When AI Labels a Message “Manipulative”: How to Check the Report Before You Rely on It

Suppose you upload six months of messages to an AI tool. A report comes back with tidy sections, a few highlighted exchanges, and a heading in red: “COERCIVE CONTROL.”

After months of struggling to explain what’s been happening, seeing those words might feel like relief. Someone, or something, has finally put a name to it.

Then you look for the messages behind the heading, and the report gives you fragments. It doesn’t say whether it read the attachments, which dates it covered, or what happened after the exchange it quoted.

That report is invented, but the question it raises applies to every AI message report.

What produced that sentence?

A useful analysis gives you somewhere to start looking. Before you rely on it, you need to be able to follow its claims back to the record.

Start with what the report actually claims

Software can help sort a long conversation, find repeated phrases, and draft a description of an exchange. That can save you from rereading months of difficult messages just to find one appointment or one argument.

Describing someone’s conduct as “manipulative,” “abusive,” or “coercive control” is a bigger step. At that point the report is interpreting the messages, and an interpretation needs checking, even when it sounds persuasive.

If you wrote the request yourself, or can see the one the tool was given, reread it. “Find proof of manipulation” starts from an assumption, and a tool can end up finding what it was asked to find. Asking it to point out relevant passages, possible readings, and missing context gives you more to examine. Either way, the label it produces isn’t a legal finding or a professional assessment.

The National Institute of Standards and Technology’s Generative AI Profile lists “confabulation” among the risks of generative AI: “the production of confidently stated but erroneous or false content.” It also warns that people may rely too heavily on these systems, or rate what they produce above other sources without good reason. Polished formatting doesn’t solve either problem.

Read the claim itself closely. “This message contains a threat” is a different claim from “this person regularly uses threats to control the other parent.” The second describes a pattern, so it needs the exchanges being counted and the period they cover.

Treat scores the same way. If a report says “92% confidence,” ask what that number measures and how the tool was tested. Without that explanation, don’t read it as a 92% chance that abuse happened.

Find out what the software read

Next, coverage. “I uploaded my history” and “the analysis considered every message in it” aren’t the same thing. Find out which files were read successfully, and whether the report covers the whole import, selected incidents, or a sample.

You should be able to identify:

• The files and conversations included.
• The first and last dates analyzed.
• Which participant is which, including any “Parent A” and “Parent B” labels.
• Whether photos, attachments, voice messages, and call transcripts were included.

If the report doesn’t say, check the import results or ask the provider. Write down anything you can’t confirm.

Then compare a few imported messages with the original file: the sender, the date, and the wording. If the source was a screenshot or a scanned PDF, look at the image beside the extracted text. A missed “not,” a wrong date, or a message credited to the wrong person changes what the analysis is working from.

Watch for double counting, too. If you uploaded overlapping exports, or several screenshots of the same exchange, check that one incident hasn’t been counted more than once.

Open the conversation behind the label

For each claim you plan to share, find the messages behind it. A useful report lets another person move from the claim to the exact messages, and from those messages to the surrounding conversation. A page number, a timestamp, or a direct link is what makes that possible.

Here’s an invented example.

The later message changes what a reader can reasonably take from the first. It says where to find the appointment details and how to reach this parent in an emergency. It doesn’t settle everything about the relationship, but it does mean the summary needs a closer look.

“Repeatedly” deserves a check of its own. Where are the other occasions? Are they separate events, or several messages from one argument?

Read far enough back to understand what was being discussed, and far enough forward to see what happened next. If a message points to another app or an attachment, follow it if you can. Include your own messages in that review, even when they make the exchange harder to explain.

Check quotation marks as well. If the report presents something as a quotation, compare it word for word with the source. A paraphrase should be labeled as one. Words nobody wrote should never become a quotation just because they appear in a generated report.

Make the gaps visible

A message archive can hold a great deal and still leave out something important. A phone call may explain why a plan changed. An attachment may hold the answer to a question. Something that happened in person may explain why a short message felt threatening.

The U.S. Department of Justice’s overview of domestic violence lists physical, sexual, emotional, economic, psychological, and technological abuse. Much of that can happen without leaving a trace in a message thread. A report can only address the material it was given. It can’t establish the whole picture of a relationship.

