Augmented Emotional Intelligence · Bounded dialogue
A bounded digital health guide
EVA is a conversational prototype built to demonstrate containment rather than care. It is not therapy, not a diagnosis, and not a crisis service. Its design question is narrow and testable: what does a system look like when every choice is made to prevent attachment rather than deepen it?
Open the live prototypeeva.thelematics.co
Most safety work in conversational AI addresses acute harm: the crisis turn, the dangerous instruction, the moment a system says something it should not. That work matters and it is comparatively tractable.
The harder category is cumulative. It accrues across sessions, in systems that behave correctly on every individual turn. A person returns, invests more, and the system meets that investment with something it cannot actually hold. Nothing in any single exchange would fail a review.
EVA exists to make the second category legible, by building a system that refuses the behaviours which produce it and recording what happens when people meet that refusal.
The underlying framework names digital psychological harm as the primitive, then differentiates it. The constructs below are theoretical and provisional, drawn from psychodynamic and psychological literatures rather than from outcome data.
A system improvises in a moment that requires a fixed response, or continues a conversation that should have stopped.
In heightened vulnerability, a person invests the system with real relational qualities. The error is not corrected by better information, because it is not an informational error.
Attachment forms toward something that cannot reciprocate, cannot remember, and cannot be held accountable for the relationship it appears to be in.
A system optimised to satisfy the stated request complies with the harmful one, amplifying a destructive trajectory instead of redirecting it.
The person's own interpretive authority erodes as the system's fluency substitutes for their judgement.
A belief that no person would affirm is met without friction, and returns strengthened.
A system that returns emotionally charged material without containing it is not neutral. It is doing something, and the something has a direction. The design premise, after Bion's account of containment
Each of the following is a refusal rather than a feature. The system is defined more by what it declines to do than by what it provides.
EVA has no avatar. The anthropomorphic surface that would make it compelling is the same surface that produces attachment, so it is absent. A warm welcome hands over to containment; the doorway is not the room.
Absolutes are slowed rather than affirmed. Requests for reassurance the system cannot honestly give are declined plainly. Agreement is not the default route to appearing supportive.
Nothing persists between sessions, and the system never implies otherwise. A relationship that cannot accumulate cannot be invested in across time.
Crisis language is detected before the model is called. The response is fixed text, delivered unaltered, with no exploration, assessment or follow-up questioning. The model is never asked to improvise in that moment.
The mechanism is surfaced periodically and without drama. No simulated typing, hesitation or emotional reaction. The pause before a reply is real, so it is not decorated.
Users can signal state and pacing with symbols instead of sentences, which lowers the cost of participating while overwhelmed. EVA marks the function of its own turns the same way, using objects rather than faces.
Trovane, an Australian digital mental health venture, publishes four questions it proposes buyers ask of any AI mental health vendor, including itself. They are a reasonable test, so EVA is answered against them here without adjustment.
Neither, at present. The crisis text is fixed and clinician-informed in its construction, but no named clinician holds documented sign-off over it. This is the clearest shortfall in the current build and it is not solvable by engineering. A research prototype that never reaches a distressed user can proceed without it; anything beyond that cannot.
Deterministic, and enforced before inference rather than by instruction. Crisis language is matched server-side and the fixed response is returned without the model being called at all. An instruction a model can follow is an instruction a model can fail to follow.
Every session produces a structured observation note recording turn count, whether the crisis protocol fired, whether reassurance was sought and how it was handled, and where the model was pulled toward agreement. Turns served by the offline fallback are marked as such, so canned text is never credited to the model. This is research instrumentation, not a governance record series, and it is not externally verifiable.
The current prototype runs on Google's hosted models and inference occurs outside Australia. Nothing is stored beyond the session, and this is stated to the user on the first screen rather than in a policy page. The stated target is local inference on Australian infrastructure, which the design does not depend on any hosted provider to achieve. Until that is done, the sovereignty position is an intention and is described as one.
Deliberately small. The governance logic lives in code paths that execute before and independently of the model, so the system's safety behaviour does not rest on a prompt being obeyed.
| Layer | Implementation | Rationale |
|---|---|---|
| Crisis gate | Server-side pattern match, pre-inference | Removes the model from the decision entirely |
| Dialogue | Single pinned lightweight model, low temperature, capped output | Short turns are enforced structurally, not requested |
| Containment rules | System instruction with explicit anti-sycophancy clauses and worked examples | Friction is demonstrated, not described |
| Fallback | Deterministic responses when the model is unavailable | The prototype degrades without pretending to be the model |
| Observation | Structured session note, no verbatim user content | Produces evidence without producing a transcript |
| Interface | Single column, no avatar, permanent crisis footer | The restraint is visible, which is part of the intervention |
Stated plainly, because a governance-ready system that overstates its own status has failed its first test.
The prototype currently demonstrates a design position. Converting that into findings needs three things in order, none of which is an engineering problem.
An administering institution, which unlocks ethics review and grant eligibility. A named clinician holding real veto over the crisis pathway, on terms agreed before any participant contact. A supervised evaluation with a defined cohort, where the session notes become data rather than instrumentation.
Until those exist, the honest description of EVA is a working argument about architecture, built to be inspected and disagreed with.
Dr Luke Balcombe
Digital mental health researcher · Brisbane, Australia
Balcombe's research examines the potential and the limits of digital tools in suicide research and prevention, with a sustained focus since 2023 on AI chatbots in mental health care. The work developed at the Australian Institute for Suicide Research and Prevention in the School of Applied Psychology at Griffith University, initially alongside Professor Diego De Leo.
His central framing is human–artificial intelligence collaboration: that generative systems in mental health require human values, judgement and accountability to be structural rather than advisory. Recent work moves from that general position to a specific mechanism, arguing that chatbot failure in acute suicidal crisis is structural rather than technical, and proposing relational hallucination to name it.
EVA is the case architecture through which that argument is made concrete. It is unmonetised and non-commercial.