The Department of Veterans Affairs last month moved to alter how it evaluates disability ratings, announcing a new rule allowing it to factor in the effects of medication when determining what kind of benefits a disabled veteran could receive.
The reaction was immediate and fierce. Veterans, advocates, and major service organizations understood exactly what was at stake. Disability compensation is not a reward for being unable to function. It is not unemployment insurance. It is recognition that military service caused lasting harm. If a veteran takes medication that helps control symptoms, that does not erase the underlying injury. If a veteran manages to keep working despite pain, trauma, or physical limitation, that is not an excuse to pay them less.
The VA backed off that rule amid the backlash. But now, they have a new plan on how to heed Project 2025’s call to limit benefits, after the conservative manifesto claimed that “expand[ing] disability benefits to large populations without adequate planning have caused an erosion of veterans’ trust in the VA enterprise.”
The new plan is artificial intelligence, which the VA wants to use to scan for fraud. Trump has cited fraud as justification for gutting the VA, but the department’s fraud rate is less than one one-hundredth of one percent. Large Language Models (LLMs) can be helpful to large government organizations — I know, I wrote the white paper the VA used to help build their AI policy — but they are bad for finding things like fraud. This is in part because LLMs, unlike humans, will not quit when they don’t find any fraud. They could even just make up fraud, something developers call LLM hallucinations.
The problems with setting LLMs loose in a system where actual fraud is incredibly rare are self-evident.
The VA’s plan is to use Disability Benefit’s Questionaries (DBQ) to identify so-called signs of fraud. These questionnaires are a critical part of the disability claims process. They are often completed by private medical providers and used by the VA to determine the severity of a service-connected condition and the compensation a veteran receives. Under the new plan, the VA will use AI to review massive numbers of these forms and flag submissions that appear suspicious. This could include forms with missing fields, repetitive language, altered documents, or forms with a provider located more than 100 miles away from the veteran.
On paper, that may sound reasonable. No one supports fraud. Veterans do not want scam artists clogging the system, exploiting the public, or undermining trust in claims that are legitimate. If there are mills churning out fake forms or unethical providers exaggerating conditions, they should be investigated and shut down. But that is not where this story ends. Because once the VA gives a machine the power to label a claim as suspicious, the burden almost always shifts back onto the veteran — and that is where the real danger begins.
A claim flagged by an algorithm is not proof of fraud. A veteran might see a doctor far from home because specialists are limited, because telehealth has expanded, or because they moved during the course of treatment. Similar language across forms may reflect standard medical terminology, not deception. Missing information may be a clerical mistake. A document may be incomplete for reasons that have nothing to do with dishonesty. But once a system marks a DBQ as potentially problematic, the veteran is no longer simply filing a claim. They are defending themselves against an invisible accusation. After all, the AI model does not know anything about the veteran, it is just an LLM being told to find fraud.
This will mean new exams, new delays, new scrutiny, and new opportunities for the bureaucracy to wear people down. This creates a culture of fear. Veterans will give up. And this is by design.
The larger pattern here is impossible to ignore.
First, DOGE came in and fired tens of thousands of veterans from federal service. Then, the VA tried multiple schemes to make changes to the benefit system through legislation. When members of Congress weren’t willing to use political capital on screwing veterans, Trump cried “fraud” in the VA system, which didn’t actually exist. The VA then pursued a rule change that would have made it easier to reduce disability ratings by redefining how conditions are evaluated. When that failed, the conversation shifted to fraud detection, automation, and AI.
The language changes, but the practical result is the same: fewer veterans receiving the compensation they earned, and more power concentrated in a system that too often treats veterans as problems to be pushed into the private market rather than as people to be served, people who dedicated their lives to defending the United States.
Veterans earned benefits through service, injury, sacrifice, and, in too many cases, permanent damage that will never fully go away. If Trump and the GOP do not want to care for veterans, they need to stop creating so many of them.
