You are handed an AR 15-6. Forty pages. The 15-6 investigation findings substantiate the allegation, the credibility analysis explains why the complaining witness is believable and you are not, and the recommendations line up like dominoes: a GOMOR, a referred OER, a referral to the SJA. You read it twice.
Something is off, and at first, you can’t name it. The prose is clean. Too clean. Every paragraph the same length, every transition the same shape, every sentence marching in the same cadence. It reads the way a hotel breakfast tastes — technically complete, faintly synthetic, assembled by someone following a recipe rather than cooking.
Then it lands. This doesn’t sound like an officer who sat across a table from a witness and made up his mind. It sounds like a machine.
I’ve now seen this more than once, and I expect to see it a great deal more. So here is the question every defense counsel is going to have to answer soon, if you haven’t already: what do you do when you suspect the investigating officer used a generative AI tool to write, construct, or draft the very investigation being used against your client? Not to fix a comma. Not to check spelling. To do the thinking: to judge credibility, weigh the evidence, choose among competing findings, and shape the 15-6 investigation findings and recommendations that follow.
Two questions follow from that, and they have to be taken in order: how to tell if a document was written by AI, and what to do about it once you believe it was. That is not a technology gripe. It is, I’d argue, a defect in the investigation itself — legal and ethical — and if you treat it as a curiosity instead of a fault line, you will waste the one moment you had to protect your client.
What Counts as AI-Written 15-6 Investigation Findings
Let me define the term I’ll use, and then hold it still.
AI-written Findings:
The situation where the analytic core of an investigation — the credibility determinations, the findings of fact, and the recommendations — was generated by a large language model (LLM) rather than composed by the appointed human officer, whether or not the officer later signed it as his own.
Notice what that definition excludes:
- An officer who runs his own draft through a spell-checker has done nothing wrong.
- An officer who asks a tool to reformat a timeline he built has done nothing wrong.
The problem isn’t that a computer touched the document. The problem is that a computer did the deciding. Keep those two things apart, because the government will try to collapse them, and the distinction is the whole case.
AI Cannot Do What AR 15-6 Investigation Findings Require
Here is the mechanism, because vibes won’t win this.
How a Large Language Model Produces “Findings”
A large language model does not reason from evidence to a conclusion. It predicts likely text. Feed it a record and an instruction, and it produces the words that statistically tend to follow that kind of instruction. That’s it. It is a very good imitation of judgment and none of the thing itself.
What AR 15-6 Requires of the Investigating Officer
Now hold that next to what AR 15-6 actually requires. Every actor the regulation contemplates is a person with assigned duties: an IO who “will conduct a fair and impartial” investigation, an assistant IO if one is appointed, a legal advisor who reviews, an approval authority who may approve, disapprove, or modify. A commercial model appears nowhere in that scheme.
Paragraph 1-9(b):
Paragraph 1-9(b) commands the investigating officer to weigh conflicting evidence and make a reasoned judgment, drawing on demeanor, bias, motive, and, in the regulation’s own words, “sound judgment and life experience.”
Paragraph 3-7(a):
Paragraph 3-7(a) defines a finding as a conclusion of fact by the IO, supportable by the evidence.
Paragraph 3-7(b):
Paragraph 3-7(b) requires the officer to say why the finding he made is more credible and more probable than the alternatives, and to cite record evidence for it.
Paragraph 3-6(a):
Paragraph 3-6(a) tells us why a particular officer was picked in the first place: his education, training, experience, demonstrated sound judgment, and temperament. He was chosen because he was him.
What the LLM Cannot Do
A model has none of that. It never sat in the room. It never watched a witness decide whether to look you in the eye. It has:
- no life experience
- no temperament
- no concept of the customs of the service
- no capacity to be impartial or partial
The further the credibility analysis moves from the officer’s own reasoning, the harder it is to call it his judgment at all. At the far end of that spectrum you have something else entirely: software output wearing his signature block.
