Eoghan Whitfield is the Technology Transfer Director at a federal research laboratory inside a civilian cabinet department: a staff of 18, an annual R&D budget of $340 million, and a portfolio of 47 AI-related inventions at various stages of patenting, licensing, and collaborative development with private-sector partners. Three years of internal research had produced a predictive maintenance algorithm for critical infrastructure monitoring that outperformed every commercially available system the agency had evaluated. Several private companies wanted it. So did two other federal agencies. A state public utility commission asked whether it could be adapted for grid resilience monitoring. Nobody at the laboratory was quite sure what the legal path was for making any of that happen.

The statutory frame was not missing. The Stevenson-Wydler Technology Innovation Act of 1980 and its subsequent amendments define the framework for federal laboratory technology transfer. The Bayh-Dole Act governs intellectual property rights in federally funded research more broadly. The laboratory's Cooperative Research and Development Agreement authority allows collaborative development with private partners. But when Eoghan tried to design a transfer arrangement for a model rather than for a machine, he found gaps the drafters of those statutes had no reason to anticipate. He spent seven months working out what a principled framework for AI technology transfer looks like. This lesson is what he found.

Why AI Transfer Is Legally Distinct

Technology transfer is the process by which innovations developed with government resources move to private-sector or other public-sector users. For traditional inventions it has a well developed legal apparatus: patents, licenses, cooperative development agreements. That apparatus was designed around discrete artifacts, a new material, a manufacturing process, a piece of equipment, each of which can be described completely in a document and handed over. AI fits it imperfectly, and the mismatch is not cosmetic. Three characteristics of a trained model break assumptions that the traditional instruments quietly rely on.

First, a model is not a finished product at the point of transfer. It is a function of its training data and its operating environment. A predictive maintenance algorithm trained on federal laboratory infrastructure data will behave differently on a utility company's grid, which has different sensor configurations, different failure modes, and different data quality. Transferring the model is not like transferring a blueprint. It is transferring a starting point that the recipient has to adapt, revalidate, and monitor in a setting the government never tested. The agreement should say who is responsible when that adaptation goes badly.

That property has a commercial consequence worth noticing before terms are drafted. A patent license ties payment to practising a claimed method, so if the model's value sits mostly in weights learned from government data rather than in a novel technique, a royalty basis anchored to the patent may collect very little while the recipient captures most of the benefit. The mismatch is not a reason to abandon patenting. It is a reason to be honest, early, about where the value actually lives, because that answer determines which instrument you are drafting and what the government is being compensated for.

Second, government training data is often inseparable from the model's capability. The laboratory's algorithm is as good as it is because it learned from three years of federal infrastructure monitoring that no private company could have collected independently. So the questions multiply. Does the training data travel with the model in some form? Do restrictions attached to the data under the authority that collected it also constrain uses of the model? If the recipient fine-tunes on their own data, how much of the resulting capability derives from the government data and how much from theirs? Silence on these points is not neutrality. It is a dispute deferred.

Third, models can be reproduced at near-zero marginal cost. A licensed physical technology has natural limits on unauthorized replication, because you cannot easily copy specialized equipment in a garage. A licensed model can be copied indefinitely by any recipient with adequate technical capability. License terms that are routine for hardware, restricting use to one facility, one application, or a fixed number of units, are much harder to enforce over a software asset. The clause still matters, because it defines the breach, but it should not be mistaken for a technical control.

The Three Transfer Mechanisms

Federal technology transfer law gives Eoghan's laboratory three principal mechanisms, each with different intellectual property implications and different suitability for AI. Choosing among them is the first substantive decision in any transfer, and it is easier if you start from what the government is actually trying to move: a patentable method, a continuing research relationship, or a finished artifact.

A patent license grants the right to use a patented invention in exchange for royalties. For AI, patents may cover specific algorithmic innovations, but they do not protect the trained model artifact itself and say nothing about training data embedded in model parameters. That makes a patent license appropriate when the thing of value is a method someone else could independently implement, and inappropriate when the value sits in the weights rather than in the technique.

A Cooperative Research and Development Agreement, or CRADA, is a formal partnership under which a federal laboratory and a private-sector partner jointly conduct research and development, sharing resources and intellectual property. It is the most flexible mechanism for AI, because it accommodates the thing an AI model actually does over time: improve. The laboratory can keep developing the model, the partner can contribute operational data from their environment, and the improved model can be owned jointly or licensed back to the government on terms agreed in advance rather than argued about later. CRADAs also carry approval requirements above certain cost thresholds and disclosure obligations, so confirm which apply to your laboratory before you promise anyone a timeline.

