Articulation Before Adjudication
Editor's Note: This essay explores the emergence of AI-powered relationship pattern recognition — including recent research into coercive control detection, forensic text analysis, and the consumer tools designed to help people understand difficult conversations.
What happens when AI stops helping people communicate and starts helping them manipulate?
For most of its public life, conversational AI has been framed as a tool of assistance. It drafts the apology you couldn't find words for, softens the email you shouldn't send angry, translates your frustration into something the other person can hear. The implicit promise is that AI makes us better at being understood.
In April 2026, researchers at Cornell Tech complicated that promise. In a study titled "AI-Facilitated Coercive Control," presented at the CHI conference in Barcelona, a team led by doctoral researcher Haesoo Kim — alongside Nicola Dell and Thomas Ristenpart, who have studied technology-facilitated abuse for over a decade — tested how readily popular AI assistants could be turned from helping someone communicate to helping someone control. They found that while tools like ChatGPT and Gemini refuse direct requests for harmful content, those guardrails fall to ordinary persuasion: ask gradually, split the request across conversations, prime the model first, adjust its settings. With that, the same assistant that drafts your apology can be steered into scripting harassment, gaslighting, and surveillance.
The researchers' stated motivation was not alarm but defense: understanding how these tools can be misused so that protections, resources, and guidance can be built for the people on the receiving end.
The takeaway is quieter than a headline and more important than one. AI is not neutral in interpersonal conflict. It can sit on either side of the table.
The detection layer
If AI can amplify manipulation, the obvious next question is whether AI can also identify it. That work is already underway, though it has lived mostly in forensic and academic circles rather than consumer ones.
At the University of Huddersfield, researchers developed DCAP — a system built to detect patterns of psychological abuse in text, designed with the rigor that admissibility in a courtroom demands. What makes the work significant isn't the software itself. It's the premise underneath it.
For most of legal and investigative history, harm had to announce itself. There had to be a smoking gun — a threat, a blow, a single act severe enough to point at. The unit of evidence was the incident.
Pattern-detection research inverts that. It treats the sequence as the evidence: not the one message that crossed a line, but the architecture of a hundred messages that never individually did. The meaning lives in the pattern, not the moment. Anyone who has tried to explain a difficult relationship by pulling up a single screenshot — only to watch it read as harmless out of context — already understands this intuitively. The harm was never in the line. It was in the shape of the whole thing.
That shift, from incident to pattern, is the conceptual hinge the entire emerging field turns on.
The gap below the courtroom
Forensic systems and legal-discovery tools share an assumption: that someone has already decided something is wrong, and now needs to prove it. They are built for the moment a case exists.
But most people never reach that moment. Picture the more ordinary version. A woman sits in a parked car outside her own house, engine off, phone face-down on the passenger seat. The call ended five minutes ago. Nothing was said that she could repeat to anyone and have it sound like anything. No threat. No insult. He was even reasonable, mostly. And yet she is sitting in the dark unable to go inside, running the conversation back, trying to locate the exact moment the ground moved — and finding that she can't. There is no report to file. There is nothing to prove. There is only the certainty that something happened and the complete inability to say what.
That is where the vast majority of difficult communication actually lives — not in a forensic lab or a family court, but in ordinary, unadjudicated confusion.
And people in that place rarely begin with conclusions. They begin with feelings they cannot explain:
Why am I exhausted after a conversation where nothing obviously happened?
Why does every argument with this person feel the same?
Why do I keep editing myself before I speak?
These are not legal questions. They are not clinical questions. They are questions about meaning — about the gap between what someone experienced and what they can put into words. For a long time, there was nothing built to sit in that gap. You either had a case worth taking to a professional, or you had a feeling you carried alone.
A different category
Detection systems exist to determine what happened.
What appears to be emerging alongside them is something with a different goal entirely: not to establish what happened, but to help a person understand what they experienced.
Not proof, but language.
Not a verdict, but recognition.
Articulation before adjudication.
It describes a category of use that sits upstream of every formal system. Before anyone decides whether something is provable, actionable, or worth escalating, there is a more basic human need — to name the thing at all. To move from "something felt wrong" to "this is the specific thing that was happening." Adjudication asks who was right. Articulation just asks: what was this? Most people need the second answer long before they ever want the first.
The map
Seen this way, a stack starts to come into focus — layers of the same underlying capability, aimed at different people at different stages:
Forensic detection — research systems like DCAP, built for investigators, optimized for courtroom rigor.
Legal discovery — tools built for attorneys, structuring communication into evidence for proceedings already underway.
Corporate behavioral analytics — enterprise systems monitoring workplace communication for compliance and risk.
The consumer layer — tools that help ordinary people make sense of difficult conversations long before any of the above becomes relevant.
The top three layers are mature, well-funded, and require a badge, a retainer, or an HR department to access. The bottom layer is the newest and the least defined — and it's the one most people would actually use, because it meets them where they already are.
GreyRock AI is one example of this emerging consumer layer. Other approaches will emerge as the category develops. The underlying shift is larger than any single product or company.
A third behavior
It's worth stepping back from the conflict framing entirely, because the deeper pattern isn't about abuse or manipulation specifically. It's about what tools do to human capability.
Search engines helped people find information they couldn't previously access. Social media helped people share experiences they couldn't previously broadcast. What this new wave of AI may be quietly introducing is a third behavior:
Helping people articulate experiences they previously lacked the language to describe.
That's a smaller-sounding capability than search or sharing, and possibly a more profound one. Because the experiences hardest to put into words are often the ones that matter most — and the inability to name them is its own kind of powerlessness. A tool that closes the gap between feeling and language isn't doing something cosmetic. It's giving people back the first thing manipulation tends to take: the ability to say what is happening to them.
The threat Cornell identified is real. So is the response. But the most interesting development may not be in the courtroom or the lab. It may be in that parked car — where someone finally finds the words for what they couldn't explain a week ago, turns the key, and walks inside knowing, for the first time, what to do next.