That is the tell you felt on the second read. The report is smooth because nobody struggled with it. Real reasoning has seams: a witness who was mostly credible except on one point, a finding the officer reached reluctantly, a recommendation he softened because the soldier’s record earned it. Machine prose sands all of that away. It reads, to borrow a phrase, like a confession written by someone who wasn’t there.
The Prompt Decides the 15-6 Findings
Say the instruction typed into the AI tool was, in substance, “write findings substantiating the allegation and recommending a GOMOR.” The tool will do exactly that. It will marshal the record toward the requested result, resolve every doubt in the requested direction, and quietly suppress the contrary inferences a fair officer would have wrestled with.
The requirement that the officer explain why his finding is “more credible and more probable” than the alternatives becomes theater, because the alternatives were never genuinely considered. They were instructed away before the first sentence was written.
Change one word in that instruction and the “findings” change. Same evidence. Different prompt. Different 15-6 investigation findings and recommendations. That is the heart of it: the output is a function of the instruction, and the instruction is invisible. You are not rebutting an officer’s reasoning. You are rebutting whatever somebody told AI to say, and you can’t see what they told it.
The Duty the Investigating Officer Cannot Delegate
Strip away the technology, and this is an old problem in new clothes.
What AR 15-6 Lets the IO Delegate
An investigating officer is personally commissioned. The appointment memorandum speaks in the first and second person, “I am appointing you,” and directs that officer, and only that officer, to reach the AR 15-6 investigation findings according to his own concept of fairness.
The regulation lets him share the work with one thing and one thing only: another qualified human assistant, designated by the appointing authority and bound by the same impartiality rules.
Why AR 15-6 Judgment Cannot Be Delegated to AI
What it nowhere contemplates is handing the judgment function to a commercial product the appointing authority never selected, and no one can hold accountable.
So the problem, as I’d frame it, isn’t “he used AI.” It’s that he accepted a personal duty to exercise his own judgment, let something else exercise it for him, and represented the result as his. If that sits wrong with you, it should. It’s the same wrong as a judge who lets a clerk decide the case, or an expert who signs a report a technician wrote — except worse, because the “technician” here has no accountability, no oath, and no ability to have formed a view about anything at all.
I don’t say this to grandstand. I say it because if we let this slide once, it becomes the way business gets done, and the next soldier inherits a system where the deciding official has quietly outsourced the deciding.
What AI Detectors Can and Cannot Prove
Here’s the strongest argument against everything I just said, and any honest counsel has to meet it head-on.
Why a Single Detector Result Proves Nothing
You can’t prove it. AI-detection tools are probabilistic, not forensic. They return likelihoods, not verdicts, and the good ones are candid about false positives. A single detector flagging a document proves very little, and if you march into a legal advisor’s office waving one printout with a percentage on it, you will deserve the skepticism you get. That objection is formally correct. Detectors are, on their own, as thin as a probable-cause affidavit written at 1700 on the Thursday before a four-day weekend.
Concede it. Then explain why it misses the premise.
What Works Instead of a Detector Score
You are not trying to prove AI authorship to a criminal standard from your desk. You can’t, and you don’t have to. Two things carry the weight instead:
Convergence: When Independent Detectors Agree
First, convergence: when multiple independently built detectors, using different methods, localize the AI signal in the same place, the credibility analysis, the findings, the recommendations, the “one tool got it wrong” answer loses its force. Different tools don’t usually make the same mistake in the same spot.
Question Authorship, Then Demand the Author, Tool, and Prompts
Second, and more important, the decisive facts are not in your hands. The identity of the true author, the tool used, and above all the instructions given to it are uniquely within the government’s control. Your burden isn’t to prove authorship. It’s to show a real, substantiated question about it and then demand that the party holding the answers produce them. A respondent cannot be made to rebut a black box. Put the box in front of the people who built it and ask them to open it.