A software license agreement, sometimes called a software use license under government intellectual property frameworks, grants the right to use a specific version of the government-developed software, meaning the trained model and its inference code, for a specified purpose, in a specified domain, for a specified period. It suits straightforward transfer where the government wants a finished model in the hands of one or more recipients with no continuing joint research. Its terms have to address the training data question explicitly rather than by implication.

The Three Ownership Layers

Intellectual property management for AI developed with federal funding has to address three layers, and confusing them is the most common drafting error. The algorithm or method may be patentable. The trained model artifact is protected as software under copyright. The training data may carry statutory restrictions from the programs under which it was collected. A license that speaks to only one layer leaves the other two ungoverned, which is how a laboratory ends up discovering after deployment that it licensed a method and inadvertently released a dataset.

Working the layers in order also settles arguments that otherwise circle. Start with the method, because whether anything is patentable determines if a patent license is even on the table. Move to the artifact, because copyright in the trained model and its inference code is what a software license actually conveys, and because the artifact is the thing a recipient will run in production. Finish with the data, because that layer is the one whose restrictions originate outside the technology transfer office entirely and cannot be negotiated away by either party at the table.

Under the Stevenson-Wydler framework, federal laboratories retain ownership of technologies developed primarily with federal funding and license them to private parties. For AI, the license should specify the permitted uses of the model; whether the recipient may fine-tune it on their own data and whether the fine-tuned model remains subject to the original terms; and what the recipient must disclose to their own customers about the model's government provenance. The general principle is that commercialization is encouraged, because the purpose of technology transfer is to generate benefit from public research investment.

The limitation on that principle is that the government does not walk away. It retains march-in rights, the authority to require additional licenses if public health, safety, or security needs require it, and the right to use the technology for government purposes without royalties. Neither of those is self-executing. March-in is a reserved authority that has to be invoked through a process by people who are paying attention, and a royalty-free government use right is only useful if someone in the agency knows it exists. Preserving them in the agreement is what makes them available at all, which is a lower claim than protection and a more honest one.

Technology transfer for AI is not about giving away what the government built. It is about ensuring that public investment in government research produces the broadest defensible benefit: in the commercial market, in other government programs, and in the communities that funded the research in the first place.

What the License Must Say About Data

The training data clause is where AI transfer agreements are won or lost, and it is usually the shortest paragraph in the draft. Three questions belong in writing. What data is embedded in the model? What restrictions from that data's original collection authority apply to uses of the model? And what must the recipient disclose about the model's government origin when they deploy it? None of those can be answered by the technology transfer office alone; they require the program office that collected the data and the counsel who advised on that collection.

The order of those conversations matters as much as their content. A technology transfer office that drafts first and consults second will produce a document the program office cannot approve, and the rework lands after a partner has already been told the terms. Consulting first produces a shorter draft with fewer aspirational clauses, because the people who know what the data can support have removed the options that were never available. It also builds the record you will want later, when somebody asks on what basis the laboratory concluded that this model could go to this recipient at all.

Fine-tuning deserves its own sentence rather than an inference. If the recipient trains further on their own data, the resulting model is partly theirs and partly derived from the government artifact, and the parties will read an ambiguous clause in opposite directions once the fine-tuned model is commercially valuable. Decide in advance whether fine-tuned descendants remain subject to the original terms, and say so. Provenance disclosure matters most when the recipient's customers are themselves government agencies, whose procurement processes may be affected by knowing that the product they are evaluating rests on a model their sister agency built.

The Recipients Who Are Not Companies

Eoghan's queue was not made up only of companies. Two other federal agencies wanted the model, and a state public utility commission wanted to know whether it could be adapted for grid resilience monitoring. Those requests are easy to treat as administrative afterthoughts, because no royalty is at stake and nobody is negotiating hard. That is exactly why they go wrong. A transfer to another public body raises the same three layers as a commercial one: the method, the trained artifact, and the training data with whatever restrictions its collecting authority attached. The absence of money does not remove any of them.

It also raises the adaptation problem in a sharper form. A sister agency receiving the model is likely to assume that because both parties are government, the data questions are already settled and the validation already done. Neither is safe. The receiving agency's infrastructure, sensors, and failure modes differ from the laboratory's, so performance has to be re-established in the new setting, and the receiving agency's authority to hold the underlying data may differ from the laboratory's. Write the same terms you would write for a company, then adjust the fee and the formality rather than the substance.

The practical benefit of that discipline shows up in the commercial negotiation. If a laboratory has already decided how it transfers to other public bodies, it can offer a commercial partner exclusivity without conceding the public paths, because those paths are documented practice rather than a position invented under pressure. Deciding the public case first also tends to sharpen the drafting, since nobody is distracted by the royalty schedule while working out what the data clause actually means.

