Using Language Models as an Analytical Writing Partner

Partner, not a panacea

Even the most well-read among us has probably only consumed a fraction of what is available on most topics. This is why we favor discourse over simple writing, because engaging with other writers and thinkers can help pull in their knowledge to supplement our own and truly move the conversation forward.

A large language model (LLM) can play an important role in this by speeding up the process and bringing in larger data than any writer, or group of writers, can possibly recall or parse.  Furthermore, now that “everything” has been digitized, how confident can we be that “everything” is actually all-encompassing versus a collection of what just happens to be easily digestible or cataloged? An LLM offers the opportunity to sift through and synthesize all of this data, including pulling in sources that might otherwise be inaccessible, and help transform ideas into workable discursive documents.

Given the Chief of Staff of the Army’s recent recognition of an article on modern coastal artillery, I asked for an LLM’s help to build on this idea. The key takeaways from this experiment include:

• Using language models is a skill and requires iteration

• Language models help most with reviewing and synthesizing literature

• Language models cannot produce publication quality essays on novel topics

While not a systematic experiment, this post explores the technology’s strengths and weaknesses to get the best out of language models. 

Initial exploration

Using Google’s Gemini, I started with a straightforward prompt that indicated specificity to military topics. 

Using these sources, [list of URLs] can you summarize and analyze how past capabilities related to coast artillery can inform how a multi-domain task force can use precision fires, integrated when possible with other capabilities, to pursue an area denial strategy?

I purposely picked a range of sources from across several decades to assess how well Gemini could sift through the sources. Gemini framed the problem thusly:

Absolutely, I can help you analyze the articles to see how past coast artillery capabilities can inform multi-domain task forces using precision fires for area denial. Here are some examples of what the model produced:

Coast Artillery in the Past (Articles 1, 2, 4, 5)

Coast artillery's primary function was to defend shores from naval attacks using large, immobile guns (Articles 1, 4, 5).

These guns lacked precision compared to modern weaponry but could inflict significant damage on approaching ships (Articles 1, 2).

Their effectiveness waned due to advancements in aircraft and missiles (Articles 2, 5).

Precision Fires and Multi-Domain Task Forces (Articles 3, 7, 8)

Multi-domain task forces combine air, land, sea, space, and cyber capabilities for integrated operations (Article 3).

Precision fires involve using guided missiles and munitions for accurate strikes (Articles 3, 7).

This approach minimizes collateral damage and allows for more strategic attacks (Articles 3, 7).

Also, because it was asked to conduct “analysis” rather than simple summary, it continued on to provide some caveats and clarifications to this initial analysis:

Coast artillery's area denial strategy can be adapted using modern weapon systems.

Precision fires offer greater accuracy and flexibility compared to fixed guns.

Multi-domain task forces allow for a more comprehensive approach to area denial.

Limitations

The articles don't discuss the cost or feasibility of implementing a large-scale coastal defense system with precision fires.

The effectiveness of such a system against modern weaponry remains debatable (Articles 6, 9).

None of this, at first glance, is incorrect, but neither is it particularly specific or compelling. A close reading of the text shows that it immediately invokes a number of key terms, especially those given in the prompt, but also those it noted were common in the provided texts. Gemini itself lacks special expertise, but adopts my framework to answer the question. 

This is where using AI becomes a skill. The model outpaces our ability to gather and analyze data, but it does not necessarily know what the data means in context, nor can it embrace the nuance behind doctrinal terminology when those terms are specific to a context divorced from larger, general purpose language sets. This is why while nothing in the LLM’s output is technically incorrect, it lacks prescriptive specificity or original thought. In short, the language models cannot introduce new concepts or engage in ways beyond what the data already can or has done. 

Iterating 

To build connections in new directions, I iterated with the tool. In this case I provided it a new, more specific source and asked it to narrow its previous discussion to one capability in particular, cyber:

Can you expand on "Connecting past to the present" by connecting this article specifically to using cyber as an area denial capability?

Response:

Sure, let's see how the concept of using cyber capabilities for area denial connects to the past coast artillery and multi-domain task forces discussed previously.

While the articles we reviewed (1, 2, 4, 5) focused on physical defenses with coast artillery, the concept of area denial using firepower transcends those specific weapons. Here's how it applies in the context of multi-domain task forces and cyber capabilities, based on the new article you provided:

Multi-Domain Operations and Integrated Effects

The article you linked emphasizes that successful multi-domain operations rely on integrating capabilities across different domains (air, land, sea, space, and cyber).

Just like using a combination of artillery fire, radio jamming, and radar hacking (as described in the article), cyber capabilities can be a powerful tool for area denial when employed strategically within a multi-domain framework.

