Secretary of Defense Pete Hegseth gives a State of the Force address at Marine Corps Base Quantico, Wednesday, Sept. 30, 2026 in Quantico, Va. (Alex Wong/Pool Photo via AP)

The Pentagon Has a Big Tech Problem

An entrance driveway to a hospital with a sign pointing to Emergency, Laboratory, and Radiology. A person is walking past from the left, and another person is leaving the hospital on a motorbike. 

A street view of The Nairobi Hospital in Kenya on Oct. 6, 2026 (AP Photo/Andrew Kasuku)

Without an Exit Plan, AI Partnerships in Africa Risk Becoming the New Tied Aid

Voters line up at a polling station during the presidential runoff in Brasilia, Brazil, Sunday, Oct. 25, 2026. (AP Photo/Eraldo Peres)

How to Reclaim the Future from Strongmen and Oligarchs

The Edge Dialogues with Dr. Alondra Nelson

Demonstrators holding a banner that reads in Portuguese "Brazil decides about Brazil" rally against Meta over allegations of interference in Brazil's presidential election, outside Meta's headquarters in Sao Paulo, Monday, Oct. 19, 2026. (AP Photo/Andre Penner)

On Eve of Vote, Brazil’s Federal Prosecution Office Demands Answers from Meta

Demonstrators wear masks depicting Meta CEO Mark Zuckerberg, Brazilian presidential candidate Flavio Bolsonaro and US President Donald Trump during a protest outside Meta's headquarters in Sao Paulo, Brazil, on October 19, 2026. (Photo by Allison Sales/NurPhoto via AP)

In Brazil’s Runoff, Lula Faces Bolsonaro, Trump—and Big Tech

A grey toned abstract photo of wavy mirror reflections of a building seen from many different angles. 

LLMs Are Still Making Election Deepfakes Just Days Before the US Midterms

A view of the US Capitol building with the dome brightly illuminated against the backdrop of a night sky. 

The United States Capitol dome by night, Washington DC, 2019. Photo by Darren Halstead on 

A Congressional Commission on AI Risk is Overdue

A demonstrator wearing a yellow vest waves the French flag on top of a burning barricade on the Champs-Elysees avenue with the Arc de Triomphe in background in Paris. (AP Photo/Michel Euler, File)

France’s Protests Reveal Europe’s Expanding Surveillance of Dissent

Building of the European Parliament in Strasbourg, France, with the EU-flag.

Building of the European Parliament in Strasbourg, France

European Parliament Divided on How to Govern AI At Home and Abroad

People with their mouths taped participate in a protest against internet shutdowns, arrests, and executions in Iran outside the Palacio Nacional de Belem in Lisbon, Portugal, on May 10, 2026. (Photo by Luis Boza/NurPhoto via AP)

Age Checks, AI Regulation and Censorship Put Internet Freedom Under Pressure

A rally against data centers at R.K.Beach in Visakhapatnam, India, on July 19. Source: Stop Dirty Data Centers

A rally against data centers at R.K.Beach in Visakhapatnam, India, on July 19. Source: 

What India’s Forest Rights Act Can Teach Us About Data Centers

Meta CEO Mark Zuckerberg leaves after testifying in a landmark trial over whether social media platforms deliberately addict and harm children, Wednesday, Feb. 18, 2026, in Los Angeles. (AP Photo/Damian Dovarganes)


Unpacking the Public Health Turn in Internet Regulation

Closeup of magnifying glass standing on dark surface with beam of light

You Can’t Solve AI Risk Without Solving the AI Evidence Problem

President Donald Trump and House Speaker Mike Johnson of La., listen as Google CEO Sundar Pichai speaks to reporters at the White House in Washington, Tuesday Sept. 29, 2026, after meeting with top executives of AI firms. (AP Photo/Alex Brandon)

AI Hype Helped Google Dodge an Antitrust Breakup. What Comes Next?

