How to use Deep Research to examine President Trump's federalization of the National Guard

A lot of people are unaware that modern AIs can examine legal issues of considerable complexity and produce long-form outputs and derivative products such as websites, quizzes, flashcards, simulated legal briefs, and much more. This entry shows what "Deep Research" from Google Gemini can accomplish. The particular task is to explore the legal controversy over President Trump's federalization of the National Guard and the use of the Marines following sometimes violent protests in Los Angeles over aggressive enforcement of immigration laws by the federal government. I am going to walk you through the process step by step.

Step 1: Getting the prompt right

First, I had my own ideas of what I wanted. This is one part of the collaboration between myself and AI in which I was in full control. Here was my rather crude query that I gave to ChatGPT o3. I wanted a "thinking model" and turned on web search because moving from crude query to something more refined is, in this case, a complex matter.

Here is a draft query arising out of recent legal challenges to President Trump's federalization of the California national guard and use of the marines to address immigration protests in Los Angeles. Read it and see if there are things I might add or clarify. %%% I want you to write a 4000 word blog entry explaining this lawsuit to an audience of upper level law students. Provide basic citations and, if possible, links to publicly available versions of cases and statutes. The tone should be objective, professional, and direct. Assume the law students have taken a basic constitutional law class and have some familiarity with the political question doctrine, the tenth amendment, the first amendment, and executive immunity. Your article should provide the historical and statutory context for federal use of force within the united states (both the national guard and regular branches of the armed forces such as the marines). It should briefly review what happened with protests in Los Angeles and elsewhere in California. You should then analyze the theories under which Newsom proceeded and how the district court reacted to each theory. What was the overall structure of the opinion. You will also note that the Ninth Circuit issues an administrative stay of the injunction. Why might it have done so. What do you see as the likely outcome on appeal to the Ninth Circuit and, perhaps ultimately, the Supreme Court. Pay particular attention to the political question doctrine and the ability of courts to interpret things like the insurrection act. I would remind the reader of the Zivotofsky I decision. Your response should briefly mention ties between the outcome of this case and cases under the Alien Enemies Act in terms of presidential power to find invasions or rebellions. It should discuss implications on protest and political debate depending on how the case comes out.

o3 came back with a decent suggestion for how to proceed but in reading its proposal, I realized there were some issues (such as the assertion that it would create "footnotes"), so I then provided some additional guidance. I spent time here because optimizing the prompt is crucial for getting good results from a complex query. And I don't want to waste time, tokens and electricity on a response that doesn't really meet my needs.

Can most AIs really produce classical footnotes? I don't think so. They can do endnotes. Also, maybe your should provide more of an overview of what we want before launching into this very detailed specification. Finally, I am thinking that 4000 words may be too short for our main product. I'd like to go to 8000 words. We can then create smaller summaries from the longer product. Could you rewrite the prompt in light of these suggestions.

After some thinking, o3 came back with the following comment and a revised prompt.

Generative-AI systems are excellent at producing long-form prose but still struggle with inline Bluebook-style footnotes; current research and user testing show that they mis-handle small caps, supra/infra cross-references, and page-specific pincites (sites.usc.edu, directory.lawnext.com). Because endnotes avoid those formatting pitfalls while remaining perfectly serviceable for web publication—indeed, The Bluebook itself contemplates endnote citation in many contexts (tarlton.law.utexas.edu, legalbluebook.com)—the revised prompt below tells the model to generate numbered endnotes collected in a single section after the conclusion. At your request it also (i) opens with a concise overview that orients the writer before plunging into particulars and (ii) doubles the target length to roughly 8 000 words so that you may later excerpt shorter pieces. All other structural, doctrinal, and stylistic safeguards remain, but word-count allocations have been scaled accordingly.

