Detecting emotional change, distress and gradual manipulation early enough to act on it. UK patent application pending.
Most safety systems classify a conversation, or a message, as safe or unsafe. That framing misses the way real harm actually develops, which is gradual. Each individual message looks acceptable; the trajectory across fifty of them does not. Classifying at the message level cannot see a trajectory, and classifying at the conversation level sees it far too late.
Our approach analyses individual chunks of a conversation for emotional signal, and tracks how that signal moves over time. What is being detected is not a forbidden word but a direction of travel: withdrawal, escalation, dependency, distress, and the slow reframing that characterises manipulation.
The system can therefore raise a concern about a conversation in which no single message would ever be flagged.
A UK patent application has been filed and is pending on the Emotional Watcher system, which performs this chunk-level emotional and manipulation detection and surfaces it to a supervising layer. Intellectual property is held by Mofy AI Ltd. See Patents.
The applied case we care most about is a household. A parent should be able to decide what is worth being told about, receive it as a plain-language concern rather than a raw log, and know that the system is watching for a trajectory rather than scanning for banned words.
The aim is not to keep AI away from children. It is to make it possible to bring AI into a family at all.
Emotional signal reads differently across people, and a system calibrated on a neurotypical average will misread a large minority of its users in both directions. Calibration across differing communication styles is an open and central part of this programme.
Deployed as the Guardian Watcher layer in Ask Zai.