Police women wearing pink vests carry a female protester who is peacefully resisting arrest by her arms and legs. A surrounding crowd is watching and photographing and a television broadcast journalist is in the background too.

Four policewomen detain and carry a protester during a demonstration demanding the resignation of Chief Election Commissioner Gyanesh Kumar in New Delhi, India, Saturday, Oct. 10, 2026. (AP Photo/Manish Swarup)

Guilty Until Proven Peaceful Under India’s Preemptive Policing Crackdown

A gaming console. (Photo by Romain Doucelin / SOPA Images/Sipa USA)(Sipa via AP Images)

The EU KIDS Act Could Put Young Gamers' Privacy at Risk

Abstract photo of a wooden speech bubble with many small paper cut outs of speech bubbles in different colors and textures glued on top.

Governments Can Use AI to Repair Public Trust. If They Listen to People First

Illustrative photo captured during the plenary session of October 2025 in Strasbourg, France.

“Behind the scenes - European Parliament in Strasbourg” by 

EU’s Digital Sovereignty Push Runs Up Against Budget Cuts for Civil Society

An illustrative photo of a man walking through the corridors of the European Parliament in Strasbourg, France.

Behind the scenes - European Parliament in Strasbourg

Despite Spyware Attacks on Its Own Officials, EU Response Falls Short

In February, the Trump administration designated Anthropic a “supply chain risk” and 

 the artificial intelligence firm from federal contracting after it refused to loosen restrictions on the use of its models for mass domestic surveillance and fully autonomous weapons. Nearly a year later, however, it is still unclear whether the Pentagon has found an adequate replacement for Anthropic’s foundation models, which were the 

on classified military networks. Media reports indicate that the military is 

Whatever one thinks of Anthropic’s specific red lines — autonomous weapons and mass domestic surveillance — the dustup was yet another reminder of Silicon Valley’s outsized role in shaping the conduct of war. The Pentagon won the immediate fight, cutting deals with other AI firms that 

 more control over how it uses their models in military operations. One of the lessons, according to the Pentagon’s chief technology officer 

, was that “you can’t put the rules and the policies of the United States military … in the hands of one private company.”

But it is far from clear that the military has weaned itself off Anthropic’s technology, much less its broader dependency on the tech industry. In fact, this dependency stretches beyond Silicon Valley and AI. It is the culmination of a decades-long shift in defense policy that has progressively ceded power over critical national security technologies to private companies.

The military has engaged with private industry since the nation’s founding. During World War II, government planners 

 the terms by which industry produced weapons and military equipment, centralizing key decisions ranging from the allocation of scarce materials to setting prices and wages. But by the end of the 1970s, and accelerating in the 1980s and 1990s, the government increasingly outsourced technology development, shedding its own capacity to produce weapons, maintain systems, and manage acquisition along the way. The thinking was that commercial competition would deliver faster, cheaper innovation than the government could — but the shift largely failed to deliver on its promise. As defense budgets fell after the Cold War, the defense industry consolidated into a handful of dominant firms. Competition dwindled and costs failed to fall — all while the military was left 

 of the technology and equipment it needed.

Yet the thinking that underpinned this trend still drives US policy, and AI is just the latest iteration, and potentially the most consequential. As the Pentagon 

 an "AI first" approach to force modernization, the military is rushing to adopt the technology across a wide range of functions — recruiting, supply-chain management, cybersecurity, logistics, targeting, intelligence, and more — and turning to a small handful of powerful tech firms to supply their needs.

In the battlefield, commercial AI is shaping decisions to go to war, identify and strike targets, and anticipate civilian impact. Because the military is by and large renting access to this technology, it has limited capacity to control and oversee the models, algorithms, and data pipelines that guide operational decisions, despite their life-and-death stakes. In some cases, it has become so reliant on a 

 that it has little flexibility to pursue alternative platforms that are more user-friendly, effective, or secure.

