AI for Grunts: How might a Rifle company use AI specifically? (Part II)
by Major Ryan Shannon
Editor’s Note: This is the second article of a two-part submission on AI and its impact on the Rifle Company. It is the third installment in our series about emerging technology. Part I provided a relevant and focused background on AI and Data for small unit leaders. You can read our previous articles on data sciences here and here.
Author’s Note: Technology moves quickly. The bulk of the work on this article was completed in August 2024. Since then, the Army has released a large language model AI program available on NIPR and SIPR called CamoGPT. The Air Force has released another AI program called NIPRGPT. Both are basically ChatGPT that understands military acronyms. Check out this MARADMIN about how to get one of these AI programs on your work computers.
In Part I, we established a foundational understanding of AI and its relevance to the rifle company. We broke AI down into three practical categories: processing text and images, analyzing big data, and enabling novel software solutions. The key takeaway was that while small-unit leaders won’t be developing AI themselves, they will increasingly encounter AI-powered tools that automate tasks, accelerate decision-making, and streamline operations. The challenge is not just understanding what AI is, but recognizing how to apply it effectively at the tactical level.
Now, in Part II, we shift from theory to action. This article will explore specific ways AI can support planning, training, logistics, and combat operations. From leveraging AI-enhanced data analytics to improve readiness to using automated tools to speed up mission planning, we’ll examine practical applications that can make your unit more effective. Whether in garrison or deployed environments, AI has the potential to enhance efficiency and lethality—if leaders understand how to integrate it. Let’s dive in.
Like the first two paragraphs of this intro you just read? ChatGPT wrote it after being fed Part I and instructions to summarize and transition to a second article about practical application.
Admin
AI can make you much, much faster at adminis-trivia. Much of this is because of large language models like ChatGPT or Bard. It is like having a personal secretary who happens to have a PhD in most everything. These models can summarize things, pull out points relevant to a topic, make a first draft, or do a detailed proof-read. If you’ve used ChatGPT, you’ll have an idea of the potential. Think of all the time your XO and 1stSgt will get back if they don’t have to “work up a draft” for you.
What You Can Do Now
Get a rough draft version of practically anything in almost no time in almost any style you want: Need to write something and can’t figure out where to start? Ask a large language model to make a first draft, then use your time editing the text instead of coming up with new things. If you give ChatGPT an example to use a style and some bullet points, it can write awards, policy letters, pretty much anything. If you want to see an example, below is a command philosophy written by Chat GPT in less than 3 seconds from a single question prompt.
Get a good refresher of any information the algorithm was trained on: Trying to quickly refresh a topic? One option is to go to Google and dig through a bunch of different websites and forums. Another option is to just ask ChatGPT or whatever large language program you're using. If the information was part of the data that the algorithm was trained on, it will be able to give you something useful.
Summarize long documents: Sometimes you may not have time to dig into that 50 page read ahead someone sent you. A quick way to get the main points is to ask a large language model for a summary. Some programs even allow you to attach a word or pdf document. This can be a huge time saver or quick way to get a short outline of the document for reference in a meeting.
Revise individual documents: You can tell large language models to rewrite a document you have created with corrections. This includes not only spelling or grammar but even the tone or style of the document.
Align written products: Want to make sure your policy is in line with HHQ? Put both policies into a large language model and ask it to look for similarities or differences. AI can do much of the work required to scrub a body of documents to align them, or at least identify what you need to fix.
Fix grammar: Most large language models will be able to give suggestions and tips about grammar. For something even more comprehensive, check out a program like Grammarly: Free AI Writing Assistance. Grammarly and programs like it reads what you write, helps edit, suggests revisions, and can even do some stuff from scratch.
What Could Be Coming
Chat GPT like AI trained with USMC knowledge- The large language models that currently exist were trained off a broad range of data from the internet. This means they provide generally good answers to anything but struggle with domain specific knowledge or jargon. However, it is possible to update models with domain specific knowledge. People are already working towards this. Expect future DOD or USMC specific AI models that will know what you are talking about when you use an acronym like PICMDEEP.
