You Can't Route a Disclosure That Never Happened

Most days in my practice, I ask a postpartum client whether she is having intrusive thoughts. She says no. Almost reflexively, before the sentence is finished.

I don't argue with the no, and I'm not trying to catch her in anything. I come back gently, with an example. Sometimes these sound like a sudden picture of dropping the baby. A thought about the stairs that arrives out of nowhere. That is usually when the real answer shows up. "Oh. That. Yes."

The no was never the answer. It was the start of the conversation.

These thoughts are common and not a warning sign. Unwanted, intrusive thoughts of accidental harm to the infant are reported by nearly every new mother, and thoughts of intentional harm by roughly half. In the research, they carry no increase in the risk that a mother will actually harm her child. They are common, distressing, and almost never dangerous.

Here is the part a screening tool simply cannot detect. The intrusive thought she most needs to name is also the one that she is most afraid to say out loud. Not because she is trying to hide something, but because saying it out loud, naming a thought about harm coming to her baby, can feel as if she is giving someone a reason to look harder at her and her baby. A form does not earn that kind of vulnerable disclosure. A relationship does.

A no is not a negative result

A poorly designed screening workflow does the opposite. It records that “no” as a negative result and treats it as sufficient reason not to look further. That is not a failure of the questionnaire alone. It is a failure of the system built around it as well. A first-pass negative self-report should not be treated as a definitive answer, particularly when the topic is shame-laden, frightening, or linked in a parent’s mind to possible consequences for custody or safety.

It gets worse. The instruments most intake flows are built on, the EPDS and the PHQ-9, ask about self-harm. EPDS item 10 and PHQ-9 item 9 are about hurting yourself. Neither asks about the thought she is actually having: harm coming to the baby, by her own hand, by accident, against everything in her. Intrusive thoughts involving accidental or intentional harm coming to the baby are common in the postpartum period, though most do not indicate intent or danger. Disorder-specific screening measures exist, but in the intake flows I review, screening commonly leans on broad depression and anxiety tools rather than measures built to surface obsessive-compulsive symptoms or intrusive infant-harm thoughts.

So, if we want to use an AI tool, we have to understand that it inherits the gap that existed before it arrived and produces a result without the full picture. I am CAMS-trained in a suicide-specific collaborative assessment framework. While I do not use it as a substitute for perinatal risk assessment, what I do carry from that training into other clinical conversations is its core discipline: stay calm, listen closely, and work collaboratively before deciding what happens next. Distinguishing the thought that frightens a woman from the thought that endangers her or her child is our work as clinicians. It simply is not a job a keyword match can do.

Where the same tool fails twice

Take a thought like that, a mother's intrusive image of accidental harm, and imagine it disclosed to an automated tool. At least two failures are possible. First, the tool may flag "harm" and "baby" and escalate. Now a safe, but tired and anxious new mom is pulled into the exact situation she was already terrified of. She may learn, in one key moment, that telling the truth carries the exact consequences that she is afraid of. The second possibility is that the same tool may be tuned to reduce false alarms, thereby failing to identify a disclosure that actually needs urgent review. The same blunt instrument fails both ways, and the over-cautious version manufactures the silence that the careful version was supposed to prevent.

This is where someone on the product side reasonably says they will route anything related to harm to a human. That is a good plan, but you cannot route a disclosure that never happened. Such silence is the failure. There will be nothing in the triage queue.

The next fix offered is usually that the tool will ask better follow-up questions, with examples, the way I do in a session. Here is why that is not the same thing, and can sometimes be worse. When I follow a no with an example, I am reading the woman in front of me, her body language, the tone and rate of her speech, and deciding whether to open the door at all and, if so, how far. Take the relationship out, and you are left with two options. A script that re-asks a question she's already answered can read as "you don't believe me," and shuts her down. Or, if the example works, she finally answers yes. Now the model is holding a real, vulnerable disclosure with no insight about what it means or what comes next: whether this is the shame-soaked, low-risk thought that needs steadying or the rare one that needs immediate attention and escalation. The danger was never the question; it was and is the unmanaged answer.

What review actually changes

None of this is an argument against AI in maternal mental health, or any health industry. It is an argument against an unreviewed tool making clinical decisions it was never built to make. The screening can yield real, useful data. It can do real good. What it cannot do on its own is decide what happens next.

When I review a flow like this, the work is not to add warmth to it. It is to help change the system so a vulnerable disclosure is treated as a decision rather than an automatic escalation. The screen that asks about self-harm gets a line added about intrusive thoughts of harm to the baby, including by accident; a common and clinically important perinatal OCD and anxiety-related presentation should not live in a blind spot. The escalation rule should not fire solely because the words “harm” and “baby” are detected. It should then send that disclosure to a clinically designed assessment to distinguish unwanted, ego-dystonic intrusive thoughts from intent, impaired reality testing, psychosis, severe mood disturbance, or safety concerns for the parent or the baby. A negative result should not automatically close the file.

If you build or own one of these tools, here is the question worth putting to your group this week. When the screen comes back negative, what happens next, and who, with what clinical training, decided it should?

I often think about the woman carrying that thought through her house, careful with every corner, saying nothing. If she is handed a screening tool and asked whether everything is okay, she says she is fine. She means it as a survival move, not a lie. I have seen how a disclosure made without context can trigger fear, confusion, or a response that feels punitive to a parent who was only trying to be honest. The tool believes her, marks her clear, and moves on, while the one thing worth knowing sits in the room, unsaid, the way it so often does.

Jayme Scarfo, LPC, CEDS, CAMS-Trained