There is a result from an AI abuse detection tool that nobody talks about. Not the false positive, where the tool flags something that turns out to be nothing. The false negative: where the tool finds nothing, and the person using it concludes that nothing is wrong.
For most applications, a false negative is a minor inconvenience. For someone in a coercive relationship who has spent months being told they imagine things, a tool that returns no result is not a minor inconvenience. It becomes one more piece of evidence in a case that is already being built against their own perception.
This is the clinical problem with AI manipulation detection that the product announcements do not address.
What the accuracy numbers actually measure
The most cited figure in AI gaslighting detection right now is ninety-four percent accuracy. It appears prominently in at least one consumer tool’s marketing and is circulating in practitioner discussions as evidence that the technology has matured.
I went and read the research it comes from.
The ninety-four percent figure is drawn from a study on detecting basic emotional tone in text: joy, anger, sadness, fear, disgust, neutral. That is a classifiable problem with a relatively clear ground truth and decades of training data behind it. It is not the same task as detecting gaslighting.
Detecting gaslighting requires reading context across time. It requires understanding the relationship between what was said and how it was received, between this message and the one three months ago, between the explicit content of a statement and its function within a pattern of control. The model in the same article built specifically for that task achieved seventy-two percent recall on actual gaslighting cases.
Seventy-two percent recall means approximately one in four gaslighting cases received no flag.
One number is doing the marketing. The other number is doing the clinical work.
Why a missed result is not neutral
In most clinical contexts, a test that misses one in four cases is a problem worth noting and managing. In the specific context of coercive control, it is a problem with a particular and serious character.
The primary mechanism of coercive control is the systematic erosion of the target’s capacity to trust their own perception. This is not a side effect of the controlling behaviour. It is the method. Gaslighting, reality distortion, the attribution of the target’s legitimate concerns to their own instability: these work precisely because they accumulate. Each individual instance is deniable. The pattern, over time, produces someone who no longer trusts what they see, hear, or feel.
A person who has been in this process for months or years and then uses an AI tool that returns no finding is not receiving a neutral result. They are receiving confirmation of what they have already been told: that they imagined it, that they overreacted, that nothing is wrong.
The tool has just become one more voice in the chorus their abuser has been conducting.
This is not an argument against AI detection tools. It is an argument for practitioners understanding the difference between what a tool measures and what it claims to measure, and for building that understanding into the way they introduce tools to clients.
The practitioner’s responsibility in the gap
When a client arrives having already used an AI detection tool and received no result, the practitioner faces a specific clinical situation. The client may present the negative result as evidence. They may use it to explain why they stayed, why they dismissed their own concerns, why they stopped telling people what was happening.
A practitioner who does not understand recall figures, who does not know which task the tool’s accuracy claim applies to, who treats “the AI didn’t find anything” as clinically meaningful, is not equipped to work with that presentation.
The questions a practitioner needs to be able to ask are: what task was this tool trained on? What does its recall figure look like specifically for the behaviour being assessed? How much data did the client give it, and does the tool account for confidence levels based on data volume? What does it mean clinically when a low-recall tool returns nothing?
These are not obscure technical questions. They are the same questions a practitioner would ask about any assessment instrument: what does this tool actually measure, how reliably, and what are the limits of its findings?
The design choices that change the equation
Not all AI detection tools are built to the same standard, and the difference matters more in this context than in almost any other.
A forensic tool designed specifically for coercive control contexts needs, at minimum, to be trained on the correct task rather than a proxy task, to return a confidence level alongside its findings so that the clinician and the person using it understand the limits of the analysis, and to analyse language longitudinally across a body of communication rather than a single message, since that is where the pattern actually lives.
A tool that scores ninety-four percent on emotional tone detection, built for general relationship wellness use, is a different instrument from a forensic tool trained on the documented frameworks for coercive control and returning explicit confidence levels alongside its findings. Both may be described as “AI abuse detection.” They are not doing the same job.
Practitioners who recommend tools to clients have a responsibility to understand which instrument they are recommending, what it was actually built to detect, and what a negative finding from that instrument means and does not mean.
The result that says nothing was found deserves as much clinical attention as the result that says something was.