Write a short note about the gaps. For example: “This report covers messages from the parenting app, January through June. It doesn’t include calls, text messages, or the attachments missing from the export.”

That note helps whoever reads the report judge its scope. A blank stretch of time should stay marked as a gap unless you can show what happened during it.

The same limit applies when a tool finds nothing concerning. An unflagged exchange doesn’t mean the situation is safe. You can take a concern seriously, and talk with a trained advocate, without waiting for software to give it a label. In the US, the National Domestic Violence Hotline offers free, confidential support around the clock at 1-800-799-7233.

Keep the original and a record of your review

Save the original export separately from the AI report. Make corrections and notes in a working copy, and keep the report exactly as it first appeared, so you can always tell what the software wrote and what you changed.

The Safety Net Project at the National Network to End Domestic Violence has practical documentation tips, including how to preserve messages and the identifying information that goes with them. It warns, for example, that “a forwarded email will not have the same information” as the original.

For the AI analysis itself, keep a short record: the tool you used, the date, the files you submitted, and any instructions or settings you chose. Note the model or version if the service shows it. You don’t need to guess at what a provider doesn’t disclose. Just mark it as unavailable.

Your review notes should explain the corrections that matter. “Changed ‘no response’ because a reply appears in the next day’s email” is useful. So is “Sender can’t be confirmed from this cropped screenshot.”

Keep your own knowledge identifiable. If you remember a call the report couldn’t see, describe it separately, as your recollection. If a teacher told you something, note that it’s the teacher’s account. A reader should be able to distinguish the messages, the software’s interpretation, and the information you added.

These notes help you explain how the report was checked. On their own, they don’t show that its conclusions are right.

Check privacy before another upload

It can be tempting to send the same archive to several services and compare their answers. Before you do, find out where each one sends your information and what happens to it afterward.

Look for specific answers: who processes the content, how long it’s kept, whether it’s used for training, and who can open a shared report. A “no training” promise answers only one of those questions. Check the policy for the specific feature and account type you plan to use.

Our own security overview takes the desktop app, optional Cloud AI, and the tools that run in your browser one at a time, because the answers differ.

Think about where you save the report, too. Safety Net’s cloud safety guide explains that if your cloud account is shared or linked to someone else’s, they may be able to see what’s stored there. If you think you’re being monitored, talk with an advocate about safety before changing accounts or settings.

If the material includes messages with your attorney, or anything a court order says you can’t share, ask your attorney before uploading it.

Ask a question someone can answer

Before you send the report to your attorney, write down what you want help with. “Can you review these three exchanges about access to appointment information?” gives them a specific question and a place to start.

Include the relevant messages, enough context to follow them, and your note about what the analysis left out. Mark clearly which parts the AI wrote and which corrections are yours. Ask whether they want the report itself, or would rather have a short timeline with references to the source messages.

For an example of how one family court handles AI, look to Australia. On May 29, 2026, the Federal Circuit and Family Court of Australia issued a practice direction on AI that applies to everyone using those courts, including people without a lawyer. It requires them to verify AI output and to be ready to explain, if asked, which tool they used and how they checked its work. It also says not to assume a tool is reliable because it’s paid or private.

That direction applies only in those Australian courts. For your own case, ask your attorney, or your court’s help desk if you don’t have one, which rules apply and whether material an AI tool produced should be used at all. A report’s title or score can’t answer that question.

Where CustodyComms fits

CustodyComms keeps finding an exchange and interpreting it as separate steps. In the desktop app, the Incident Finder runs on your computer and points you to exchanges worth reading. If you choose, Cloud AI writes up an incident you select, and you can edit every writeup.

When Message Review flags an exchange, its category tags name the keyword group that matched, such as “Coercive Control keywords.” That tells you which words it found. It isn’t a finding about what anyone did.

You can try Message Review on one export to see which exchanges it surfaces. Unlike the desktop app, it isn’t local: the text from your file is sent to OpenAI in the United States to produce the report. Read what the page says about eligibility and privacy before you upload.

Then apply the same checks to our writeup that you would to any other AI report. Open the messages, read the context, and correct what the description gets wrong.

Go back to that red heading. Once you’ve checked it, it’s no longer the answer. It’s a starting point: the exchange that prompted it, a note of what the report couldn’t see, and a specific question for the person helping you.