Cyber Capabilities for Area Denial

Here are some ways cyber capabilities can be used for area denial within a multi-domain task force:

Disrupting Enemy Communications and Command & Control: Cyberattacks can cripple an enemy's ability to communicate and coordinate their forces, hindering their operations in a designated area.

Disabling Critical Infrastructure: Cyber attacks can target power grids, communication networks, and other essential infrastructure, disrupting an enemy's ability to function effectively within a specific area.

Deception and Misinformation: Cyber operations can manipulate information systems to confuse and mislead enemies, making it difficult for them to operate effectively in a particular region.

Key Advantages

Precision and Selectivity: Cyberattacks can be more precise than traditional forms of area denial (like shelling a coastline) and can target specific systems or infrastructure.

Denial and Disruption: Cyberattacks can disrupt or deny an enemy's access to critical resources and information within a designated area.

Non-Kinetic Effects: Cyberattacks can achieve area denial effects without causing physical destruction or casualties.

Challenges and Considerations

Attribution: Attributing cyberattacks to a specific source can be difficult, potentially leading to unintended consequences.

Escalation Risks: Large-scale cyberattacks could escalate tensions and lead to broader conflict.

Defense Measures: Adversaries are constantly developing defenses against cyberattacks, so their effectiveness can vary.

Connecting to Coast Artillery

While the tools and methods differ vastly, the underlying goal of hindering enemy movement or operations within a specific area remains the same.

In the past, coast artillery batteries were used to physically block enemy ships from entering a harbor. Today, cyber capabilities can be used to disrupt an enemy's ability to communicate, navigate, or coordinate operations in a designated maritime area.

Conclusion

By integrating cyber capabilities with other domains in a multi-domain task force, it's possible to achieve area denial in a more sophisticated and targeted way compared to traditional methods like coast artillery. However, cyber warfare comes with its own challenges and considerations.

Finally, to see if I could close the loop on this, I asked one more question:

Can you recommend an article already written on this topic I can engage with to explore this idea more fully?

Response:

Yes, absolutely! Here are a few resources that explore the use of cyber capabilities within a multi-domain area denial strategy:

Preparing for Multidomain Warfare: Lessons from Space/Cyber Operations - Air University (https://www.airuniversity.af.edu/Portals/10/ASPJ/journals/Volume-32_Issue-3/V-Harris.pdf) - While the focus is on space/cyber, this article provides excellent insights into how cyber capabilities can be leveraged to degrade an adversary's ability to operate and the importance of integration in a multi-domain world.

Two explanations for the paucity of cyber-military, cross-domain operations | Journal of Cybersecurity | Oxford Academic (https://academic.oup.com/cybersecurity/article/8/1/tyac002/6610005) - This article analyzes the reasons behind the relatively limited examples of integrated cyber-military operations. It explores the technical and operational challenges, providing a valuable perspective when considering the practical implementation of cyber in a multi-domain strategy.

Strategic Anti-Access/Area denial in cyberspace - CCDCOE (https://ccdcoe.org/uploads/2018/10/Art-11-Strategic-Anti-Access-Area-Denial-in-Cyberspace.pdf) -  Focuses specifically on the concept of area denial in cyberspace. Although slightly dated, it provides an understanding of how cyber operations can hinder an enemy's freedom of movement and action.”

The previous response gets much closer to what we would consider a stand-alone essay. The language model explores far more specific options at the conceptual level, though it still falls short in terms of describing specific platforms, units and their functions, or even historical or geopolitical context, which an Army officer writing an essay would almost instinctually include while illustrating their argument. This last one brings us back from the abstract to how practitioners are already working on this problem, while maintaining awareness of how the conversation started.

Using this LLM to work through the initial phases of brainstorming moves us from a very broad topic (comparing coast artillery capabilities to modern MDO in a defensive context) to a more specific one (how cyber specifically could support MDO in this context) very quickly. Furthermore, once we have that specific area of interest, we can rapidly solve the conundrum of “is this an original idea” and make sure we are engaging with scholarship already done in the area. It is still up to a professional to find their angle within this conversation, and more fully pursue it within the context of that scholarship, i.e. engaging in the discourse rather than writing stand-alone content in a vacuum.

LLMs as a partner, not a panacea 

This exercise shows how AI can largely only replicate things on which we have data, while new ideas still require input and critical thinking. No field is perhaps more complex than military science, and so becoming effective AI operators is essential to ensuring this emerging field works with us rather than parallel to our at odds with what we seek to achieve. Embracing LLMs as writing partners rather than writers themselves is one way to do this.