What to Do Before Responding to the AR 15-6 Findings
Here is the operating rule I’d give any counsel who calls me with that queasy second-read feeling. Do these things, in this order, before you write a single word of rebuttal to the AR 15-6 investigation findings.
Preserve: Send a Litigation Hold Immediately
Preserve first. Send an immediate litigation hold and preservation demand. The evidence that proves or disproves AI authorship is fragile and mostly digital: the original electronic file with its hidden metadata and edit history, the account logs of whatever AI tool was used, the prompt-and-output sessions, the device itself.
That evidence can evaporate faster than a chain of custody gets muddy after too many handoffs. Your hold letter should name it specifically and warn, in plain terms, that spoliation will be argued if it disappears.
Demand Full AI Disclosure
Then demand disclosure, all of it. Every AI system, model, and version used. Every exhibit or summary fed to the tool. Whether the tool operated under a no-train, no-retain configuration or dumped your client’s file into someone’s training data. Every prompt and every follow-up instruction, in sequence, because the sequence is the clearest record of intent.
And the identity of the human author of each section, so the record shows who decided what. Ask, too, for the governing AI-use policies at every relevant level of command, and any correspondence about them. If there was a rule, you want it. If there wasn’t, that silence tells its own story.
Insist the Government Fund the Forensic Exam
Then insist the government, not your client, fund and conduct the forensic examination that settles authorship, and insist it happen before any response on the merits is due.
Do Not Respond on the Merits Yet
Responding now is premature and futile:
Requiring a response before authorship is settled shifts a burden onto the respondent that the regulation never put there. So, ask for the extension. Ask for tolling. Reserve every right. Put a real deadline on the government’s written confirmation, and calendar it.
AI-Written 15-6 Investigation Findings Cannot Stand
AR 15-6 investigation findings draw their legal force from one thing: that a specific human officer, chosen for his judgment, actually exercised it. Take that away, and the report is not a weak investigation. It is not an investigation. The legal advisor’s sufficiency review assumes the findings are the officer’s reasoned conclusions; if they’re a machine’s, the review is reviewing nothing. The approval authority is signing off on a judgment nobody made.
So, when the document in your hands reads too smooth, too even, too much like it was written by someone who never met your client, believe the feeling, and then do the work it demands. Don’t argue the merits of findings whose author you can’t identify. Make the government prove a human decided the case. If they can’t, or won’t, the report cannot lawfully support adverse action, and you should say so, in writing, in a room where you’d be proud to have your name on the letter.
Because at the end of the day, your client is entitled to a decision made by a person. Not a prediction of what a decision might have looked like.
FAQs
How Can I Tell if a Document Was Written by AI?
You cannot tell from an investigation report’s prose alone. Uniform cadence, generic credibility language, unexplained factual certainty, and detector results can justify an inquiry, but the stronger evidence is the original electronic file, its edit history, account logs, prompts, outputs, and testimony from the supposed author.
What Should I Do if the Government Refuses to Disclose AI Prompts?
If the government refuses to disclose the prompts, ask for a written denial, the stated legal basis, preservation of the complete digital record, and an extension of every response deadline. Then explain in writing why a merits rebuttal cannot fairly answer reasoning whose instructions and author remain concealed.
Who Can Help Me Challenge AR 15-6 Investigation Findings Written by AI?
The Law Office of Will M. Helixon helps Soldiers challenge AR 15-6 investigation findings that may have been produced by generative AI. The work begins with preservation, disclosure demands, authorship evidence, and a disciplined attack on any judgment the appointed officer did not personally make.
Can an AI Detector Prove Who Wrote the AR 15-6 Investigation Findings?
No. A detector can support further inquiry, but authorship must be tested through native files, edit history, prompts, account records, and evidence from the supposed author.
Every case is different. Past results reflect the facts, law, and advocacy specific to that matter and do not predict the outcome of yours. The regulatory duties described here come from AR 15-6, Procedures for Administrative Investigations and Boards of Officers (22 June 2025).
