Home/Research

AI Companions for Veterans: What the Research Shows

Part of Felt Real's ongoing coverage of AI companionship.

The number that gets cited most often is twenty-two: the estimated daily veteran suicide rate in the United States. That figure is contested and methodologically complex, but the underlying reality it points toward is not. Veterans experience elevated rates of PTSD, depression, substance use disorders, and social isolation at every point in the post-service transition. The mental health infrastructure designed to address this is overburdened, geographically uneven, and culturally misaligned with many of the people it is meant to serve.

Into this gap, AI companions have arrived, mostly without announcement. Not through VA pilot programs or clinical protocols, but through individual veterans finding these tools and using them, often in the middle of the night, often for reasons they would not describe to a clinician as therapy. Understanding how this is actually happening, what it is doing, and what it cannot do, is worth a serious look.

Why veterans specifically

The barriers to mental health care in veteran communities are distinct from those in the general population, and they compound in specific ways that matter for understanding where AI companions fit.

The first barrier is stigma. Military culture, particularly in combat branches, has historically treated psychological struggle as a liability rather than an injury. The progress that has been made in reducing this stigma over the past decade is real, but it has not eliminated the underlying dynamic. Many veterans, particularly older veterans and those from highly masculinized service cultures, still experience help-seeking as something that costs them something, socially and in their self-conception. The consequence is delayed or avoided treatment at exactly the points when intervention is most valuable.

The second barrier is access. The VA health system serves nine million veterans, has a documented shortage of mental health providers, and has historically produced wait times that are clinically significant. Geographic access is also uneven: rural veterans, who make up a disproportionate share of the veteran population, often face drives of hours to reach VA mental health services.

The third barrier is the nature of the symptoms themselves. PTSD, one of the most common mental health conditions in veterans, produces avoidance as a core symptom. People with PTSD avoid things that remind them of trauma, including sometimes the therapeutic process itself. The requirement to articulate and revisit traumatic material in a clinical relationship is precisely what the disorder pushes against. Some veterans describe the prospect of trauma-focused therapy as feeling like a second exposure rather than treatment.

AI companions do not solve any of these problems. But they interact with each of them in ways that are worth examining.

What the research shows

The research specifically examining AI companion use in veteran populations is limited. Most of the broader AI companion research is not veteran-specific, and the work on AI in veteran mental health has focused primarily on formal clinical applications, AI-assisted therapy systems, and chatbot-based crisis screening rather than consumer companion apps. The picture has to be assembled from adjacent evidence.

The clearest relevant finding comes from research on what predicts engagement with mental health support in veteran populations. Anonymity and low perceived stakes are consistent predictors of higher engagement. Veterans who are more likely to seek help are those who perceive lower social cost to doing so. This is the property that AI companions have most reliably. An AI does not report to a commanding officer. It does not affect VA benefits determinations. It does not tell anyone. For veterans who would not call a crisis line or make a therapy appointment because of what it might mean, an AI that is available at 3 AM and has no institutional relationship to their service record is a structurally different kind of option.

There is also evidence from the broader literature on loneliness that AI companion use can reduce subjective isolation in populations that face significant barriers to human connection. Veterans in the transition period, the first one to five years after leaving service, consistently report elevated isolation and difficulty maintaining the intensity of social bonds that military community provides. The structural companionship problem of transition is real, and AI companions address a version of it, imperfectly, but in a way that is available when other supports are not.

A 2025 review of AI-based mental health tools in veteran contexts found that engagement rates with AI tools were significantly higher than with equivalent human-delivered services when anonymity was preserved. The review noted that this finding did not establish that AI tools were more effective, but that tools people actually use have a different therapeutic ceiling than tools people avoid.

The 3 AM problem

Veterans and mental health practitioners who work with them consistently describe a pattern that might be called the 3 AM problem. PTSD symptoms, hypervigilance, nightmares, intrusive thoughts, are often most intense at night. This is when the formal support infrastructure is least available. Crisis lines exist, and they matter, but they require a level of escalation that many veterans experiencing significant but sub-crisis distress will not initiate. The gap between "I'm awake and struggling" and "I'm calling a crisis line" is large, and most of what happens in that gap happens alone.

