The era of GenAI.mil is here. Users have mixed reactions and many questions.

A little over a week after GenAI.mil’s rollout, service members and defense officials have met the Pentagon’s newest hub for commercial AI tools with mixed reception — and many questions.

While the military has been exploring and using large language model AI systems for more than two years, the Defense Department’s latest, large-scale generative AI push has hit the organization fast and expansively, meaning troops and civilian employees who may not have been exposed to these models before are now being told to use them in their daily operations.

The rapidly emerging field of genAI and associated LLMs — which generate convincing but not always correct software code, images, audio and other media from human prompts — present both promise and complex risks to U.S. national security and military operations. 

Last week, DOD leaders unveiled the new GenAI.mil platform to supply industry-built AI assets to almost its entire workforce, directly on their government desktops. 

Google Cloud’s Gemini for Government products are the first to be offered in the one-stop-shop, and others will soon follow. But sources said the Pentagon has yet to release comprehensive guides or formal training for the platform, beyond prohibitions on uploading personal data into it.

This breakneck launch has left some military members with concerns about privacy and safety in the absence of clear guidelines for use, while others have expressed excitement about the platform, already employing it for various tasks down to the tactical level. 

DefenseScoop spoke to multiple current and former defense personnel about the GenAI.mil deployment — from senior officials to junior service members and industry experts. The majority requested anonymity to avoid retaliation because they weren’t authorized to speak to the press.

Most grappled with how much trust (if any) the platform’s human operators should give it, including about data security and privacy. For some, that contention amounted to a question that broader society is also wrestling over: where should AI use begin and human thinking end? 

With GenAI.mil, the DOD has done little to answer that so far. The Pentagon’s research and engineering undersecretariat, which owns the portfolio for GenAI.mil, did not provide interviews or answer detailed questions sent Wednesday from DefenseScoop by publication. 

“The rollout of GenAI.mil is a major step forward, but adoption has to happen hand-in-hand with discipline and transparency,” said Tyler Saltsman, an Army veteran and founder of EdgeRunner AI who has been involved in deploying AI systems for the military. 

“If we get this right, AI becomes a force multiplier,” he said. “If we get it wrong, we create a new class of vulnerabilities.”

Now what?

As service members and military officials logged into their government computers on Dec. 9, they were met by an uncanny pop-up message. 

Under a stylized GenAI.mil logo, a quote about how “victory belongs to those who embrace real innovation” and a link to a supposedly government website, sat the line “I want YOU to use AI.”

“For people in my office, when we first saw that pop-up, we wondered if our computers had been hacked and if this was a legitimate new software, or if this was maybe something nefarious by an adversary,” said one senior Army official. 

Defense Secretary Pete Hegseth posted a video on social media and sent out a memorandum to the force about GenAI.mil that day, quelling fears of a cyberattack, but doing little in the way of answering a slew of questions that began to permeate the force about the program. 

“I expect every member of the department to login, learn it, and incorporate it into your workflows immediately,” he wrote. “AI should be in your battle rhythm every single day. It should be your teammate.” 

His directive struck multiple personnel as too vague for what is shaping up to be one of the most consequential technologies of the 21st century, now being widely pushed across the military. 

Several officials who spoke to DefenseScoop said they knew that GenAI.mil was going to be released for months, but they didn’t expect it to be set in motion in the way that it was. Some suggested that those in the military AI space who focus on end user implementation efforts are “beyond upset” about the new genAI hub.

According to a defense official familiar with the department’s LLM implementation over the last few years, the new platform looks “like a collection of stovepiped commercial products that you lock yourself into — that is disastrous.”

And those who had never meaningfully used LLMs before are now being told by the DOD’s top chief to do so without clear guidance or forewarning.

“It has literally been shoved into our face,” said one senior enlisted leader in an Air Force line unit, referencing “I want YOU to use AI” posters of Hegseth pointing like Uncle Sam that have been photographed around the Pentagon. “It literally just showed up one day [in] an email saying ‘I want you to use it, this is the cutting edge yada yada yada’ while having zero [training or] addressing anybody’s concerns about this.”

Combined with previous skepticism about artificial intelligence, GenAI.mil’s forceful launch and disclarity around usage for military purposes, that senior enlisted leader decided not to use it for now. He also said that other troops he knows are wary of its efficacy. 

A military reservist DefenseScoop spoke to said that he was surprised at what little information he received about GenAI.mil from the top. He said that users should know that their queries may not be accurate or private — concerns several other personnel said they wanted better answers on.

Even those who have previously used LLMs for military purposes said the Pentagon should implement a training program, as it does with a vast majority of its equipment, systems or processes, especially if AI use ends up being truly mandatory for DOD personnel. 

“I don’t feel forced to use it,” said the Air Force leader. “Until they come out with a very specific left and right — this is how you’re going to use it, in what specific workflows — I’m not going to.”