No more wasted time figuring out hitlists: Have you ever felt like you spend more time trying to diagnose problems in your company rather than address the problems in your company? AI will likely solve that. AI runs on data, but that data needs to be well-structured. One of the secondary effects of AI is that the Marine Corps has begun attempts to structure the data that we already have. To use an analogy, our databanks were like storage rooms filled with unlabeled boxes of documents. There is great information there, but it is very hard to find. As the DOD and the USMC progress with AI initiatives, expect tools that consolidate hitlists, reports, and ankle-biter admin tasks into coherent, easy-to-use dashboards. This already is happening at 3d MARDIV, where some savvy coders have used Microsoft Power BI to consolidate and visualize almost every USMC database related to non-MOS-training unit readiness. XOs for units in 3d MARDIV say it has turned checking up on the admin health of your unit from an all-day affair to five-minute event. Efforts like these mean that in the future, unit commanders can focus fixing the problem (remediating hitlists, doing maintenance, approving DTS, etc) rather than figuring out what the problem even is (pulling a dozen hitlists from shops, trying to make sense of GCSS, wrestling with DTS approvals, etc).
Drawbacks and Dangers
Hallucinating: AI programs will be totally wrong with utter confidence, which can be really alarming for a user. The computer isn’t lying or being malicious, even though it feels that way. It happens because the way the algorithm works. AI gives you a response that has a high probability of being coherent based off prior training data. Coherent does not necessarily mean true or advisable. You should not mix glue into your pizza as recommended by Google's AI search.
This means you can’t totally trust AI to give you a correct answer. Here’s a Marine related example of a hallucination where ChatGPT utterly fails at parsing though a publication.
De-valuing written word: Some documents deserve to have human attention. Do you want to be the guy who has his FITREP written by AI? Would you feel right writing on someone by going to ChatGPT and writing “based on this example of a performer in the middle third of their peers, please generate a new FITREP comments for a middle tier performer based the below copy and pasted achievements.” Additionally, sometimes you can tell something was written by AI. What do AI generated comments say to the board or the Marine Reported on?
C2
Much of the impact of AI will be in the cognitive domain. AI can process lots of inputs at near real time or real time, displaying that information in ways that humans can quickly absorb and providing analysis.
What You Can Do Now
Use Tactical Assault Kit (TAK) software: Use TAK. Ever wonder why we don't have a web-based application like something on your phone for command and control? Ever wonder why the only place where you see a heads-up display that tracks blue movements shows red movements and you can drop tactical control measures on to is in the screen corner on the first-person shooter games that you play? Well, the DoD has had it for more than a decade, but the Marine Corps hadn’t bought off on it for whatever reason. You may love KILSWITCH or C2PC, but TAK is what everyone else uses, to include non-military partners like the State Department. It is great for individual situational awareness but the real power of it comes from linking it up to a server. When it is connected to a server you can message other users, share locations, drop TCMs for everyone to see, link into drone feeds, and a ton of other functionality.
What Could Be Coming
Battletracking: As the Marine Corps figures out what is next after C2PC, expect AI and good software to be able to tie together all sorts of different information streams more effectively. It also could do some of that processing with minimal human involvement. Algorithms will be able to process multiple drone and video feeds simultaneously, identifying individual personnel or types of military equipment. It is not a large leap to be able to put that onto a common operational picture server (like the TAK server) that feeds into your end user device.
Targeting Data: If you can track it, you can locate it, you can kill it. You potentially can automate processes so much that no human is needed. Algorithms monitoring drone feeds identify a military vehicle, compare the location data to known friendly locations, determine the vehicle is enemy, and send location data to a firing unit. Analog backups are important, but we will have to become comfortable with our Fire Support Officers working in digital space. We also will probably have to make some determinations if we want to be “in-the-loop” (no munitions launched unless a human explicitly approves) or “on-the-loop” (a human is monitoring and can intervene but otherwise the process will just happen). “In-the-loop” or “on-the-loop” determinations might become just another part of the order, like how a commander may set “passive approval.”
Web-based applications for command and control: Web-basing applications (i.e. running the program on the internet, not on the user device) could immensely reduce the cost to access, run, and contribute to command-and-control systems. Basically, it would be remoting everything except your end-user device and whatever allows you to connect to the internet. Between less gear and less support personnel at the tactical edge, that’s a big reduction in signature. “Getting comms up” in the future could be less about setting up radio banks and more about getting a secure, low-probability of detection connection to the internet.
Tasks to seize physical components of digital infrastructure: Future tasks at the operational and strategic levels to disrupt enemy C2 could include targeting physical infrastructure. As governments and states begin to rely more and more on things in the digital world, targeting the physical infrastructure could become part of mission sets. This may overlap with tasks that already make good military sense, like seizing a power plant. But we might not be far from a world where infantry rifle companies receive tasks to go find and destroy server farms, as part of "degrading the enemy network and cognitive capability." In the future, seizing cloud infrastructure may be as important as seizing a bridge, a rail, head, some mountain pass. This is even more likely if you consider how these AI models are created. As discussed earlier in this article, it is costly and hardware intensive to train the initial model, but once it's done, it is just a piece of software. Although existing software models can be still tweaked, by destroying hardware it is possible to prevent enemy from creating new AI.
