There is a moment that happens in nearly every first session with someone trying to make sense of a controlling relationship. They reach for their phone. They scroll, sometimes for a long time, looking for the message that will explain it. Then they find one, hold it out, and watch your face while you read.
And the message is nothing. “I never said that.” “You’re remembering it wrong.” “I was only joking, you’re so sensitive.” Ordinary sentences that pass through a room without disturbing the air.
What happens next matters enormously. If you respond as most people outside the room respond, with a slight puzzlement, the client concludes that they have wasted your time and possibly invented the entire problem. If you understand why the message looks like nothing, you can do the actual work.
The research has caught up
Until recently, the explanation for this was clinical intuition. Practitioners knew the harm was cumulative because we saw it repeatedly, but knowing it and being able to demonstrate it are separate problems.
Research presented at CHI 2026, one of the most significant venues in human-computer interaction, has now stated the mechanism in technical terms. The study, on AI-facilitated coercive control by Haesoo Kim and colleagues, constructed scenarios combining established coercive control tactics with the capabilities of current conversational AI systems, and tested them. Alongside its findings on how these tools can be misused, it identifies where detection would have to happen: in the analysis of multi-turn and multi-session patterns of interaction, because that is the level at which the hallmarks of coercive control become visible.
Read that carefully, because it contains the whole thing. The identifying features of coercive control are not present within a single exchange. They emerge across many exchanges over time. This is not a claim about how difficult detection is. It is a claim about where the phenomenon exists.
Why the deniable version is the one that lasts
Consider what has to be true for a controlling relationship to persist for years.
If every exchange were plainly abusive, it would be identified early. Friends would notice. Colleagues would notice. The person experiencing it would have straightforward language for what was happening and unambiguous evidence when they used it. Relationships of that kind get named and, more often than not, ended.
The version that persists is the one where each moment is individually defensible. Where every objection can be met with an explanation that sounds reasonable to anyone hearing it fresh. Where the person raising the concern ends up apologising, again, for a reason they cannot quite reconstruct afterwards.
Deniability is not incidental to this dynamic. It is load-bearing. The pattern survives precisely because no individual element of it will bear the weight of an accusation. Which means that when a client cannot find the smoking gun, the absence of one is not evidence that they are mistaken. It is close to diagnostic.
What the pattern actually looks like
If the unit of evidence is the sequence, the practical question becomes what to examine in it.
Frequency and distribution. Reality-denial appearing four times in three years is a difficult relationship. The same phrase appearing weekly for eleven months, clustered around any attempt to raise a problem, is a structure.
Accountability direction. Track what happens when the client raises a concern. In healthy conflict the direction of accountability moves around. In controlled dynamics it moves reliably one way, and by the end of the exchange the person who raised the issue is the one apologising. That reliability is more informative than any single instance.
Range contraction. Compare what the client could say without consequence in month one against month twenty. The narrowing is usually invisible from inside because it happens gradually, and stark when the two periods are placed side by side.
Escalation after boundaries. What happens in the exchanges immediately following an attempt to set a limit. A pattern of intensification there is one of the more reliable signals available.
None of these can be read from one message. All of them are readable across a body of correspondence, and all of them are the kind of thing a third party can be shown.
What this means for tools
Almost every product currently marketed for this purpose analyses individual messages. You submit a conversation. It scores it for manipulative language. It flags the concerning parts and returns a result.
The models doing that scoring are improving quickly, and it will not solve the problem. The limitation is not accuracy. It is that the tool is examining a different object from the one that matters. A relationship can consist almost entirely of individually innocuous exchanges and still be tightly and damagingly controlled. A perfect single-message classifier would return nothing of interest on such a relationship, and it would be functioning exactly as designed.
This is why practitioners should ask a different first question when evaluating anything in this category. Not how accurate is it. What does it look at. A tool that scores messages and a tool that analyses sequences across time are not competitors at different quality levels. They are answering different questions, and only one of those questions is the one your clients are actually asking.
The work this changes
The shift in clinical practice is straightforward, if not always easy.
Stop hunting for the worst message with the client. It reinforces the belief that the harm needs to be locatable in a single moment, and when the search fails, and it will fail, they take that as confirmation that they have exaggerated.
Build the sequence instead. Take the full correspondence over a defined period. Establish what recurs, how often, in what circumstances, and what the direction of travel is. That is the object which needs to become visible, both to the client and to whoever else needs convincing.
This is laborious to do by hand, which is a considerable part of why it is so rarely done. It is exactly the kind of task that pattern analysis across a documented record is suited to, and it is what Validate was built for: examining a body of correspondence, identifying where documented patterns of coercive control appear across it, and showing the specific exchanges that support each finding. The output is not a claim about how anyone feels. It is a description of what is in the record, with the record attached.
The client scrolling through their phone in that first session is not looking for the wrong message. They are looking for the wrong kind of thing. There is no single message, there was never going to be one, and the sooner that is said out loud in the room the sooner the real work can start.