Revised Author-Prompt

Overview (≈ 250 words, not counted toward the 8 000-word main text)

Purpose. Produce an ~8 000-word, law-review-quality blog article that dissects Newsom v. Trump, the pending challenge to President Trump’s June 2025 federalization of the California National Guard and deployment of U.S. Marines in response to immigration-related protests in Los Angeles.
Audience. Upper-level law students who have completed Constitutional Law I; they know the basics of the political-question doctrine, the Tenth Amendment, the First Amendment, Youngstown’s tripartite framework, and executive immunity.
Value-Add. The piece must supply deep statutory and historical context (Insurrection Act, Posse Comitatus Act, prior domestic troop deployments), walk through the district court’s TRO/preliminary-injunction analysis, unpack the Ninth Circuit’s administrative stay, and forecast likely outcomes—anchoring every move in precedent and policy. End with concrete implications for protest rights and separation-of-powers doctrine.

Global Requirements

• Length. ~8 000 words (± 500), exclusive of endnotes.
• Citation format. Bluebook-compliant endnotes numbered sequentially and placed after the conclusion. Hyperlink each primary authority to a free source (LII, govinfo, Justia).
• Headings. Use markdown: “##” for major parts, “###” for subparts. Boldface only in headings (user preference).
• Sources to be discussed (non-exhaustive).
  • Statutes: Insurrection Act, 10 U.S.C. §§ 251-255 (uscode.house.gov, law.cornell.edu); Posse Comitatus Act, 18 U.S.C. § 1385 (law.cornell.edu, en.wikipedia.org); Alien Enemies Act, 50 U.S.C. § 21 (law.cornell.edu).
  • Supreme Court cases: Martin v. Mott (supreme.justia.com, tile.loc.gov); Sterling v. Constantin (supreme.justia.com, caselaw.findlaw.com); Ex parte Milligan (oyez.org, en.wikipedia.org); Youngstown Sheet & Tube; Duncan v. Kahanamoku; Baker v. Carr (six-factor test) (law.cornell.edu); Winter v. NRDC (preliminary-injunction standard) (supreme.justia.com); Zivotofsky I (supreme.justia.com, casebriefs.com).
• Endnote policy. Quote or closely paraphrase with pincites; include parenthetical explanations where helpful.
• Style. Objective, professional, and direct; no rhetorical flourishes, no clichés.

Detailed Outline & Word-Budget

(Total ≈ 8 000 words)

Additional Author Instructions

1. Endnotes only – no inline footnotes.
2. Apply Winter v. NRDC’s four-part test explicitly when critiquing both the TRO and the preliminary injunction (supreme.justia.com).
3. When tackling the political-question doctrine, work through each Baker factor and signal which side it favors (law.cornell.edu).
4. In forecasting appellate outcomes, draw on Ninth-Circuit national-security precedents and Supreme Court Rule 10 criteria for cert-worthiness.
5. Conclude with a brief note on how different rulings could expand or chill protest activity in major cities.

You can see this is far more sophisticated and specific prompt than I had initially provided. It's very much what I was hoping for.

Step 2: Getting Deep Research to Develop a Research Plan

I next shifted over the Gemini 2.5 Pro (preview), which is the best model I have yet found for doing the sort of complex legal research requested here. I turned on the Deep Research button and loaded up all the court documents I could find online from the Newsom v. Trump case so far. (To my shame, I did this part of the project without AI assistance). I cut and pasted o3's prompt into Gemini. (I probably could have done everything within Gemini, but I wanted to give o3 an opportunity). Here's what my Gemini screen looked like.

And now, I let it rip. Here's the plan it came up with after a few minutes of thinking. (This is not the workflow to use if you are in a big rush, but the results tend to be great).

The plan, as usual, was quite detailed and looked good. But to just show you how you can tweak the plan, I then pressed the Edit plan button and added the following.

This is basically fine but I would like you to start your response with a 600 word executive summary. It should be written like a high quality newspaper article in the Wall Street Journal or Washington Post.

Gemini came back to me with a revised plan. I was then ready to press the "Start Research" button. Immediately, I get a display showing what Gemini Deep Research is up to. There are two things I can look at. First, I can look at its progress. Here's what I see at the beginning.

I can also look at the web sites Deep Research is probing. Here's a snippet of what it is up to. You can see it examined 88 web sites. If you want to see its entire thought process, click here.