Mixed reviews

In late 2022, Open AI launched an early version of ChatGPT as a free preview to the public. 

That chatbot quickly became a worldwide phenomenon. Other companies soon followed. The conversational technologies generated a lot of hype early on, particularly for their ability to de-bug software code, summarize complex documents, produce hyper-personalized videos and automate other historically time-consuming tasks. 

However, experts also warned about severe — known and unpredictable — threats that genAI poses to humanity, public trust and democratic processes. Some of those include the potential for LLMs to reflect or amplify biases found in their training data; disclose sensitive info to unauthorized parties; be manipulated to share data by malicious prompts; and cause humans to experience long-term cognitive changes.

DOD leaders urged the workforce against joining a movement that was growing at that time, in which some experts called for a pause on genAI so that they could first learn more about it and its possible consequences. The Pentagon argued that such a halt could let America’s adversaries get ahead in developing and using the technology.

In August 2023, Pentagon leaders under the Biden administration set up Task Force Lima within the Chief Digital and AI Office (CDAO) to quickly explore potential ramifications and promising use cases for safely integrating genAI. Lima issued guidance early on to direct all DOD components’ paths ahead with the disruptive technology. 

Broadly, the individual military services’ policies and approaches to genAI systems — including but not limited to NIPRGPT, CamoGPT and AskSage — have been fragmented and complicated by data ownership, acquisition issues and other complexities.  

The CDAO announced plans to sunset Task Force Lima in late 2024. But “AI dominance” remains a top priority for the U.S. government under the second Trump administration.

In summer 2025, the Pentagon awarded OpenAI, Anthropic, Google and xAI individual contracts, each worth up to $200 million, for access to some of the most advanced commercial AI capabilities, like large language models, agentic AI workflows, cloud-based infrastructure and more. 

Now, GenAI.mil is here, initially with Google’s products. Officials told DefenseScoop that the platform has a navigational panel that suggests xAI’s Grok, OpenAI’s ChatGPT and Anthropic’s Claude assets are coming soon. 

“This feels a lot different [compared to previous AI rollouts] because this one got so much more attention — and it also came with a department-level mandate that said ‘You will incorporate this into your daily workflows,’”  a senior Army official told DefenseScoop, noting other programs felt more optional.

He said that Gemini’s responses to his prompts read more like natural human language than some of the other models.

“I can tell you not all models that we’ve been given are created equal, and this is one of the better, if not the best, model we’ve received so far,” he noted. 

Not everyone feels the same way. Every LLM is built for certain purposes and comes with its own pros and cons or strengths and weaknesses.

“It is not that one model performs better or worse than another for military work,” a defense official said. “They are all good at different things. What we really focus on is the degree of editorializing and stylistic influence a model injects into its outputs.”

They also suggested that systems in the military’s existing LLM arsenal, like AskSage and CamoGPT, offer more security and applications than what’s currently available through GenAI.mil.

“People are furious that it brings less capability than NIPRGPT — which is being decommissioned — even had,” the defense official said.

Expanding the user base

While the DOD has already been using both unclassified or classified LLMs like CamoGPT or SIPRGPT, GenAI.mil’s expansive launch seems to have tapped into a different military user base. 

Multiple troops who serve in line units said that they or their peers were experimenting with GenAI.mil. Some had never used other DOD models before or had only used public versions for various administrative tasks, the latter point invoking concern from one official about possible data spillage to commercial models.

Example use cases DefenseScoop heard about included configuring awards packages, emails, memorandums, performance reviews and order writing. Some sources said that personnel had already been using commercial models for these tasks. But they noted how GenAI.mil so far allows users to input controller unclassified information, or CUI, and that since it was so heavily advertised, more people are aware of it in comparison to older DOD models. 

An infantry corporal, who thought GenAI.mill would be a “net positive” and had only used commercial models before, observed that some of his younger peers seemed more ready to embrace the model than their older leaders, likening them to “teachers telling you not to use Wikipedia whenever you’re citing your references.” He attributed this perspective to younger people being exposed to internet technologies for virtually their entire lives.

“I wrote out entire [standard operating procedures] with the AI aid and it covered shit that I didn’t even think of or consider,” he said, noting how he drafted an outline, prompted the LLM for specific content he wanted included and told it “exactly what I’m wanting it to achieve.” 

“I was able to aid my entire company by helping write out SOPs and providing feedback from the AI perspective to even my superiors,” he said. “I’ve gotten a lot of kudos for that.”

The corporal, like other military personnel, warned that these models are not always correct, and that it was crucial to proofread and fact-check whatever it spits out.

Google’s press release on the day of the Pentagon’s announcement said that its model is designed to supply the DOD workforce with “an edge” through natural language conversation and retrieval-augmented generation. 

RAG refers to a technique that essentially makes chatbots and other models more reliable, by enabling them to look up information from a specific set of relevant data sources before answering a question.