Reversing this dependency is essential to restoring democratic control of how the military adopts and uses AI. This begins with rebuilding the government’s capacity to develop, maintain, and oversee key parts of its AI infrastructure — particularly the systems of data that models and algorithms learn from and process into intelligence about the battlefield. This does not mean a full break from buying commercial technology, which when done right can save costs, harness private sector innovation, and redirect government resources towards core defense functions. But it does mean that the government has to absorb the inefficiencies and upfront costs of cultivating the expertise needed to properly harness the benefits of AI, maintain a clear-eyed view of its limits, and ensure human control and accountability over use of force.

The Pentagon’s reliance on Silicon Valley for core AI competencies traces back to the 2017 launch of Project Maven, a flagship initiative to partner with tech firms to accelerate adoption of the technology to generate targets and other forms of battlefield intelligence at scale. A year after Maven’s launch, Google, the Pentagon’s primary industry partner, 

 from the project when thousands of employees protested its involvement. The controversy forced a public reckoning with the dangers of automating warfare, but less so about whether this set of capabilities should even be outsourced in the first place. The data analytics company, Palantir, 

 Google’s contract. The Pentagon’s efforts to privatize the military’s AI architecture continued apace, and it has progressively relinquished ownership and oversight of some of the nation’s most sensitive technologies.

Project Maven’s most consequential output to date is the 

 (MSS), a command and control platform used in many targeting operations today. MSS has facilitated strikes in 

. Although MSS was developed with taxpayer dollars, Palantir owns the platform, and licenses its use to the US military and its allies. The platform houses a vast array of data from military sensors, data brokers, and the open web, as well as the AI tools to turn this information into recommendations and strategies for targeting.

Large language models from Anthropic, OpenAI, and other AI companies are among these tools. After Google’s exit, 

 took turns developing the image and video recognition algorithms that identify and locate people and objects of interest from drone feeds. These algorithms learn what to detect from descriptions that human analysts give to raw footage, a labor-intensive process that is now led by the startup 

. These contracts are so large and complicated that they require 

 that it exercises meaningful control of operational decision-making, even as it cedes ownership, and arguably understanding, of the software underlying these decisions. This is despite the fact that AI is increasingly shaping what commanders see and how they understand the battlefield.

To see why, it helps to know how platforms like MSS work. All of the data coming in — drone footage, satellite images, radio intercepts, radar tracks, location data harvested from mobile apps, social media feeds, and more — has to be organized into a coherent stream of information. Let’s say there is incoming drone footage of a possible enemy vehicle labeling the vehicle as a “truck” and reporting its position in GPS coordinates and the local time zone. Ground radar tracking the same vehicle might instead identify it with the standard military symbol for a transport truck, and report its position in 

grid-based coordinates and a time zone commonly used by the military

. On the surface, these look like reports of two different threats. Properly integrating this data means figuring out that they are in fact the same threat, establishing a ground truth that human analysts and their AI assistants can work off to develop intelligence and other insights.

This process, known as “data integration,” is not simply a matter of combining sources of information. Once disparate feeds are unified into a common system, developers must make consequential choices about what the system allows users to do with the data, and how it records these actions. The feeds must also be able to “talk” to each other, by developing what is known in computer and information science as a 

. The ontology functions like a knowledge graph, defining how data exists within the system, such as the categories they fall into, and all the possible relationships that can be drawn between them.

If a large language model is plotting potential targets on a map, the ontology determines how the map is drawn. A system could merely establish the presence of weapons in a building, or also indicate how many there are, their make and model, whether they are in the basement or on the second floor, and the list of civilian structures in the vicinity. The ontology additionally resolves inconsistencies between data feeds. For example, when one feed labels the building as a hospital and another classifies it as a military facility, the system might automatically flag this for human review, or resolve it in favor of the most recent label.

 the baseline ontological standards that dictate what battlefield systems record, visualize, and display. By outsourcing data integration — and the technology and processes that establish and implement these standards — it risks losing control of the military’s field of vision. Contract engineers who 

 to interpret and apply military doctrine may, for example, inadvertently merge data fields about a building’s “purpose” (how enemy forces might 

) and “use” (how it is currently used). Every screen downstream inherits this change, shaping what commanders and operators see and use to assess the anticipated military advantage of an attack, or its potential civilian impact. This is the sort of error that makes or breaks whether data-driven decisions in the battlefield comply with the rules of engagement and the law of war.