Equity and Access in Transfer Terms

When a government laboratory develops an AI system with significant public benefit, the transfer framework should include provisions ensuring broad access rather than commercial exclusivity for whoever arrives first. Exclusive licensing periods should be limited in duration, should carve out government use and non-commercial research, and should include public interest provisions allowing the government to require non-exclusive licensing if the public benefit of the technology is not being adequately realized. Those provisions are cheap to include at signature and effectively impossible to add afterwards.

Eoghan's laboratory negotiated a five-year exclusive commercial license for one of its models, with a parallel non-exclusive license to all state and local governments and public utilities for a nominal administrative fee. The commercial partner objected, which was predictable and reasonable from their side. The laboratory's position was that the model was built with public funds for a public purpose, and that commercial licensing supplements that purpose rather than replacing it. Writing the parallel license into the same negotiation, rather than promising to consider it later, is what made the position hold.

The reason to fight that argument at signature is that public interest provisions are only meaningful while they are still cheap. Before signature, a duration limit and a non-exclusive path for public bodies are terms among terms. After signature, the same provisions require reopening a deal the partner has already built a product plan around, and the agency is asking for something rather than offering it. The asymmetry is entirely about timing, and a laboratory that understands it can be generous on the commercial terms it does not care about while holding the two or three that determine whether the public keeps access.

Anti-Patterns

• Treating a license clause as a technical control. A term restricting a model to one facility or one application defines a breach; it does not prevent a copy. Where replication would be catastrophic rather than merely undesirable, the question is whether to transfer the weights at all, or to transfer access to a hosted service instead, not how strongly to word the clause.
• Assuming march-in rights protect the public automatically. They are a reserved authority, not a safety net. If nobody in the agency is monitoring whether the licensed technology is actually reaching the public, the right sits unused, and unused authority protects no one.
• Licensing the method and ignoring the weights. A patent license on an algorithm leaves the trained artifact and the data inside it ungoverned. Name all three layers in every agreement.
• Leaving fine-tuning to interpretation. Ambiguity about whether a fine-tuned descendant remains under the original terms is only ambiguous until the descendant becomes valuable. Then it is a dispute with two sincere readings and no cheap resolution.
• Letting the first licensee define the public benefit. Exclusivity that carries no duration limit, no government use carve-out, and no non-exclusive path for other public bodies converts a public research investment into a private one and calls it commercialization.
• Promising a schedule before checking the approvals. Cooperative agreements carry approval thresholds and disclosure obligations that vary by laboratory and by cost. Find out which ones apply to yours before a partner builds a product plan around your estimate.

Practice Prompts

• Sort your portfolio by mechanism. Take the AI assets your organization holds and assign each to a patent license, a cooperative research agreement, or a software license, based on whether the value sits in a patentable method, in a continuing research relationship, or in a finished artifact. Note the ones that resist classification and say why.
• Trace the data back to its authority. For one model you might transfer, list every dataset used in training and the collection authority behind each. Then write, in plain sentences, which restrictions from those authorities you believe follow the model into a recipient's hands.
• Draft the fine-tuning clause. Write the sentence that says whether a recipient may fine-tune, and whether the fine-tuned model remains subject to your terms. Then have a colleague argue the opposite reading of your own draft.
• Design the parallel access path. For a model with public benefit beyond the commercial market, draft the non-exclusive license you would offer to other public bodies alongside a commercial exclusive, including the fee basis and the eligibility test.

Reflection

• Of the AI assets your organization has transferred or is preparing to transfer, how many have a written answer to the training data question, and how many rely on the assumption that it will not come up?
• Who in your organization would notice if a licensed technology stopped reaching the public it was meant to serve? If the answer is nobody, what is your reserved authority worth?
• Where has your organization confused a contractual restriction with a technical one, in this domain or another?
• If your best model were licensed exclusively tomorrow, which other public bodies would lose access, and would anyone have raised that during the negotiation?
• What would have to change for a transfer agreement to be drafted with the program office that collected the data in the room, rather than reviewed by them afterwards?

Glossary

• Technology transfer. The process by which innovations developed with government resources move to private-sector or other public-sector users, through patents, licenses, or cooperative development agreements.
• Patent license. An agreement granting the right to practise a patented invention in exchange for royalty payments. For AI it reaches the claimed method, not the trained artifact or the data embedded in it.
• Cooperative Research and Development Agreement. A formal partnership under which a federal laboratory and a non-federal partner jointly conduct research and development, sharing resources and intellectual property on terms agreed in advance.
• Software license agreement. An agreement granting use of a specific version of government-developed software, including a trained model and its inference code, for a specified purpose, domain, and period.
• March-in rights. The reserved government authority to require additional licenses where public health, safety, or security needs require it. An authority that must be invoked, not an automatic protection.
• Government purpose use. The retained right to use a transferred technology for the government's own purposes without paying royalties to the licensee.