Prior Defense Department-supplied AI options rely on RAG datasets that are created locally, or specifically shared with individual users or user groups. This approach works well, according to a defense official, because it allows people to deliberately curate datasets that contain accurate information tailored to their specific purposes. 

The official suggested that Google’s offering in the platform has an internal dataset that appears to consist of cached, scraped web content that is enabled by default. Google reportedly places citation markers next to content pulled. 

“So functionally, Google has effectively included an internal dataset that — unless explicitly disabled — draws from web-based data sources. While it is likely curated to offer more authoritative sources, this is still catastrophic for accuracy in any context,” the defense official said. 

They noted that when DOD users work with critical military products that require verifiable sources, and in operations where accuracy is vital, they do not want LLMs to draw from uncontrolled data sources that aren’t already embedded in the trained and weighted model.

“Although there is a setting that allows this behavior to be turned off, no training whatsoever has been provided on the tool,” the defense official told DefenseScoop. “The average user has little to no experience with AI or LLMs, and they would not know to look for a setting to turn cached data off. Nor would they understand the implications of pulling data from a cached version of the web and how that practice can contaminate or degrade model outputs.”

A representative from Google said GenAI.mil employs multimodal RAG and linked to a document about the program’s index curation, but did not answer specific questions from DefenseScoop.

Trust issues

Before he was CEO and founder of EdgeRunner AI, Saltsman served as an Army officer and deployed to places where units operated under jamming, degraded comms, and near-peer surveillance.

Now, his company builds AI systems for real-world military settings where it can’t be assumed that cloud access, secure networks, or even steady power will be accessible.

“A big part of our job is ensuring the model’s behavior aligns with U.S. values, laws of armed conflict, and the specific operational cultures of each branch. There’s no one-size-fits-all military, and there’s no one-size-fits-all AI,” he noted.

Regarding GenAI.mil, Saltsman said the biggest theme he’s hearing across the services is uncertainty about what’s actually inside the model. 

“For most commercial generative AI systems, the training data is a black box. Operators don’t know who wrote the source material, whether it reflects U.S. values or adversarial influence, or whether it was curated for military accuracy. With warfighters, you can’t have that kind of ambiguity. Garbage in, garbage out — and in this environment, ‘garbage out’ can put Americans at risk,” he told DefenseScoop.

Saltsman is also hearing concerns about data leakage through prompts. In his view, even on a .mil domain, people naturally paste in too much context to get a better answer, such as planning details, internal emails, or tactics, techniques, and procedures. 

Several service members and officials who spoke to DefenseScoop echoed his alarm, specifically about where their inquiries were going, who was able to see them and whether their data was going to be used to inform the commercial models housed in GenAI.mil. 

Some said that without extremely clear guidance on retention and logging, this could present a new vector for unintentional exposure.

“There is also the trust issue,” Saltsman noted. “A model that speaks confidently can be wrong, hallucinate, or reflect values that aren’t ours. Warfighters are rightly cautious about taking advice from a system whose internal worldview they can’t inspect.”

Separately, the military reservist told DefenseScoop that — while the chatbot “feels like a friend” — users must recognize that its outputs are not always true or private.

In one section, there’s information about policies and terms of use, but in the reservist’s view, the onus is on the user to find resources and explanations about how to engage with the elements.

He pointed out that some reservists who he has worked with might have little to no exposure to generative AI, or a full grasp of the technology’s potential implications.

The reservist used Google’s initially available model on the GenAI.mil to assist in producing 60-day training plans. The outputs were fine, he said, but they didn’t include doctrinal terms and they presented some “AI slop” — or low-quality, generic digital content.

Although a lot of refinement and LLM training is needed, he still expects GenAI.mil offerings to increase efficiency in some of his drill activities down the line.

So far, many use cases for GenAI.mil that DefenseScoop was told about have revolved around administrative tasks. 

Some personnel said that they were anxious about troops not learning how to communicate or complete them on their own now that AI is an option. But they also worried about where the lines should be drawn on not using these models, especially because they are known for producing mistakes or “hallucinating” information.

Should they be used to create risk assessment worksheets, which help identify possible safety concerns for even the most mundane rifle range? What about using AI to help create weather analyses for drone flights? And what if the technology injects a mistake that causes an injury or equipment mishap or even death, but no one catches it? Who’s to blame?

“AI shouldn’t be the only thing that’s making the decisions on the strategic or operational or tactical level,” the infantry corporal said. “That human element, it definitely needs to be maintained.”

“For me, being lower enlisted, if I’m looking at the officers or the senior NCOs appointed above me and I start to see that behavior,” he said of leaders who would use AI to make decisions for them, “I would definitely not like that because that human element is also the thing that’s going to help me get home to my wife and kid in wartime.”

Updated on Dec. 18, 2025, at 6:30 PM: This story has been updated to note that a representative from Google said GenAI.mil employs multimodal RAG and linked to a document about the program’s index curation, but did not answer specific questions from DefenseScoop.