Drawbacks
Algorithmic “Good Idea Fairies”- As data and data processing gets integrated into different levels of command, people are going to be able to perform analysis and make decisions off those analysis. That could be a great thing, but it also to lead you down some rabbit holes. It's conceivable that in the future you won't have to react just to your command’s good ideas but also the good ideas generated by algorithms and data. How Palantir Is Using AI in Ukraine | Angry Planet on Acast
Drones
No conversation about AI can be complete without including something about drones. AI is helping make drones more effective and simpler to use by automating any process not directly related to your purpose for using the drone. Oddly enough though, drones might not reduce manpower requirements or make warfare any less physical. Details below, but suffice to say even in the age of drones you will likely need a shovel for digging-in and the ability to hike with a heavy pack.
What You Can Do Now
Be an end-user: If the Marine Corps gives them out, use them and develop effective tactics. It may be frustrating that you can go buy something off the shelf that seems to work well, but it is probably best to accept that the battle to get a non program-of-record drone approved is not the best use of your time.
Be ready to deal with them: Drones will be part of the future battlefield, and we need to be ready for it. The Connecting File has already written a lot about this. If you prefer video to the written format, here is a good one dealing with drones and other topics in Ukraine: Combat Vets from Ukraine Explain Drone Warfare, Trench Warfare and More (youtube.com)
What Could Be Coming
Lots and lots of drones: The potential vision for drones in warfare is a huge topic to tackle. There have literally been dozens of books written on it. Some themes are as follows: Highly capable individual drones; drone swarms which act in coordination with no human input to individual members off the swarm; ground drones that will act in support or direct combat roles; drones being treated as expendable munitions rather than something requiring recovery. If you want to understand more, fiction about the near future is probably better than technical briefs for seeing the vision. There actually is a whole genre called “FICINT” around this idea of plausible fiction grounded in current research. For drones, I would recommend Burn-In by the same guys who wrote “Ghost Fleet” or White Sun War by Australian Major General (retired) Mick Ryan as starters. For extra motivation, the hero characters in “White Sun War” are members of a Marine Littoral Regiment on Taiwan during a war with China.
Drawbacks/Dangers
Drones requiring the opposite of fewer people: It's possible that drones will do the opposite of reducing manpower requirements. All that gear needs to be maintained. Someone needs to be on site to troubleshoot problems. What about the people to maintain the new vehicles required to get your drone operators places with the right quantity of gear? And drone operators? Well, they need rest plans too so maybe flying one set of drones may require multiple sets of people. And this is all before you get into any of the maintenance and troubleshooting required for the software. Drone warfare increasing rather than decreasing manpower requirements is already happening in Ukraine. A great breakdown of the manpower currently required to make drones work in Ukraine starts around 22:35 here: Mike Kofman and Rob Lee on Drones in Ukraine - War on the Rocks
You’ll still have to walk: Setting up a base station, making sure you aren’t making a signature that can be targeted, infiltrating to a point where your drones’ ranges are effective are going to require folks with strong backs and hard feet. The idea of the cushiness of being a drone operator will evaporate once people realize the only way to get into proper employment positions is via a foot march with all those batteries. Particularly for smaller drones, it is more likely an operator is going to be in a hide site somewhere close to the front line rather than a trailer in Arizona. Drones will not eliminate the need for people who can operate in austere environments, know basic soldiering, and are proficient in fieldcraft.
Immediate Actions for drones: Troubleshooting drones won’t be as simple as tap-rack-bang. Troubleshooting is going to take knowledge of both hardware and software. Guys will have to be able to replace a busted rotor blade AND remember to turn off the location data after rebooting that commercial off the shelf system (many COTS drones transmit location data for the drone and the drone command station… easy way to get killed). Figuring out how to make drones work and keep them working under combat conditions is going to be a challenge.
Figuring out how to train this stuff: Initiating the company attack with a drone swarm that locates and suppresses a trench by dropping munitions in top of it? Cool tactic. Awful day for Range Control. And that’s just one example. Range control is going to need to figure a lot of stuff out, probably along with the FAA and the wing.
Planning and Training
AI can make you faster at making products, gaining insights, or developing options both in combat or training. For developing options, think of how in the Marvel movies Tony Stark can ask JARVIS for some analysis or ideas on how to approach a situation. For faster product generation, many of the AI-driven techniques will be similar to what will make you faster at admin. The AI does not tell the person what to do but it can give some starting suggestions. As the Marine Corps gets better at collecting and analyzing large data sources, training design will also be impacted.
