OpenAI Is Funding the Creation of Biological Data to Solve Medicine's AI Data Hunger Problem

When Medical AI Hits a Data Wall

Large language models are surging ahead in coding, writing, and general Q&A — but the moment they step into highly specialized domains like biomedicine, performance often falls short. The reason isn't complicated: high-quality, structured biological data is extremely scarce, and that data is the critical fuel needed to train reliable medical AI.

Leading labs like OpenAI have recognized that public web text alone can't feed medical AI. The most valuable information in biology — detailed drug development processes, clinical safety data, manufacturing strategies — is typically held tightly as core trade secrets, never disclosed publicly. This creates a paradox: the knowledge most in need of being learned is precisely the hardest to access.

A Bold Idea: Mining Data from Bankrupt Companies

Clinical trial policy analyst Ruxandra Teslo has proposed an imaginative solution: acquiring data from failed biotech companies. Her approach involves participating in the liquidation proceedings of bankrupt firms to obtain detailed regulatory filings, manufacturing strategies, and safety data — the kinds of documents normally treated as precious trade secrets.

The elegance of this idea lies in the biotech industry's notoriously high failure rate. Vast numbers of companies shut down after burning through their funding, but the experimental data, trial records, and hard-won lessons accumulated during their R&D remain. If these "sunken assets" could be recovered and structured into a format suitable for model training, they would dramatically enrich the knowledge base of medical AI. Data from failed cases is especially valuable — it can show models which paths don't work, and this type of negative information is severely underrepresented in public literature.

Why Failure Data Is Equally Precious

In drug development, the vast majority of candidate molecules ultimately fail. But published papers and successfully approved drugs only tell the "survivors'" stories, leaving AI models with little understanding of failure patterns. The complete archives left behind by bankrupt companies fill exactly that gap, allowing models to understand the true complexity and risk of real-world drug development.

This phenomenon is known in statistics as "Survivorship Bias." In drug development, bringing a new drug from candidate molecule to market typically takes 10–15 years, costs over $1 billion, and the overall clinical trial success rate is below 10%. Vast amounts of failed trial data are permanently lost due to competitive confidentiality, publication bias (journals favor positive results), or company closures. This means AI models trained on available data have only "seen" success stories, leaving them severely under-equipped to recognize failure patterns — and prone to systematic errors when predicting toxicity, side effects, or compound druggability. Complete archives from bankrupt companies — including records of trials terminated due to safety signals and regulatory filings that failed review — are precisely the key material needed to fill this cognitive blind spot.

OpenAI Chooses to Actively Fund Data Creation

Faced with data scarcity, OpenAI's answer is to directly pay to create new data, rather than relying solely on collecting existing data. This shift means AI companies are no longer passive data consumers — they are actively investing in the production of high-quality data.

This approach is especially important in biomedicine. Unlike web text, which can be scraped at scale, reliable biological data requires real experiments, rigorous documentation, and expert annotation — expensive and irreplaceable. The willingness to pay for it reflects a reassessment by leading labs of the value of domain-specific data. As general-purpose data approaches saturation, high-quality data in specialized fields is becoming the new competitive frontier.

"Paying to create data" is not an entirely new concept in the AI industry, but its application in biomedicine is particularly complex. Common approaches include commissioning specialized annotation (e.g., having pathologists label medical images), funding synthetic data generation (using computational simulations to produce data on molecular structure-activity relationships), and signing data-sharing agreements with hospitals and CROs (Contract Research Organizations). What makes OpenAI's move notable is its scale and intent — incorporating data production into core strategic budgets rather than treating it as a scattered partnership project. There's also a clear defensive business logic: proprietary datasets produced in-house cannot be replicated by competitors, creating a data-layer moat whose long-term value may even exceed that of model architecture advances.

What the Data Race Signals for the Industry

OpenAI's investment in biological data reflects a deeper trend in AI development: gaps in model architecture are narrowing, while data quality and uniqueness increasingly determine outcomes. Whoever can access proprietary data that others cannot may be able to build a nearly insurmountable advantage in specific domains.

From bidding on data from bankrupt companies to actively paying to generate new data, these explorations all point in the same direction — breaking into hard science domains like biomedicine no longer depends on larger models, but on better and scarcer data. This race around data may define the future landscape of medical AI more than the competition over model parameters ever could.

Conclusion

Progress in medical AI has long been constrained by the data bottleneck. Whether mining value from the archives of failed companies or having institutions like OpenAI directly fund data creation, both approaches are ultimately answering the same question: how do we help AI truly understand the complex reality of biology? The answer may lie not in smarter algorithms, but in how much we're willing to invest in high-quality data.