Step 3: The Report

Within about 10 minutes, Deep Research was done and had produced its report in a "Canvas."Here's a screen capture of part of what it produced. You can see it is indeed starting with the simulated newspaper article I requested.

Step 4: Hallucination Checking

Before proceeding further, I wanted to use AI to detect hallucinations in the report generated by Deep Research. I thus fed the report to ChatGPT o3 Deep Research (same name, slightly different product) with a lengthy prompt that begins as follows:

You are a Senior Legal Editor for a major academic press specializing in legal publications. Your work is meticulous, precise, and focused solely on factual and legal accuracy. Your task is to review the attached legal text—whether a textbook excerpt, law-review article, white paper, litigation brief, or similar document—intended for an audience with foundational legal knowledge. Context: The document explains legal doctrine through analysis of cases, statutes, and other authorities. The style is scholarly yet accessible. The author has made interpretive choices, which you should respect. Your role is not to critique the author’s analysis or tone, but to ensure that every factual and legal predicate is flawlessly accurate. Primary Objective: Perform a rigorous fact-check of the provided text to identify and correct any errors, with particular attention to inaccuracies involving case law, statutory citations, and other legal authorities. Your verification must be grounded in primary legal sources and highly reliable secondary sources.

You can see the full query and output here. Somewhat to my surprise, after 11 minutes of research, ChatGPT concluded "All statements of fact and law in the document were verified as accurate." I'm skeptical that ChatGPT's enthusiasm is warranted, i.e. that Gemini's report is hallucination-free, particularly since ChatGPT says it examined only 20 sources. And were this in the "real world" for a court or client, I would do some additional checks using authoritative databases and my own eyes. But since this is but a mere blog entry outlining a process, we move on.

Step 5: Derivative Products

Create a Google Doc

We now go back to the output from Gemini Deep Research. The "Export" and "Create" buttons on right are now crucial to the remainder of the process. I like to immediately take the output and save it in a Google Docs. I can do that from the drop down menu that appears when I press "Export." I do this so I can use all the tools one has available with respect to a document in the Google Docs system. For example, by putting the material into Google Docs, I get a full outline of the report. Another advantage is that I can get a real word count; Deep Research went a little overboard, producing 13,000 words instead of the requested 8,000. That's OK because it's easy to use AI to condense a document if I want something smaller. A third advantage is it's a lot easier to read a long document in Google Docs than within the Gemini browser window. Here's what my Doc looks like.

With my Google Doc version of the output in place, I now examine that amazing little "Create" button at the top right of the Deep Research output. Here's the drop-down menu resulting from pressing the button.

Having now read the resulting article in a convenient format, I would say it is quite good and certainly well written, supported, and organized. Its primary strength is the detailed examination of the specific statutes governing the domestic use of military force. It effectively distinguishes between the Insurrection Act and the less-litigated 10 U.S.C. § 12406, which is central to the case. The article clearly articulates the key legal questions: whether the protests constituted a "rebellion," whether the President was "unable with the regular forces to execute the laws," and the legal weight of the mandate that orders "shall be issued through the governors of the States". The analysis is further strengthened by its specific use of historical precedents. It effectively contrasts the 2025 deployment with the 1970 Postal Strike—the key historical instance where the "unable to execute laws" prong was invoked—and differentiates it from Civil Rights-era deployments that used the Insurrection Act to enforce federal rights over state objections.

The document useful for multiple audiences.

• For law students: It serves as a practical case study in statutory interpretation, the political question doctrine, and Tenth Amendment challenges.
• For journalists: It supplies the necessary legal context to report accurately on the specific arguments and procedural steps, such as the district court's TRO and the Ninth Circuit's administrative stay.
• For legislators: It highlights how statutory ambiguities can precipitate significant federal-state conflicts, demonstrating a clear case for potential legislative review and clarification of executive emergency powers.

Moreover, contrary to what some people will tell you based on old information, modern AI can and does contain citation to sources. So you can go back and check whether what the document is saying is accurate. Indeed, the fact that it is sourced exposes a possible weakness of the report. It relies too a little too heavily on the opinion of Judge Breyer to interpret the facts of the case. That's probably my fault, however, for failing to point it to more pro-Trump accounts of events in Los Angeles and the ability of local law enforcement to respond capably.