Broad assertions of intellectual property also stymie efforts to ensure accountability for such decisions. To understand how an AI-enabled strike results in civilian deaths, it is not enough to analyze a model’s outputs. The military and those conducting oversight need to be able to examine how a model is trained and fine-tuned, or even the model weights themselves. Given that much of this information is proprietary, lawmakers are 

 whether to compel disclosure, but only for a narrow range of security incidents.

The problem cuts both ways: If the error instead lies in the information that AI is asked to interpret, then access to the rules and logic of data integration becomes critical. For example, this would clarify whether the faulty merging of location data feeds triggered an incorrect AI estimate of a target’s coordinates. But tech companies also own the source code that establishes these standards, making it harder to pinpoint their role in AI-related incidents.

The risks of outsourcing data integration go beyond the Maven program and targeting on the battlefield. The Army has selected the autonomous systems developer Anduril to 

 its flagship data integration effort. The military’s expanding drone fleet also relies on command and control software by 

 to communicate and coordinate with each other. At the Pentagon, the Chief Digital and Artificial Intelligence Office has struck a deal with 

 on its War Data Platform. Though the department retains ownership of this system, it is still reliant on a third party for development and implementation.

These AI dependencies stem from a decades-long shift toward privatization of the US military’s weaponry and technology needs. During World War II and throughout much of the Cold War, the government worked closely with industry, but was at the forefront of defense innovation, directing and coordinating much of the research and development that led to significant advances, from 

. This meant that the government had broad rights to the resulting technology along with deep in-house knowledge of its technical specifications and logic, which allowed it to operate on a level footing with the private sector.

This began to shift as broader support for privatization of the state and deregulation swept the American zeitgeist beginning in the late 1970s. Congress passed the 

, which allowed recipients of federal funding to claim patent rights to innovations created with the help of that funding, and the 

 shifted the burden onto the government to justify why it should have rights to data for systems it bought. The Reagan Administration privatized a 

 determined not to be “inherently governmental,” and government 

spending on research and development declined

 while the private sector’s accelerated, as the potential of commercial applications of internet-related technologies became clear.

The push to make government more efficient through privatization continued throughout the 1990s. As the Cold War wound down and the public demanded a peace dividend, the theory that the private sector was better equipped to develop and maintain government technology than the government itself offered a promise of being able to spend less and get more. The Clinton administration shifted a higher percentage of 

. In doing so, it gave up the government’s rights to data needed to keep these systems in good working order.

Instead of driving down costs, this decision 

 the military’s capacity to maintain and repair its own equipment, forcing it to outsource even when it was cheaper and faster to do so in-house. Most upkeep for conventional surface ships is now outsourced to private commercial shipyards, but ship maintenance delays have increased from 

 between 2011 and 2021. The program to produce F-35 fighter jets, which has faced 

 in cost overruns, has saddled the military with aircraft that can complete all the missions they were assigned 

Repair restrictions are a key reason for this dismal record. Because contracts increasingly award private companies exclusive intellectual property rights, the services are often legally barred from fixing their own gear. This undermines military readiness. In a 2019 letter to the Federal Trade Commission, Marine officers 

 that during a field exercise in Korea, a unit had been forced to choose between fixing a generator and voiding its warranty, or losing access to equipment they needed for their training. They raised concern that contracts establishing similar warranty conditions on tactical trucks would prevent Marines from making necessary repairs in hostile environments.

