Every week, a lab somewhere feeds a protein sequence into AlphaFold Server and gets back a structure prediction that used to take months to solve by X-ray crystallography. Chemists run candidate molecules through Schrödinger’s software and generate leads that once required years of wet-lab screening. Universities, biotech and pharma companies now treat these and a myriad of other artificial intelligence tools as standard equipment. Researchers at companies routinely click through license agreements for the same tools without reading the fine print, overlooking restrictions that may limit or outright prohibit the use of program outputs for commercial purposes. In this article, we briefly review some of such restrictions and the issues that arise therefrom.

Many scientists working at academic, research, or nonprofit institutions assume that because they work at a university or nonprofit, they are covered by the "non-commercial" or "academic" license most platforms offer. Often they are, but even where that is the case, the provisions of these license agreements may interfere with the translation of the innovation in question from the benchtop into practice. Sponsored research or a university spinout can run afoul of a non-commercial license even when the person clicking "accept" is a tenured professor with no commercial title at all, and a corporate R&D team can just as easily violate an academic-only license simply because no one flagged the restriction before the results made it into a product.

Take AlphaFold Server for example, the most widely used of these tools, and a free, web-based research tool created, hosted, and operated directly by Google DeepMind. Google’s Additional Terms limit the use of AlphaFold Server to non-commercial use by individuals and non-commercial institutions such as universities, non-profits, research institutes, educational bodies and government bodies, and for journalism. While it may seem that academic labs fit into this list, the same terms bar use “in connection with commercial activities,” and they specifically name research done on behalf of a commercial organization. Picture a university lab running a sponsored-research agreement with a biotech partner. The professor uploads a target sequence to AlphaFold Server to help design a binder. The lab is academic, but the funding is not. If the sponsor gets rights to what comes out, the project may no longer qualify as non-commercial use, regardless of the professor’s employer.

AlphaFold’s Output Terms go further, prohibiting any use of outputs or derivatives that give a commercial organization rights in those outputs or derivatives. Thus, publishing research results in a journal or other public media would be fine, but expecting or representing that a company owns rights to the resulting structure prediction would not. Notably, this impacts the ability of these non-commercial institutions to grant licenses to their innovations. 

Schrödinger’s Non-Commercial License presents a similar scenario. Its Academic Restrictions let Qualified Non-Commercial Users run the software for academic or research purposes, but bar any use “directly or indirectly” for a project that supports, or is supported by, commercial efforts or a commercial enterprise. Schrödinger carves out basic research funded by a commercial enterprise with no commercial or proprietary interest in the outcome — a narrow lane that does not apply to much sponsored research.

Other AI platforms regulate use, not based on academic or commercial status, but rather on whether the subscriber is using the platform for its own internal drug discovery purposes or for providing services for a third party’s benefit, including providing services or reselling results to a third party. That creates exposure for contract research organizations, sponsored-research projects and potentially, university spinouts.

The common thread: owning your inputs and outputs does not erase use restrictions. Terms like “derivatives,” “output” and “commercial” are often left undefined, or defined loosely enough that reasonable people may disagree about such terms’ boundaries. Companies and institutions should evaluate and weigh the ownership risks, project by project, before they upload anything into AI models.

Confidentiality is the second area that should be considered, and the relevant license terms vary more than most people expect. Don’t assume that one platform treats your data the same way another does. The AlphaFold Server Terms of Service do not have a conventional confidentiality clause, despite committing not to share your content or output without approval. Instead, Google’s AlphaFold Server Privacy Notice permits Google to keep a time-stamped “User Record” of the sequences you input and the outputs AlphaFold generates for an extended period, to monitor compliance and investigate violations of Google’s terms. Deleting your Google account or your AlphaFold Server history does not delete that record.

Schrödinger treats uploaded data as confidential information and requires strict confidentiality from both sides. But its hosted-software terms let Schrödinger use, compile, store and process such data to run the software, and let it collect anonymized analytics for its own internal purposes. Other tools may permit the platform to derive and utilize aggregated, de-identified benchmarking data. This last carve-out deserves some scrutiny. A novel compound structure or an unpublished sequence can remain identifiable even inside an aggregated dataset, simply because so few labs work on it. For a high-value program, the safer approach would be to skip the upload of the sensitive data altogether, absent clearance from an attorney.

A third issue gets less attention than it deserves, but it too can impact commercialization: mandatory disclosure and fees. Under Schrödinger’s Non-Commercial License, results, inventions or discoveries generated using the software must be publicly disclosed—not kept confidential or proprietary—unless the licensee pays in advance to upgrade to a full commercial license at Schrödinger’s then-current pricing. The same section also imposes a punitive patent-filing fee if the licensee has not subscribed for a commercial license before filing a patent application that contains a claim based on output from the platform. Tech transfer offices should investigate what AI tools have been used in research and what restrictions they may impose before filing patent applications.

None of this means these AI tools are too risky to use. It means that whether you are doing research for a company or for an academic or research institution, you need to be aware of the terms of use that may apply to the tools you use. Before your team uploads a sequence, a compound or a dataset, map out which platform you’re using, pull every referenced agreement—not just the Terms of Use, but also the Privacy Notice and any other linked terms—and map the specific use against the license. Executives, general counsel and tech transfer offices who build these steps into processes and sensitize their organizations to these issues now will save themselves a much harder conversation later.

Summer Associate Halla Daoui co-authored this publication.