Somewhere in Kansas, a grandmother uses a pharmacy app to manage prescriptions, receive refill reminders, and arrange pickup.
That is the territory Juan Caro and Peter Brown map in their recent argument about American AI policy [1]. Caro is a former Deputy Assistant Secretary of Commerce; Brown is a retired Rear Admiral and former Homeland Security advisor [1]. Their claim is not that AI is dangerous. It is that AI is useful — genuinely, concretely useful — and that the usefulness is precisely why its data exhaust deserves protection. The two ideas are not in tension. They are the same idea.
What the Tool Actually Does
The pharmacy app is not a metaphor. It manages prescriptions, sends refill reminders, and arranges pickup. An AI assistant layered on top could help the user understand dosage instructions and prepare questions for her pharmacist. That is a real gain: less time on hold, fewer missed doses, more independence. The benefit is not abstract. It shows up in the ordinary hours of a day.
This is technology serving a human purpose — helping someone stay independent and spend more time with the people she loves [1]. The phrasing matters because it locates the value where it actually lives: not in the model’s benchmark scores, but in the grandmother’s afternoon. An AI that explains a drug interaction in plain language has done something a pamphlet cannot. It has answered a question she did not know how to ask.
The same pattern repeats elsewhere. Entrepreneurs use AI to compete with larger firms that have more staff. Workers use it to develop skills at their own pace. Caregivers use it to navigate days that are already too full. In each case, the gain is specific: a task made lighter, a decision made clearer, a person left with more room to act. That is the first thing to hold onto. The tool lifts.
The Data Behind the Lift
Here is where the story turns. To make the reminder arrive on time, the app needs to know what the prescription is. To help with the pharmacist conversation, it needs the medical history. The information that makes the tool useful is the same information that makes it valuable to someone else.
In 2023, Duke University researchers purchased sensitive information about American military personnel, veterans, and their families for as little as 12 cents [1]. The records included health, financial, religious, and demographic details. No hack was required. The data was for sale on the open market.
That finding is the hinge of the argument. It shows that the threat does not depend on a foreign government breaking into anything. Commercial data markets can provide another route. China does not have to steal every piece of information it wants when some of it is available for purchase.
The danger is not in any single record. A refill date is mundane. A ZIP code is mundane. A religious affiliation is mundane. But combined, they become revealing. Health information, location histories, and financial details together can expose routines, relationships, and vulnerabilities. AI makes that combination easier to perform at scale. An adversary could use the resulting picture to identify military families, target government personnel, or sharpen influence operations. Information collected to make life easier could become a tool for surveillance and coercion.
The Policy Already in Place
Washington has not ignored this. The Justice Department’s Data Security Program, effective since April 2025, prohibits or restricts certain transactions that give countries of concern — including China — access to Americans’ bulk sensitive personal data and government-related data. That is a foundation, not a finished structure.
That foundation should be enforced and strengthened while preserving beneficial AI development [1]. First, give Americans meaningful control: companies should clearly explain what sensitive information they collect, why they need it, and whether they share it. Services should collect only what they require and offer real choices about unrelated uses. A daily prescription reminder should not require surrendering control over an intimate medical history.
Second, make national security protections follow the data. Regulators should scrutinize intermediaries, ownership, and onward transfers so adversaries cannot evade restrictions by routing purchases through another country. Enforcement should address deliberate efforts to split transactions to avoid applicable thresholds.
Third, establish workable federal standards for privacy-preserving data use. Removing names alone may not prevent identification. Standards should require testing for reidentification risk, documentation of lawful sourcing, safeguards against misuse, and accountability for subsequent transfers.
The Landscape That Separates Knowledge From Assumption
The Duke findings are specific: sensitive data on military personnel and their families was purchasable for 12 cents [1]. The Data Security Program is specific: certain bulk transactions are now restricted [1]. What we assume is that these measures are sufficient. The gap between the two is where the argument lives.
Finally, Commerce and Justice should coordinate their authorities and assess whether targeted export controls could address sensitive AI-training data transfers beyond existing protections. Additional controls, the authors write, should address demonstrable national security risks and provide clear compliance guidance. The word “demonstrable” is doing work here. It keeps the policy anchored to evidence rather than fear.
The landscape is not a single law or a single agency. It is a set of overlapping authorities that must be aligned. Privacy and American strength, the authors argue, belong in the same conversation. Protecting personal information preserves citizens’ freedom to act while denying adversaries opportunities to exploit them. That is a framing choice, and it is a deliberate one: it refuses the trade-off that says security requires surveillance.
What Follows
The consequence is straightforward. If the data that powers useful AI is also the data that foreign adversaries want, then the protections must be built into the tool, not bolted on after. Caro and Brown put it directly: privacy should be built into innovation, giving consumers confidence and companies greater certainty. American researchers and businesses get clearer rules. Consumers get confidence. The grandmother gets her refill reminder without wondering who else is reading it.
The alternative is a future where the same app that helps someone manage her prescriptions also feeds a system she never agreed to. That is not a hypothetical. The Duke study showed the market exists. The Data Security Program showed the government can act. What remains is enforcement and coordination.
Americans should not have to become data-security experts to refill a prescription or use a digital assistant. Businesses and government must make responsible choices possible and enforce protections against foreign exploitation. The ambition, as Caro and Brown frame it, should be powerful American technology worthy of public trust.
The grandmother in Kansas should use her pharmacy app confidently. The entrepreneur should grow a business. The worker should master new tools. Their information should serve their ambitions, not Beijing’s. More agency. Greater purpose. Stronger security. That is the future the policy is meant to protect — and the reason the tool is worth defending in the first place.