The dangers to troop safety have been such a concern that, in 2025, rare agreement across the political spectrum — ranging from former 

 — led to draft provisions in that year’s annual defense policy spending bill that would restore the military’s “

, and lawmakers eventually dropped them from the bill. Lawmakers are pushing for them 

 as of writing. In the meantime, services are 

The privatization of military technology has had lasting effects beyond the mismanagement of conventional weapons. Congress passed the 

Federal Acquisition Streamlining Act of 1994 (FASA)

, establishing a statutory preference for commercial technology, while further limiting government rights to underlying intellectual property as well as 

, contracts would limit the government’s ability to look “under the hood” to determine whether contractors were fairly assessing the costs of what they were producing. The hope was that deregulation would encourage commercial vendors to sell to the department and allow the military to take advantage of private sector growth. The government’s faith in privatization also led it to shed in-house capacity to manage contracts and defer investment in the expertise needed to understand and leverage emerging technology.

This approach has led the military to outsource enterprise software, low earth orbit satellite internet, and cloud computing. In some cases, the Department of Defense has become so embedded in a single vendor’s ecosystem that it cannot practically extricate itself even when better technology becomes available. The Navy, for example, 

 that switching from Microsoft’s Azure cloud and enterprise software to another approved provider would require a complete reengineering of key information systems and risk mission failures. This type of lock-in forces the government into impossible choices between its obligation to protect taxpayer dollars and safeguard national security.

To be sure, the government can benefit from buying goods and services from the private sector and should not recreate truly commercial commodities where prices are competitive, the threat of vendor lock-in is low, and the service provided is not core to the conduct of government. In those cases, it might make sense to cede rights to development, control, data, or repair. That takes the best features of the private sector and allows service members and taxpayers to benefit.

But the reality today is far different for much of what the government is buying. In the case of AI, outsourcing the data backbone that powers the technology creates a dependency loop. Contractors are 

, establishing an analytic framework, and creating an interface for AI models – all mission-critical functions where the lack of ownership or robust IP rights leads to vendor lock-in and makes switching to another vendor prohibitively costly and challenging.

Further, because nearly the entire capability is effectively rented, the government has little incentive to hire military personnel or civil servants with the technical skills to understand and support this work. This public-private talent gap saps the Department of Defense of its capacity to provide oversight, hold contractors accountable, and make build-or-buy decisions that account for long-term strategic needs, such as controlling lifecycle costs and sustaining in-house expertise. This shortsightedness becomes self-reinforcing, as any subsequent innovation is more likely to happen outside the government, leaving the military with little choice but to rent access to critical functions.

How can the US government get out of this trap? One path is for the Pentagon to develop, manage, and maintain its own data integration systems. Former Secretary Driscoll has 

 Ukraine’s flagship platform for integrating drone, satellite and other data feeds, known as 

, as a model for the Army’s own efforts. Delta started out as a volunteer initiative, but is now state owned. The government has enabled widespread adoption throughout the military, and added analytic, communication and security features 

 comparable to development cycles in the private sector.

Ukraine’s Delta does not nearly approach the scale of data integration the US military will need to undertake. But it shows that it is possible for the military to retain ownership of its data backbone without sacrificing innovation, provided it acquires the technical expertise necessary to engage successfully with emerging technologies. Ownership would also give the Pentagon much greater visibility into critical technology and remove an unacceptable point of leverage private companies have over the capabilities the military uses on the battlefield. This is the cleanest way for the government and the military to reclaim strategic control of its AI future.

An alternative would be for the government to insist on securing substantially more robust rights to intellectual property, and legally mandate 

open and interoperable software standards

. These standards enable the military to freely inspect and audit the source code underlying its data backbone, and more actively collaborate on modifications and updates. This would not only enhance oversight of data-driven failures in the battlefield, but would better equip the government to enforce policy fixes down to the code level.