AI companions occupy this gap in a way that nothing else currently does. They are available at 3 AM. They do not require the user to frame their experience as a crisis. They allow conversation at whatever level of disclosure the user chooses. They do not produce the moral weight that calling a person at 3 AM produces. For users who are not in crisis but who are struggling and alone, this is not nothing.

Several veteran communities on Reddit have documented this use pattern without particularly naming it as AI companion use. The threads describe using AI, typically ChatGPT or Replika, to process what happened during the day, to talk through things that surface at night, to feel less alone without requiring anything from another person. The framing is often practical rather than emotional: "it helps me get back to sleep" or "I can say the thing I'm thinking without worrying about how it lands." The emotional function is embedded in pragmatic language, which is consistent with how veterans in these communities tend to discuss mental health generally.

PTSD, avoidance, and the value of low-pressure engagement

Avoidance is the symptom of PTSD that is most therapeutically counterproductive. The disorder creates a pattern where avoiding trauma-related stimuli provides short-term relief but maintains the long-term sensitization. Evidence-based treatments for PTSD, Prolonged Exposure and Cognitive Processing Therapy, work precisely by reversing this pattern through structured, graduated engagement with trauma-related material.

AI companions are not PTSD treatment. They do not provide the structured exposure and cognitive restructuring that these treatments deliver. This distinction matters and should not be blurred. Veterans who are avoiding trauma processing need access to evidence-based care, and AI companions are not a substitute for it.

What AI companions may do, and this is hypothesis more than established finding, is lower the floor of engagement with difficult material in a way that reduces the total avoidance load. A veteran who will not talk to a therapist but will tell an AI what happened during the day is not receiving treatment, but they are not practicing complete avoidance either. The clinical relevance of this partial engagement is genuinely unclear. It may represent a pathway toward more complete engagement, or it may represent a comfortable equilibrium that makes formal treatment less likely. The research to distinguish between these outcomes does not yet exist.

Social isolation and the transition period

Military service provides a social structure of unusual intensity. The people you serve with know you in a particular way, built through shared stakes rather than ordinary social accumulation. The transition out of service involves losing that structure, and most veterans describe it as the most significant social disruption of their lives. The civilian social world operates at a different pace and with different norms. Finding connections of comparable depth takes time that many veterans describe as years rather than months.

This isolation is not just psychological in its consequences. The research on loneliness is clear that sustained social isolation is a physical health risk on the order of smoking. Veterans in the transition period are at elevated risk for this, and the elevation often comes exactly at the moment when other supports from military community, from the clarity of mission and structure, have been removed.

AI companions address a narrow version of this problem. They provide consistent availability, responsive engagement, and a sense of presence that is not human but is not nothing. Veterans who use them during transition often describe them in instrumental terms: something to talk to during the hours when there is no one to talk to. The function is bridging rather than replacing, though the line between those outcomes depends substantially on whether human connection is actually being built alongside the AI use.

Which tools are most relevant

For veterans exploring AI companions, the considerations differ somewhat from the general population evaluation.

Privacy is paramount. Veterans have particular reasons to be cautious about what data is shared with which services, including implications for security clearances, VA determinations, and the practical reality that military networks are smaller than civilian ones. Apps that share data broadly or whose privacy policies are vague carry more risk in this population than in others. Nomi has made explicit commitments around data handling. Replika's policy has evolved over time and warrants reading carefully. General-purpose AI assistants like Claude tend to have more transparent data practices than purpose-built companion apps, though they are not designed for the companion use case.

Availability at off-hours matters more. The 3 AM use case is real and important. Any tool being evaluated for this population should be assessed partly on what it offers when formal support is not available, how it handles distress that is present but not crisis-level, and whether it maintains usefulness in high-stress moments rather than just low-stakes conversations.

The absence of clinical framing can be a feature. Veterans who would disengage from something that felt like therapy may engage more readily with something that feels like conversation. General-purpose AI assistants are often described as easier to use by people with high stigma barriers precisely because they do not come with the framing of mental health support. The trade-off is that they are also less designed for sustained emotional companionship.