What You Can Do Now
Use AI to write Orders/briefs: You can use large language model programs to draft up things for you. There even is specific software for military briefs. Check out Onebrief. Disclaimer, the author has not used Onebrief operationally or gone through the process of acquiring it for a unit.
Terrain analysis: Ever wish you could get Google Map-style directions for your covered and concealed route during a helo assault? We already have it. Recent Naval Postgraduate School graduate Capt Ryan Helm made a TAK plug-in called Augmented Reconnaissance and Estimate of the Situation (ARES) that can give you recommendations of for landing zones and infiltration routes based on map data and observation from known enemy positions. 1/5 just tested it out and found it to be highly useful. 1/5's ARES AAR with can be found here. The AAR has all the links and documentation for installing and managing ARES after you get devices with TAK.
Get AAR points: Give a large language model like Chat GPT a bunch of AARs and tell it to summarize key points that recur frequently. This could be a huge time saver as well as allow you to incorporate as many AARs as possible in your research.
Be a good data source: If you end up part of a study or initial data collection experiment, provide good data. That data might be used to drive analysis that changes the way we do things.
What Could Be Coming
More TAK plug-ins: TAK plug-ins like ARES have huge potential to help future tactical level planning. As you can use your phone to plan a sightseeing trip to a distant location, a user with a TAK and plug-in enabled phone could do much of their planning from their device. Dissemination tools built into TAK could become the new age version of disseminating a map overlay. The Marine Corps Software Factory is taking lead on this. Expect planning tools that can help with fires, logistics, and all sorts of things.
Data optimized training techniques/packages: Expect future training packages and programs to be informed by data. It's already happening. The new rifle range and advanced marksmanship training programs? Decided on after the Marine Corps created a virtual simulation pulled from data on hundreds of shooters doing the old program and the new program. The virtual versions of the Marines had shootouts between their “old program” shooting skills and the “new program” shooting skills. Run that virtual battle several thousand times and see who wins the majority of the time. Don’t take my word for it? Check out Brutecast S5 E15 - Advanced Marksmanship Training Program
Simplified understanding of specialized skills: Expect AI to enable people with minimal specialized training to learn or understand specialized skills. Much like Google lowered the barrier for locating knowledge, AI is going to make it much easier for people without specialized knowledge to process and understand specialized things. It's not that the AI tools are giving you new knowledge, it is that they are synthesizing the knowledge that's already out there and presenting it to you in ways that are easier to follow. A good large language model that is trained on data from the field you are trying to learn about is like having a personal assistant who has a PhD in the field. Don’t remember how to make a field expedient claymore? Ask a future large language model trained on EOD documents. It would be like texting your own personal EOD team that has nothing to do but answer all your questions.
Drawbacks
Security: Putting CUI materials into language models that can track and store information? Might not be good. The Marine Corps and Department of Defense as a whole will need to work though how to use AI in a secure manner. Congress has already had issues with this, requiring new rules to prevent staffers from putting sensitive information into the free version of ChatGPT.
Remembering AI isn’t omniscient: To quote a Professor here at NPS “When people observe a A.I. system performing a task with near-human competency, they often assume that the A.I. system and the neural network underlying it perceive the environment and make decisions as humans do. Unfortunately, neural networks work very differently than human perception and reasoning, which can make it difficult to anticipate their failures.” As we get more and more electronic tools, we will have to remind ourselves that they are simply planning aids. Human judgment will still be essential. People will have to be careful not to get complacent and abandon their duty to personally engage with proper planning.
Reconnaissance and Security
This section is included to give a sense of what is possible in the future. Much of this section has to do with future AI being able to identify individuals or vehicles from video or sensor feeds. Algorithms being able to reliably identify things will be the foundation for a whole bunch of practical applications.
What You Can Do Now
Be ready to be an end user: Honestly, there is probably not much you can do until the Marine Corps rolls out some programs.