A website

We can also create a web page based on this report. The website can provide a visually appealing and highly condensed version of the analysis. I press the "Web page" item in the "Create" menu. Within a few minutes, Gemini has generated this HTML code.

I can preview how the code is likely to render.

And critically, by pressing the Share button I can create a public version of this page as an actual website: https://g.co/gemini/share/9053d6361989 The result is something more than a mere textual report, whether written by human or machine, can provide. It's an environment for basic exploration. Here are some screen captures of what you see:

A Quiz

I can also produce a quiz. The result is usually 10 multiple choice style questions. Here's a screenshot.

And by pressing the share button, I get an online quiz that each of you (or my students) can take. Give it a try! https://g.co/gemini/share/c0a2fa2a3e8a

Flashcards

Gemini does not limit you to pre-built creations such as a webpage or quiz. You can ask it to try to build its own app based on your work. Here's what I asked.

Build a flashcard deck of about 50 flash cards, The output should be a CSV file in which the first column is the front of the flashcard and the second column is the back side. Here is what the flashcard deck should contain.

(a) front side is case name and year(taken from cases mentioned in the report); back side is a 30-50 word description of the case

(b) front side is a 30-50 word descripton of the case described in the report; back side is the case name and year

(c) front side is a statutory cite or name of a statute; back side is its key provisions

(d) front side is a hypothetical closely resembling a case mentioned in the report; back side is the closest case

(e) front side is a fake statute similar to one discussed in the report but that contains a key change; back side is an explanation of how the fake statute differs from the real one. 

Notice I did not have to do any programming. I just told Gemini what I wanted. Here's a screen shot of what Gemini came back with:

I pressed the little share icon and get back a link by which you can see the entire deck: https://g.co/gemini/share/78cbdc3f925f

I can then import the CSV file Gemini created into a flashcard app such as Anki and then begin my studies. I actually found I had a lot to learn on this topic! Here's an image of what Anki can produce.

Synthetic Podcast

Finally, I can create a synthetic podcast discussing the subject of the report: the authority of President Trump to federalize the National Guard and use the military to address unrest in Los Angeles. I could do this directly from within Gemini: one of the "Create" options is "Audio Overview." But I'd like to get a little more control. So, I take the Google Doc I exported earlier and import it into a new NotebookLM notebook. (I discussed the virtues of Google's NotebookLM in a prior blog post). I also throw into NotebookLM the same PDFs of documents in Newsom v. Trump that I gave to Gemini's Deep Research. I then "Customize" the Audio overview in NotebookLM to be "Short" and to focus mostly on the long report it was provided with. I then press the Generate button.

The first output was problematic in that the two synths appeared to uncritically accept Judge Charles Breyer's opinion. So, in round 2, I added that they should not be afraid to be skeptical of his opinion, particularly given that the Ninth Circuit had issued an administrative stay of his ruling. You can listen to the result here. By the way, although I did not pay the money to do so, you can now apparently create a video version of the synthetic podcast through jogg.ai in a process described here.

Other Derivative Products

This blog post is already too long. So I am not going to discuss today a variety of other derivative products AI can produce either from your work or, as here, its own work. But I can at least tease that future blog posts will look at creation of presentations (think Powerpoint or posters), advocacy documents, preparation for legal argument, and more.

Conclusion

It's sometimes said that AI can't handle complexity, particularly in legal issues. And certainly AI has had historic struggles with long-form output. If I had been writing this last year, I would have told you that after about 2500 words, results got pretty spotty. But as in many other areas, researchers are breaking through those boundaries. With Deep Research, you can get AI to produce fairly in depth reports. This one was 13,000 words but I've been able in other work to get book chapters of 20,000 quality words. And I will let you be the judge, but I would say the quality of the report produced here was pretty high. If I were a journalist, a law student, a legislator, or possibly even an attorney involved in the case, I would find it quite useful.