Open science once meant freely sharing research data to accelerate discovery, but generative AI has upended that ideal. Researchers who spent years advocating for transparency and accessibility are now facing an uncomfortable reality: their openly shared datasets are being harvested at scale by companies building proprietary AI systems, often without permission or attribution. This tension is forcing the scientific community to fundamentally rethink what "openness" actually means in the age of large language models (LLMs), which are AI systems trained on vast amounts of text data to generate human-like responses.

The problem is both technical and ethical. When research data is made openly available, it becomes a target for automated scraping, a process where AI developers systematically download and use that data to train their models. This extraction happens at a scale and speed that traditional open science frameworks never anticipated. Researchers describe feeling caught between two competing values: the collaborative spirit that made them champions of open data, and the need to protect their work from being repurposed in ways they never consented to or benefit from.

What Are Researchers Actually Worried About?

In-depth interviews with research data practitioners reveal a landscape of new concerns that didn't exist just a few years ago. These concerns go beyond simple data reuse; they touch on fundamental questions of control, attribution, and the purpose of scientific collaboration.

• Uncontrolled Scraping: Researchers describe helplessness as their carefully curated datasets are automatically downloaded and incorporated into AI training pipelines without their knowledge or consent.
• Re-identification Risks: When datasets are combined with other sources or processed through AI systems, individuals in the data may become identifiable again, even if the original data was anonymized.
• FAIR Principles Weaponized: The FAIR framework (Findable, Accessible, Interoperable, Reusable) was designed to make data easier for researchers to work with, but it's increasingly being reframed as a checklist for making data "AI-ready" for commercial extraction.

Rather than abandoning openness entirely, researchers are exploring what they call "recalibration." This means finding middle ground between complete secrecy and unrestricted access. Some are experimenting with selective sharing, where they make data available to trusted academic partners but restrict access to commercial entities. Others are advocating for alternative infrastructures and commons-based governance models that give communities more control over how their data is used.

How Are Researchers Adapting Their Data-Sharing Practices?

The shift toward responsible data stewardship is already underway, with researchers developing practical strategies to align openness with evolving ethical concerns. These approaches reflect what scholars call "reflexive agency," the capacity to make thoughtful decisions about data sharing that account for both the benefits of collaboration and the risks of extraction.

• Selective Sharing Agreements: Researchers are implementing tiered access systems where different versions of datasets are available to different audiences, with clearer terms about permitted uses.
• Alternative Infrastructures: Some are moving away from public repositories toward community-controlled platforms that allow researchers to set usage restrictions and monitor how their data is being accessed.
• Advocacy for Commons Governance: There's growing momentum for governance models that treat research data as a shared resource requiring collective stewardship, similar to how communities manage natural resources.

The underlying principle is that openness doesn't have to mean unrestricted access. Instead, it can mean transparency about how data will be used, accountability from those who access it, and genuine collaboration between data creators and data users. This reframing treats data sharing as a relational practice embedded in specific institutional and community contexts, rather than a one-way transfer of information.

Why This Matters Beyond Academia

The tension between open science and AI extraction has implications far beyond university research labs. As AI systems become increasingly central to decision-making in healthcare, finance, criminal justice, and other high-stakes domains, the question of where training data comes from and who benefits from it becomes a matter of public interest. If researchers lose trust in open data systems, they may become less willing to share findings, potentially slowing scientific progress. Conversely, if AI companies continue extracting data without consent or compensation, the social license for open science may erode.

The challenge is to develop governance frameworks that preserve the collaborative spirit of open science while protecting researchers from exploitative extraction. This requires alignment across multiple levels: the technical infrastructure that hosts data, the institutional policies that govern access, and the normative commitments that define what responsible data stewardship looks like in practice. Some institutions are already moving in this direction, but there's no consensus yet on what the new standard should be.

What's clear is that the era of "open means anyone can use it however they want" is ending. In its place, a more nuanced understanding of openness is emerging, one that balances transparency with accountability, accessibility with protection, and innovation with care for the communities whose data makes that innovation possible.