The Grammar of Future Use: 2015 Data-Sharing Terms Under 2026 Machine-Learning Review
Photo: Syced, CC0, via Wikimedia Commons
At the HumanRep 2015 Archive, the most noticeable change in incoming reference requests is not the subject matter but the verb. Earlier inquiries asked whether a 2015 cohort could be reanalyzed, whether a specimen series could be shared with a collaborator, or whether a data-use agreement permitted a follow-up study. Recent requests increasingly ask whether the same materials may be used to train, validate, or fine-tune a machine-learning model. The documents being consulted are unchanged. The interpretive frame is not.
This review examines the 2015 consortium’s data-use agreements and material-transfer agreements as archival objects. It does not attempt to settle their legal meaning. It asks a narrower question: what did the drafting language assume about future research, and how are US institutions in 2026 reading that language when de-identified cohort data and biobank samples are sought for AI model development?
The Documents as They Were Drafted
The 2015 templates were products of their moment. They distinguished between data transfers and material transfers, between identifiable and de-identified information, and between scholarly collaboration and commercial application. Future-use clauses typically authorized research within a broad domain—reproductive health, fertility, pregnancy, or related biomedical research—and often added phrases such as including but not limited to. Some templates permitted sharing with named collaborators; others required amendment for new institutions. Commercial use was frequently treated as a separate question, requiring notice, negotiation, or an explicit written modification.
What the archive does not contain is equally important. The agreements do not define model training. They do not mention algorithmic derivatives, model weights, embeddings, or synthetic data. They do not specify whether a trained model is a research output, a data derivative, or a new material subject to the original transfer terms. In 2015, machine learning was not yet a routine part of reproductive-health research administration, and the documents reflect that absence without necessarily revealing it.
De-identification and the Model-Training Question
US institutions reviewing such requests in 2026 begin with familiar categories. Under the Common Rule, de-identified data are generally outside the regulatory definition of human-subjects research, and HIPAA provides its own de-identification standards. But contractual terms can be stricter than federal rules. A data-use agreement may permit de-identified sharing while still limiting the purposes for which that sharing occurs. The central question is therefore not only whether the data remain identifiable, but whether model training falls inside the permission that was granted.
Archive records show that 2015 drafters thought in terms of recipients, projects, and research domains. They did not think in terms of model artifacts that could circulate after the original data had been minimized or withheld. A 2026 reviewer may ask whether training constitutes research use under the agreement, whether the resulting model is a derivative work, and whether the institution retains control over later commercial deployment. The 2015 text rarely answers these questions directly. It supplies general categories, and later reviewers supply the boundaries.
Consent at the Perimeter
The consent forms in the archive present a related problem. Many were written to support broad future research, including genetic and reproductive studies, while preserving institutional review board oversight and withdrawal procedures. Some plainly contemplated biobank governance and long-term specimen use. Few, if any, anticipated that a participant’s de-identified data might contribute to a model whose parameters could be reused across institutions or commercial settings.
That does not make the original consent meaningless. It does mean that 2026 institutions must decide whether AI model training is a species of future research already contemplated or a materially different use requiring fresh review. The archive offers evidence of intent, but not a controlling answer. Consent language was designed for a governance system built around specimens, datasets, and projects. Algorithmic reuse introduces a unit of analysis that the 2015 paperwork did not name.
How 2026 Reviewers Are Reading the Terms
Three broad interpretive postures appear in current US review practice. A literal posture treats de-identified data as within the agreement’s research permission unless a specific exclusion applies. A conservative posture treats model training as a new use because the agreement did not expressly authorize it, especially where commercial partners or foundation models are involved. A case-by-case posture examines the recipient, the model’s purpose, the sensitivity of the source cohort, and the likelihood of downstream re-identification or mission drift.
None of these postures is invented from nothing. Each draws on the same 2015 language but assigns weight differently to its silences. The archive cannot resolve the conflict because the archive was not designed to do so. Its value is diagnostic: it shows that future-use clauses functioned less as predictions than as grants of general permission, with later institutions completing the grammar.
What Foresight Could Not Supply
The 2015 consortium understood that data and specimens would travel. It built review mechanisms, collaborator obligations, and amendment procedures around that expectation. What it could not supply was foresight about scale, model opacity, and the commercial ecology of 2026 AI research. The old clauses remain usable, but they require translation. They are not relics of a simpler time so much as records of a governance approach that preferred general categories over enumerated technologies.
As of November 2026, the most honest reading of the archive is that the 2015 agreements did not see the algorithm. They did see sharing, and they left a structure for controlling it. Whether that structure bends toward permission or restriction now depends on institutional interpretation, not archival recovery.