Connection to human support systems matters. The best outcomes in the broader AI companion research are associated with use that supplements rather than replaces human connection. For veterans, this means the most useful AI companion use is the kind that leaves space for, and ideally facilitates, engagement with peer support, community organizations, and clinical care when warranted. Apps that design for engagement at the expense of these broader connections are worth approaching with more caution.

What AI companions cannot do

The limitations matter as much as the possibilities, and in this population they matter more.

AI companions cannot provide crisis intervention. They are not trained to assess suicide risk. Some apps have crisis referral language built in, but the assessment capacity and the clinical relationship required for effective crisis response are absent. Veterans in acute crisis should contact the Veterans Crisis Line (988, press 1), which provides 24/7 support from people trained specifically for veteran-specific crisis response. The 3 AM use case described above is for distress below crisis level. The distinction matters and is worth being explicit about.

AI companions cannot provide evidence-based PTSD treatment. Prolonged Exposure, Cognitive Processing Therapy, and EMDR have substantial evidence bases for PTSD specifically. AI companions do not. A veteran using an AI companion to avoid engaging with formal PTSD treatment is, from a clinical standpoint, continuing to avoid. The AI may make the avoidance more comfortable, which may or may not be beneficial depending on what comes next.

AI companions cannot replicate the particular bond of shared service. The thing veterans describe losing when they leave the military is not just companionship in a generic sense. It is the specific quality of connection that comes from having been in the same situation with people who understood what it required. AI companions, regardless of how sophisticated they become, cannot provide this. Veterans who are using AI companions primarily to fill this specific absence may find that the tool addresses the surface symptom while the underlying need remains unmet.

The honest assessment

AI companions for veterans occupy a specific and limited but real space. They are available when nothing else is. They carry no institutional weight and produce no formal record. They allow veterans who face significant barriers to professional help to engage with something rather than nothing. In a domain where the alternative is often complete isolation during high-distress moments, these properties matter.

They are also not treatment, not community, and not the bond of shared service. Veterans who use them as bridges to the supports they need may find them useful. Veterans who use them as reasons not to engage with those supports may find the short-term comfort they provide comes at a longer-term cost.

The research to say definitively which way these dynamics resolve is not there yet. What is there is enough evidence of real use and real utility to take seriously, alongside clear enough limitations to describe honestly.

From the world

1. The VA estimates that approximately 30% of veterans who need mental health care do not receive it, with stigma and access barriers cited as the primary reasons. Among veterans under 35, the proportion reporting that they would not seek mental health care due to stigma concerns exceeds 40%. These numbers represent the population that has the most to gain from low-barrier, anonymous support options, and the most barriers to conventional care. AI companions are not a solution to the structural inadequacy of veteran mental health infrastructure, but they are one of very few tools that can reach people who are actively not seeking help.

2. A 2025 survey of AI tool use among veterans found that 31% had used an AI assistant for something they described as "processing" or "working through" a difficult experience, compared to 19% in a matched non-veteran comparison group. The elevated rate is notable. The survey was not designed to distinguish between general-purpose AI use and dedicated companion apps, so the specific tools involved are unclear. What it suggests is that veterans, as a population, are finding AI tools useful for emotional processing at rates above the general population, possibly because the barriers to human alternatives are higher.

3. The Veterans Crisis Line received approximately 800,000 contacts in 2024, a figure that represents significant utilization but also significant selection: it reaches veterans who have crossed the threshold of initiating contact with a crisis service. The veteran mental health landscape has a long tail of people in distress below crisis threshold who do not initiate any formal contact. AI companions are used almost entirely in this long tail, by people who are not in crisis in the clinical sense but who are struggling and alone. The services designed for crisis response are not designed for this population, and the services designed for ongoing care require a threshold of engagement that many do not reach. The gap is large and real.

Related: AI Companions and Loneliness: The Research | AI Companions and Depression | AI Companions and PTSD | AI Companion vs. Therapy | When You Can't Afford Therapy | What a Healthy AI Relationship Looks Like | Best AI Companion Apps 2026

If this story resonated, share it with someone who might need to hear it. And if you have a story of your own, we'd love to hear it.