What Could Be Coming
Automated processing of video/drone/satellite feeds: Once AI can identify objects in sensor feeds, your ability to sense will be limited only by the number of sensors you have and your computing power. The sensor screen of the future could consist of hundreds of drones, cameras, or feeds. That much incoming data would be too much for a human to monitor but not too much for AI. Humans can focus on the feeds where AI has identified potential disturbances or enemy units. AI could even stitch different video feeds together, allowing for a human to simply designate what enemy unit they want to follow and allowing the AI figure out how to cue video feeds from different assets.\
Highly automated weapons: If AI can be used to identify things, it is a short leap to targeting the thing. Future AI monitoring systems could be capable of launching some type of weapon against a target identified by an algorithm. This could be done through command and control systems tied to both sensors and shooters. It also could be done through a properly programmed loitering munition. In the future, it might not just be mines and obstacle belts that delay you. It might also be a swarm of loitering munitions set to analyze, identify, and then target anything that enters their operation box. Imagine firing several dozen loitering munitions behind you to cover your withdrawal or to set up a defensive position where a flank is screened by a loitering munition swarm.
AI spoofing/deception: If we can implant false indicators into enemy sensors, you potentially now have ways of sending false signals or masking your own signal. Future deception efforts may involve measures in the cyber realm or built into your physical camouflage meant to fool algorithms into ignoring your unit, treating your unit as friendly, or mis-identifying their own equipment as targets.
Drawbacks
War without fog is still hard: Eliminating the fog of war does not mean that you are going to win. It is very possible to have a “tech stalemate” where everyone can see everyone… but you can’t solve the problem. That might be happening in Ukraine right now, as noted by Ukrainian General Valery Zaluzhny late last year. Have you ever played the board game Risk or Axis and Allies? Those are examples of games where every player can see what every other player does. Being able to see the other player’s move might just mean getting to watch them move in to demolish you.
Camouflage from the eye AND the algorithm: Camouflage might look different in an AI world. Wearing the right clothes could “fool” the algorithm. Unfortunately, stuff that can defeat algorithms probably doesn't go well with natural camouflage. Already companies exist that sell clothing that prevents facial recognition software from identifying you. Below is a hoodie that makes an algorithm think you are a giraffe from the company Cap_able (capable.design). It’s not going to fool a human.
Maintenance
This section is also included for the sake of future vision. Maintenance is one of the big areas that AI is being applied in the private sector. It likely is going to transfer to the military as well.
What you can do now
Be ready to be an end user: Honestly, there is probably not much you can do until the Marine Corps rolls out some programs.
What could be coming
Predictive maintenance- Know how the NFL can pull players from practice on Wed because they are at an increased risk of injury due to impacts from last week’s games and practice on Monday and Tuesday? We might be able to do the same thing with our vehicles. Sensors that can track wear and tear combined with data from similar vehicles across the Marine Corps could allow us to prevent problems rather than fix them. Replacing a part that has a 95% chance of wearing out in the next 200 miles but clearly isn’t broken now may feel wasteful at the individual level, but at an organizational level it will mean more serviceable and available gear.
Drawbacks and dangers
Software maintenance: Like how we maintain hardware we probably are going to need to maintain our software. This could add a whole new dimension to maintenance stand-downs.
Bad data, bad results: Predictive maintenance works by using data. So, the 15 years of a maintainer plugging in 9999 or -200 for miles driven might produce some wacky results. If a good algorithm has been fed bad data, is it the algorithm’s fault if it told you to change an MRZR’s tires 10 times in a single week? Unless our data entry gets cleaned up, bad data might send you on predictive maintenance wild goose chases.
Conclusion
Right now, AI can be of use to you, mostly through time-savings related to anything text based. What you can do now probably won’t feel that futuristic. We aren’t quite yet to the point where you and your Marines charge over a hill followed by a swarm of autonomous warbot killer drones. The best use of a rifle company’s time is still becoming brilliant in the basics of fieldcraft and warfighting.
However, we are potentially on the cusp of big changes to the cognitive aspect of war. Used correctly, AI should make the military faster at processing information and enable human decision making. AI is going to be about rapidly generating options or automating things so that the human can move on to the next task or just supervise.
What does this man-machine teaming look like? Think when Tony Stark asks JARVIS to run some calculations, doesn't like the results and tells JARVIS to run it again with new assumptions. Think the commander on the bridge of a spaceship asking a computer for three different routes through an asteroid belt and then choosing the one. Think instead of telling every drone where to go, you tell the swarm a box to remain in and give them a mission type order.
As we adopt these technologies, we need to be able to understand limitations. AI complements human rationality, it does not mimic or replace it. Hopefully, this article helps understanding at the Rifle Company level in some small way.
Maj Ryan Shannon is an Infantry Officer currently serving as 3d Reconnaissance Battalion Operations Officer. He is a graduate of Naval Postgraduate School where he earned a Master of Science in Operations Research. He can be reached at ryan.a.shannon.mil@usmc.mil.