10
Loading video...
By Larry Klosowski
0
Share
Newsweek is a Trust Project member
See more of our trusted coverage when you search.Prefer Newsweek on Googleto see more of our trusted coverage when you search.
A hospital may have years of records that could help researchers study a rare condition, yet have good reasons to keep those records inside its own systems. A university may have the team to lead that research but lack the computing power to run it.
Both should have a way to help develop artificial intelligence on terms they can verify.
As public agencies invest in AI research, I believe they should fund projects that let institutions pool computing power and learn from one another’s data while retaining control over what they contribute.
Read More on Leadership Insights
Stanford University’s 2026 AI Index found that companies produced more than 90 percent of the notable advanced AI models it tracked for 2025. Companies have the resources to build some of the most demanding systems. Universities and hospitals bring a different strength.
They know which questions matter in a lab or on a hospital ward, and they can help test whether a model works outside the setting where it was developed. Access to the tools for building and testing models gives those institutions more say in which problems AI addresses.
...
Shared Research Infrastructure
There is a timely opening to widen that access. In September, the National Science Foundation announced an operations center for its National Artificial Intelligence Research Resource.
The center will coordinate access to computing power, data, models, and other research tools. That gives the agency a place to test how institutions can contribute resources as well as draw on them. A university might offer time on its research computers when capacity is available, allowing another team to use that time to test a model it has already built.
Choosing the right jobs matters.
A university should be able to put its own researchers first and limit the kinds of work it accepts from outside teams. A shared project should also account for the cost of electricity and staff time. If the institution carrying those costs has little access to the results, it has little reason to keep participating.
An agreement should spell out what each partner gets back, whether that means computing time, access to the resulting model, or a share of the research funding.
Protecting Data and Contributions
The same principle applies to data. Imagine several hospitals studying a rare complication. Each may see too few cases to build a model on its own. With a technique called federated learning, the hospitals could train the model on records inside their own systems.
The project would combine updates from each hospital into a shared model while the records stayed where they were. The hospitals could then test how well that model performs on their own patients’ data.
Keeping records in place addresses one risk. The National Institute of Standards and Technology has documented how an attacker can sometimes recover details about the training data from the updates themselves or from a trained model.
One safeguard NIST describes is combining updates in a way that hides each institution’s contribution from the party that combines them. Projects should test for leaks during training and again before anyone outside the team can use the model.
Each participant should know who can pause the work if a test reveals a problem.
Making Collaboration Accountable
Clear records would also make disputes easier to resolve. If a team uses more computing time than agreed, both partners should be able to check the same record. If a hospital withdraws, the agreement should say when its systems will stop receiving new work and whether the project may keep using models already trained.
The National Science Foundation should use its new center to run a small round of shared AI projects with these safeguards written into the funding terms. Each project should document who authorized the work, what resources it used, and how partners can verify that the work stayed within the agreement.
Before expanding the program, the agency should publish how much the projects cost, which projects gained access to computing power or data they needed, and how the safeguards held up. That gives other institutions evidence they can use to decide whether joining would serve their own research or public responsibilities.
The institutions closest to the problems AI could help solve should have a hand in building it and a say in how their contributions are used.
Larry Klosowski is the founder of Citrate Network, an AI infrastructure initiative focused on federated learning, verifiable computing, and AI governance. The views expressed in this article are the author's own.
View 0 Comments
Submit Correction
View Editorial & AI Guidelines