Related Lessons

• Building Innovation Ecosystems covers the infrastructure that decides whether a laboratory has anything worth transferring.
• Government AI Venture Creation takes commercialization further, into public technology as the basis of a new venture.
• Academic and Research Partnerships works the university side, where the intellectual property split runs in the opposite direction.
• Data Sharing Agreements for AI handles the data layer directly, including restrictions that follow a dataset downstream.
• AI Contract Negotiation covers the drafting craft that turns a transfer decision into durable terms.

Closing

Eoghan's seven months did not produce a new statute or a clever workaround. They produced a habit: before any AI asset leaves the laboratory, name the three layers, trace the data to its authority, decide the fine-tuning question in writing, and preserve the public paths that exclusivity would otherwise close. None of that is exotic legal work. It is the ordinary discipline of asking what is actually being transferred, applied to an asset that does not behave like the inventions the framework was built around.

The frameworks will keep lagging the technology, and waiting for them to catch up is not a strategy. What a technology transfer office can do now is refuse to let the gaps become defaults. An unanswered question about training data becomes the recipient's answer. An unstated limit on exclusivity becomes no limit. A reserved authority nobody monitors becomes a paragraph in a file. The work is to make each of those a decision someone made on purpose.

Key Takeaways

• AI models are legally distinct from traditional patentable inventions. They are not static products, they embed training data, and they can be replicated at near-zero marginal cost. Frameworks designed for physical inventions have to be adapted to those characteristics explicitly.
• Training data restrictions can follow the model into transfer. Where a government model is trained on data collected under statutory restrictions, those restrictions may apply to uses of the trained model. License terms must state what data is embedded and which restrictions travel with it.
• Cooperative research agreements suit models that keep changing. When both the laboratory and the partner contribute research capacity and the model will continue to evolve, joint development with pre-agreed ownership terms fits better than a one-time license.
• Fine-tuning rights belong in the text. Whether the recipient may fine-tune on their own data, and whether the result stays under the original terms, must be decided before deployment rather than interpreted after it.
• March-in rights are reserved authority, not automatic protection. Preserve them expressly in AI licenses, and accept that they only function if someone is monitoring whether the technology is reaching the public.
• Exclusivity should carry a parallel public path. Commercial exclusivity for a private partner should not close access for other government entities, public utilities, and research institutions to a model built with public funds.
• Provenance disclosure is a commercialization term. Recipients building services on government models should disclose that origin to their customers, particularly where those customers are agencies whose procurement decisions the disclosure would affect.

Frequently Asked Questions

Which mechanism should we default to for an AI model?

There is no safe default, because the mechanisms answer different questions. If the value is a specific patentable method, a patent license fits. If the model will keep improving and the partner brings research capacity and operational data, a cooperative research and development agreement fits. If the artifact is finished and you simply want it used, a software license fits. Picking the instrument your office is most used to drafting, rather than the one that matches the asset, is how the wrong terms get inherited.

Can a license stop a recipient from copying the model?

It can make copying a breach. It cannot make copying difficult. Any recipient with adequate technical capability can reproduce a model at negligible cost, which is exactly the property that traditional licensed technology does not have. If unauthorized replication would be unacceptable rather than merely unwelcome, the decision to make is about what you transfer and how, not about how forcefully the clause is worded.

Does the government keep the right to use technology it licenses out?

Federal technology transfer law preserves the right to use the technology for government purposes without royalties, alongside march-in rights where public health, safety, or security needs require additional licensing. Both should be stated expressly in the agreement. Both depend on somebody in the agency knowing they exist and watching whether they are needed.

What happens to our data restrictions when the model leaves?

That is the question to answer before signing rather than after. Where a model was trained on data collected under a specific authority, restrictions from that authority may constrain uses of the model itself. Bring the program office that collected the data and the counsel who advised on the collection into the drafting, and write their answer into the license instead of leaving it to be litigated.

How do we protect public access when a company wants exclusivity?

Limit the exclusive period, carve out government use and non-commercial research, and negotiate the parallel non-exclusive license for other public bodies in the same conversation rather than promising to revisit it later. Eoghan's laboratory paired a five-year commercial exclusive with a non-exclusive license to state and local governments and public utilities at a nominal administrative fee.

Does a transfer to another agency need the same paperwork?

It needs the same substance. The three layers do not disappear because no money changes hands, and a receiving agency that assumes the data questions are settled will inherit restrictions nobody wrote down. Adjust the fee and the formality for a public recipient if that is appropriate, but keep the permitted uses, the data provenance, and the fine-tuning position exactly as explicit as they would be for a commercial licensee.