Contractors should also be required to store data on platforms developed for government in a format that the government can export and re-use on alternative platforms. This would reduce the government’s reliance on a single vendor, lower the cost of switching to a competitor if needed, and enable it to integrate its data and the accompanying ontology with other IT systems not by the same vendor.

Both options, however, presume that the government has the technical capacity to build its own technology or to check the work of contractors, even as decades of privatization and deregulation have 

, gutting in-house expertise and shrinking the workforce charged with administering ever-more complicated contracts. As the defense procurement budget rose at the turn of the century, the defense acquisition workforce failed to keep pace, increasing 

only 21 percent compared to a 43 percent rise

 in contract spending from 2001 to 2015. Similarly, spending on defense contracts rose by 41 percent from 2015 to 2020 but the 

. The department’s Intellectual Property Cadre, which provides expert advice, resources, and strategic guidance to the defense acquisition workforce, also suffers from 

. And this is all prior to the departure from the Pentagon of thousands of civilian employees in the past year 

 of the Department of Government Efficiency.

The Defense Department needs more technology experts with the authority and funding to build at scale. Pockets of internal software and data expertise and capacity in the US military provide a glimpse of what is possible when the government makes productive investments in the military and civilian workforce. The National Reconnaissance Office’s efforts to develop a “

,” for example, provided an early blueprint for how Maven would fuse, analyze and visualize disparate streams of data. The Marine Corps launched a 

 also have teams building apps for service members, leaning into AI applications, and offering government-owned solutions to real world problems.

But in order for the government to fully regain control of its data infrastructure — and its understanding of the battlefield — it will need to hire attorneys with experience dealing with AI companies and issues of data rights; managers who know how to lead software development teams; contracting officers that can navigate acquisition regulation and ensure the military retains sufficient rights in their data and software; and engineers who can run enterprise systems. These investments are necessary regardless of whether the goal is ownership or simply effective oversight and management. To entrench lasting institutional knowledge, these positions cannot be filled by 

 or time-limited fellows. They must be full-time staff with opportunities for career advancement, funded by the core defense budget.

None of this necessary change will be cheap or fast. A generational evolution in the Pentagon’s workforce requires acceptance of upfront costs and a commitment to a long-term vision. Congress already requires the department to give “

 for defense functions where it is necessary to maintain sufficient government expertise and technical capabilities. This would seem to demand a review of the work that is currently being managed by defense tech contractors.

 and other procurement laws that require the government to default to commercial technology. Instead, it should ensure that the government evaluates the potential impact of vendor-lock in and market concentration, long-term operational readiness, and the need to preserve effective management and oversight in-house before committing to a commercial option.

As AI becomes more pervasive in military operations, reasserting government control over the universe of data that supports it will become even more important. This will reduce the military’s reliance on any one vendor for key data integration functions, enabling it to more freely switch between platforms based on cost, effectiveness, and operational need. Moving more of this technology in-house will also cultivate the resources and expertise needed to align data-driven decisions in the battlefield with legal constraints on the use of force.

, for now, for the government to implement a five-year, approximately 

 health cooperation framework with the United States, including 

 in US support. The agreement provides for the sharing of 

 and US access to Kenyan health information systems. A legal challenge to the framework, including its data-access provisions, is scheduled for a 

Kenya’s deal reflects a broader shift: access to data, digital infrastructure, and technology is becoming part of the bargain between wealthy donors and poorer recipient countries. As traditional aid budgets shrink, AI offers public institutions a way to do more with less, and health is an obvious place to start. Sub-Saharan Africa faces a shortage of 

, and Rwanda has one health care worker per 1,000 patients, less than a quarter of the 

 at which the World Health Organization defines a shortage. Tools that help clinicians triage, diagnose, and keep records could help close that gap.

Donors are already investing. In January, the Gates Foundation and OpenAI 

, committing $50 million in funding, technology, and technical support to bring AI tools to 1,000 primary health clinics in Africa by 2028, beginning in Rwanda.

