan.
Tech. L. Rev., 22, 1.
Leonelli, S. (2019). Data-centric biology: A philosophical study. University of Chicago
Press.
Nguyen, C. T. (2024). Value capture. J. Ethics & Soc. Phil., 27, 469.
Savage, N. (2023). Synthetic data could be better than real data.
Susser, D., Schiff, D. S., Gerke, S., Cabrera, L. Y., Cohen, I. G., Doerr, M., ... & Wagner, J.
K. (2024). Synthetic health data: real ethical promise and peril. Hastings Center Report,
54(5), 8-13.
