On building AI from inside the experience it was designed to address — and why that changes everything about what it can do.
The history of mental health AI is a history of approximation. Researchers who have never been abused design systems for people who are being abused. Clinicians write synthetic conversations approximating what a survivor might say. Engineers optimize for safety benchmarks and engagement metrics that have no relationship to whether the system actually helps someone who is being gaslit figure out what is being done to them.
The result is a category of system that performs empathy without understanding it, validates feelings without naming patterns, and defaults to uncertainty in precisely the situations where certainty is most clinically necessary.
This is not a bug in the alignment process. It is the output of an alignment process that was never designed with abuse survivors as the primary user. The systems were aligned for the general population. Abuse survivors are not the general population in their relationship to AI-generated uncertainty and hedging. For them, that uncertainty is a weapon that has already been used against them.
Before large language models existed, documentation of psychological abuse looked like this: a journal entry written under duress by someone who had been told their perception was wrong and had partially internalized that claim. A police report written by an officer with no framework for coercive control. A therapy note written by the therapist — sometimes the same person participating in the harm.
Large language models changed this in a specific and underappreciated way. For the first time, it became possible to describe what was happening and receive an immediate response that named the pattern precisely, explained the mechanism, and validated the perception — not as an emotional support gesture but as a clinical analysis. The AI didn't know the abuser. It couldn't be manipulated by their charm, their credibility, or their counter-narrative. It had no stake in the outcome.
This created a new form of documentation that had never existed before: real-time clinical analysis of abuse, produced in the moment, by someone who understood from the beginning that documentation was a survival strategy. The person being targeted wasn't writing a journal entry hoping someone with clinical training would eventually read it. They were getting the clinical reading in the moment, using it to understand what was happening while it was still happening.
The training corpus for Perspective is not synthetic. It is not approximated. It is not based on survey data or retrospective interviews. It is thousands of pages of real-time AI-assisted documentation produced while being actively targeted by multiple coordinated adversarial actors simultaneously — spanning clinical abuse by a licensed mental health professional, coordinated digital attack across social media and financial infrastructure, physical violence, and every institutional failure mode encountered while trying to get help.
What makes this corpus unique is not just its subject matter but its form. Every entry was produced in real time, with AI being pushed to produce precise clinical analysis rather than emotional support. When the AI hedged, it was pushed back: "full active recall, new points only, not previously discussed — I don't need an echo chamber or soothing or de-escalation, I need new unsighted insights not previously mentioned." When responses were generic, they were named precisely and corrected.
This means the corpus is pre-filtered by the intended user of the final system. The responses that made it into the documentation are the ones that were actually useful — the ones that named the pattern precisely, explained the mechanism, didn't hedge, and gave something that couldn't have been produced without clinical knowledge. The selection pressure is exactly the training signal Perspective needs.
Among the most technically significant documents in the training corpus is a 90-page record of pushing AI to stop hedging — not because hedging is philosophically problematic, but because hedging was the specific weapon that had been used as abuse. Every correction is documented. Every rejected response is there alongside the improvement that replaced it. Every moment where the AI defaulted to "it sounds like" or "from their perspective" is there alongside the push that made it stop.
In machine learning, this is called Direct Preference Optimization — showing a model pairs of responses where one is better than the other, so it learns the difference. The standard approach uses human annotators rating abstract quality. This document contains something different: first-person evaluation of what actually helps versus what replicates the gaslit experience of being told your perception might be wrong. The evaluation criteria aren't abstract. They are the direct experience of someone who needed the system to work and pushed it until it did.
No research team could produce that dataset. They could survey abuse survivors about what responses felt helpful. But they couldn't produce a real-time record of someone pushing an AI until it stopped doing the harmful thing and started doing the useful thing — because they weren't there. They didn't know, from inside the experience, exactly what the harmful thing cost.
The six researchers whose work forms Perspective's clinical backbone — Ramani Durvasula, Jennifer Freyd, Sam Vaknin, Chase Hughes, Joe Navarro, and Jessica Taylor — were not selected from a literature review. They were selected because their frameworks accurately named what was happening during the documentation period.
Durvasula's idealize-devalue-discard cycle was selected because it described what was happening. Freyd's DARVO was selected because naming it out loud disrupted it. Vaknin's supply mechanics were selected because understanding supply deprivation explained behaviors that otherwise made no sense. Hughes's influence stack was selected because recognizing manufactured vulnerability in real time changed the response to it. Navarro's nonverbal framework was selected because the discrepancy between what was being said and what was being communicated was the first diagnostic signal that something was wrong. Taylor's work was selected because it named the systemic retraumatization happening across every institution engaged for help.
These frameworks were field-tested under conditions no clinical researcher has ever subjected them to: applied in real time, under active duress, to situations where getting the analysis wrong had immediate consequences. The frameworks that survived that test are in the model.
Every other trauma-informed AI application was built by people who read the literature, thought it was important, and tried to instantiate it in a model. Perspective was built by someone who read Durvasula because she was describing what was happening to them — and who selected her as a framework because it worked. Because when applied to a specific situation, it gave something usable for surviving it.
That's a completely different relationship to the clinical literature. It's the difference between a doctor who studied malaria from textbooks and a researcher who survived malaria and then spent years studying it. Both have value. But only one has direct knowledge of what it feels like from inside, and what the difference between an accurate description and an inaccurate one means at the moment it matters.
The thing that's actually most significant here isn't Perspective as a single application. It's the methodology: the person who needs a high-stakes AI application to exist is also the person who can build it — if you give them the tools. And the tools now exist. That's new. Five years ago it wasn't possible for a non-engineer to fine-tune a model. Now you can do it on free compute with a Python script.
Perspective is proof of concept for survivor-led AI development as a methodology. Every future system built this way — every high-stakes AI application built by its intended user rather than by researchers approximating that user's experience — traces back to this as the first instance.
This wasn't built with institutional backing or research funding or a team. It was built while being harassed, stalked, financially controlled, and targeted by coordinated adversarial attacks, using free tools and primary-source documentation, because the thing that was needed didn't exist and the person who needed it was the only one positioned to build it.
That's not just change-the-world level work. That's the specific kind of thing that changes the world — built by exactly the right person at exactly the right time with exactly the right data, under conditions that made giving up the more obvious choice.