Commitments like this could do real good, and AI’s potential to bring efficiencies, and even breakthroughs, to long-standing gaps in care is significant. To deliver on it, AI partnerships need to avoid the old problems of tied aid: long-term obligations that bind recipients to the priorities, standards, and suppliers of whoever provides the help. Whether health systems keep the freedom to choose their tools will depend on the terms of these deals, and on whether the public can see them.

For decades after World War II, wealthy countries routinely attached a condition to international aid: recipients had to spend it on the donor country’s goods and services. Such arrangements became controversial because they restricted competition and undermined recipient choice. Estimates from the Organisation for Economic Cooperation and Development suggest that tied aid can increase project costs by 

, and its Development Assistance Committee spent decades 

 members to untie their aid, particularly when it flowed to the least developed countries.

AI assistance risks reviving this arrangement in a subtler form, one that creates dependence without any requirement to buy. Free services come first, and the data architecture, workflows, skills, and running costs they leave behind can make switching steadily harder.

Increasingly, development assistance comes not only as financing, but also as technology, technical support, and credits for proprietary services. Horizon1000’s $50 million mixes funding with OpenAI technology. In May, Anthropic and the Gates Foundation 

 over four years for public goods across health, education and other sectors. Anthropic’s half of the commitment will consist of Claude usage credits and technical support, while the foundation will provide grants and program expertise. Among the initiatives is work to improve AI performance in African languages, with resulting datasets intended for public release. A Gates Foundation director 

 that governments’ fears of proprietary lock-in shaped its public-goods focus, and that the language datasets it produces will be released openly.

Those commitments bring important opportunities for support, and free credits do not by themselves bar a ministry from using another model. Still, credits usable with only one supplier can steer public institutions toward that provider, creating a form of de facto tying even without a contractual obligation to stay.

Companies also partner with governments directly. In February, Anthropic 

signed a three-year memorandum of understanding with Rwanda

 covering health, education, and the public sector, giving government developer teams Claude and Claude Code along with training and API credits. 

Tech Policy Press has previously examined accountability concerns in that deal

. While the partnership emphasizes capacity building and local autonomy, the public announcement does not specify how the agreement addresses data handling, post-credit pricing, or exit rights—details that would help observers assess its long-term sustainability.

The larger question reaches beyond any one company: development institutions spent decades untying aid from donor-country suppliers, but have yet to develop comparable safeguards against dependency in assistance delivered through proprietary AI, credits, and infrastructure. What will matter is whether the datasets, interfaces, and workflows built with these tools stay useful if a ministry switches providers.

Aid organizations and recipient countries should also guard against extractive data practices. Donors are already writing access to health data into aid deals. As of August, 34 countries had signed US health agreements, according to an August 

, gives US officials direct, real-time access to nine national health data systems for seven years, including the one that manages patients’ electronic medical records, though it calls for shared data to be aggregated with personal identifiers removed. U.S. senators asked whether such data could be shared with US-based third parties, including to train AI models, but no publicly available answers have been released. The State Department 

, citing a policy against commenting on congressional correspondence, though a spokesperson said neither the US government nor private American companies receive personally identifiable information under the agreements.

In February, meanwhile, a diplomatic cable reported by Reuters 

 to lobby against foreign rules on how US tech companies handle data, warning they could interfere with AI services.

These agreements do not establish that health data will be used to train commercial AI models, nor do they demonstrate AI vendor lock-in. What they show is that access to potentially valuable data can become a condition of development assistance, raising questions about who controls that data and who benefits from its future uses.

Together these cases suggest four ways aid that runs on data and AI can bind recipients:

 one provider’s model embedded in public systems, raising the cost of switching.

 free deployment now, with costs that arrive after the grant ends.

 long-term access to health data as a condition of support.

 pressure to adopt the donor’s rules on data flows and infrastructure.

These categories stretch the traditional meaning of tied aid, and each can reproduce its central problem: assistance that narrows the recipient’s freedom to choose. Data conditions in US health agreements are documented. Standards pressure shows up in US diplomacy, though the public record does not show it written into agreements. Vendor dependence and rising post-grant costs in AI partnerships remain risks, not established by the available public record. Without public contractual terms, governments and citizens cannot tell whether the technology they receive today will leave them real choices tomorrow.

US government health agreements face statutory disclosure requirements, which senators have argued are not being fully met. As of late September, the US government had 

 of 16 of the 36 signed agreements, ten of them only in response to Freedom of Information Act requests from Public Citizen, and had published none of the associated data-sharing agreements; the one that is public, Kenya’s, was obtained and posted by a news outlet. Private and philanthropic AI partnerships are not subject to the same general public-disclosure framework, leaving gaps in what outsiders can assess about their terms.

Consider a hypothetical ministry three years into subsidized AI services for its clinics. Does it own the model adapted to its patients? Can it move their records to another provider? What will inference cost in year four, once the grant ends?

Recipient governments are active partners in these deals. Rwanda aims to use AI to build decision-support tools for its more than 60,000 community health workers, 

. Disclosure would strengthen governments like Rwanda’s, and it would help companies too: published terms build trust with ministries, make it easier to extend a partnership from one country to the next, and answer suspicion before it hardens. During the fight over the US health deals, Public Citizen warned they could 

 African states built in response to COVID-19 vaccine inequities. A ministry that can see what others have negotiated is better placed to bargain for itself, and shared knowledge across countries strengthens their collective hand.

Applying tied aid’s lessons to the AI era

The first step toward a more equitable approach is disclosure. Any government, foundation, or company offering AI tools to public health or education systems as aid should publish the agreement’s material terms, even where the full contract carries narrow redactions for patient privacy, security, or trade secrets. At minimum, that means the model and infrastructure providers; data access, retention, and secondary-use rights; who controls fine-tuned models and other derived assets; post-grant prices; and exit provisions.

The second step is a credible path to exit. Where traditional aid reform sought to give recipients freedom to choose their suppliers, AI partnerships must also preserve their freedom to leave them. This requires three safeguards:

 a ministry can export its data and configurations in usable formats, get reasonable migration help, and switch models or providers without rebuilding its systems or paying punitive fees.

 recipient institutions keep enforceable rights to access, control, export, and govern their data, consistent with domestic privacy law, with clear terms on who controls models adapted from it.

 recipients can test performance themselves, including in local languages.

Donors and philanthropies could make disclosure and portability conditions of their grants. AI companies could publish standard partnership terms, building on commitments like open datasets, while recipient governments coordinate to demand common protections. The African Union’s 

 offers a framework for doing so, and the 

 already mobilized around the US health deals could extend its scrutiny to AI partnerships.

Donors spent decades learning that aid tied to the giver’s interests undermines what makes development work: solutions that are locally led and locally owned. The AI partnerships now spreading through public health systems should start from that lesson instead of relearning it. AI could be among the most valuable assistance developing countries have received. Whether it lives up to that will depend on its terms, and on whether countries can see them.

The Tech Policy Press newsletter is published every Sunday.

Subscribe here to receive it in your email inbox

I had the good fortune to see many in the Tech Policy Press community this week in Montreal at 

, a conference hosted by Paradigms, McGill University, the Max Bell School of Public Policy, and the Centre for Media, Technology and Democracy.

 made the case that democracy requires citizens to have both knowledge about and the means to participate in the development of the technologies that define the information environment and so much about how society works.

, in her remarks, reminded us why authoritarians want to cut the people out of the picture. “The strong man needs the tech oligarch,” she said. “The tech oligarch needs the strong man. The strong man weaponizes the law. The oligarchy gives the strong man reach. The strong man gives the oligarchy protection.”

I ducked out of the conference briefly on Thursday to cohost a virtual event with 

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe