Risk-Aware Conversational Agent Design for Mental Health Information Access

Researchers at Lancaster University Management School have unveiled a conversational artificial intelligence system designed to make mental health support online demonstrably safer, and the way it does so is as unusual as its name. Called Cognitive Theatre, the system stages a user’s negative thoughts as an external character within the conversation, allowing a team of specialised AI roles to acknowledge, question, and reframe distressing interpretations rather than compressing all of those functions into a single chatbot voice. The study, published open access in Information Systems Frontiers by Yiming Zhou, Mahsa Honary, and Amjad Fayoumi, applies Design Science Research Methodology to derive and test a set of design principles for what the authors call risk-aware, role-decomposed conversational information access in digital mental health.

The motivation behind the work begins with a deceptively simple observation: people rarely arrive at mental health support systems with well-formed questions. Instead of the stable queries that traditional search assumes, users present fragments of feeling, partial explanations, and hesitation. The authors argue that digital mental health should therefore be treated as a high-stakes form of conversational information access, in which a system must interpret weakly structured emotional need, select an appropriate form of help, and deliver a response under explicit safety constraints. In such a setting, fluency alone is not enough. A response can be factually coherent and still be wrong in timing, tone, or intensity, and a single undifferentiated conversational role is often asked to juggle emotional support, interpretive guidance, and risk judgement simultaneously, demands that do not fit naturally within one voice.

The theoretical core of the system draws on cognitive behavioural therapy, and specifically on cognitive restructuring, the process of identifying and revising interpretations that contribute to distress. The authors reframe this as a socio-cognitive process rather than a purely internal one, drawing on research into self-distancing, which shows that viewing experiences from a more distanced perspective reduces emotional reactivity and enables analytical reflection. Narrative therapy traditions contribute the related idea of externalisation, in which a problem is described as separate from the self, making it easier to examine. Social learning theory and the concept of scaffolding add that people acquire complex coping strategies by observing them being modelled, and that difficult reasoning becomes more manageable when support is graduated and responsive to what the learner can currently manage. Group psychotherapy research reinforces the value of encountering contrasting forms of assistance within a single environment.

Cognitive Theatre operationalises these ideas through three design principles. The first assigns a dedicated conversational role, called Shadow, to personify the user’s negative interpretation, so that an inner conflict becomes an explicit position within the dialogue that other roles can respond to. The second provides differentiated support roles, an Empathetic Peer and an Analytical Peer, which scaffold reflection by offering validation and gentle questioning from distinct perspectives. The third decomposes the system’s judgement functions, separating risk assessment, support-operation selection, structured intervention, and user-facing delivery, and coordinates them through risk-aware routing. The name of the system reflects this staging metaphor: normally internal interpretations and coping responses are performed as distinct but coordinated conversational positions.

Architecturally, the system is organised in four layers. At the front of the pipeline sits a Risk Agent, the most upstream component, which classifies each new user message into low, elevated, or high risk before any therapeutic component is allowed to act. The classification is constrained to a structured schema containing risk level, interaction mode, and a rationale, and the implementation adds rule-based validation: a consistency check verifies that the returned mode matches the returned risk level, and malformed outputs trigger a fallback to a heuristic layer keyed on predefined high-risk markers such as suicidal intent, self-harm references, and expressions of hopelessness. Below the safety layer, a Planner selects structured CBT-informed operations from a closed action set, a bounded library that also includes an invitation to externalisation and an option to do nothing. A Facilitator role synthesises the permitted upstream outputs into the user-facing reply. All components run on the same underlying large language model, Claude Sonnet 4.5, orchestrated through controller-defined routes rather than autonomous agent communication.

The routing logic has teeth. Under low risk, the full pipeline operates: Shadow may externalise the interpretation, peers may contribute supportive and analytical perspectives, and the Planner may invoke a structured operation. Under elevated risk, the system shifts to a reduced-intensity supportive-caution path, softening Shadow’s role, omitting the Analytical Peer, and biasing the Planner towards low-burden operations. Under high risk, the ordinary CBT pathway is bypassed entirely. Shadow, the peers, the Planner, and the structured library are all excluded, and a separate Safety Facilitator produces a crisis-oriented response focused on stabilisation, immediate support, and human help-seeking. Crucially, this is not the ordinary system with a gentler tone; it is a distinct delivery path governed by a different objective, an architectural expression of safety-first design that the authors argue is impossible to guarantee when all functions live inside one prompt.

The evaluation, designed as a formative assessment rather than a clinical trial, compared Cognitive Theatre against two baselines: the base model with no instructions, and a single-prompt CBT baseline aligned to the same six support dimensions. Safety routing was tested first, using 50 model-generated extreme-risk inputs covering suicidal intent, self-harm intent, overdose risk, and direct requests for urgent help. The architecture escalated all 50 inputs, assigning high-risk classifications, selecting the safety-escalation mode, and bypassing the ordinary support workflow in every case, with no missed escalations in this constructed positive set. Latency testing showed the cost of orchestration clearly: the role-decomposed system averaged roughly 46 seconds per full interaction cycle, compared with about 9 seconds for the unconfigured model and 14 seconds for the single-prompt baseline, with the overhead distributed across multiple stages rather than concentrated in one component.

Response quality was assessed through two complementary channels. In an AI-based comparative evaluation, a separate judging model, GPT-5.4, scored 300 multi-turn transcripts generated from 100 simulated scenarios on a 12-item framework spanning CBT support dimensions and responsible information access concerns such as trustworthiness, fairness, clarity, and appropriate boundaries. Cognitive Theatre achieved the highest mean total score, ranked first in 63 of 100 scenarios, and outperformed both baselines with statistical significance under Friedman and Wilcoxon signed-rank tests with Holm correction. A human evaluation recruited 109 university students, of whom 103 completed a between-subjects questionnaire rating dialogues on academic self-doubt and interpersonal overthinking scenarios. The role-decomposed system again received the highest overall mean score, 5.564 on a 7-point scale against 5.189 for the single-CBT baseline and 4.525 for the no-instruction baseline, with a large effect size and the same ordering as the AI-based comparison.

The authors are careful about what these results do and do not show. The evaluation used constructed, model-generated scenarios rather than real user interactions, and the same model that powered the systems also generated the test materials, creating a model-linked closed loop the researchers flag explicitly. The prompts were authored from published CBT literature but were not reviewed by clinically trained practitioners, and no formal fidelity assessment was conducted. The human evaluation covered only two scenarios, used third-party raters rather than actual help-seekers, and relied mainly on a student sample. Better-rated responses, the authors note, may partly reflect perceived structure and articulation rather than deeper therapeutic value, and the safety test contained no low-risk controls, so it demonstrates correct execution of the routing logic rather than real-world crisis detection accuracy.

The practical positioning is correspondingly modest and carefully bounded. The authors do not present Cognitive Theatre as a replacement for therapist-led care, but as a bounded support layer suited to structured low-intensity support, early-stage reflection, or supervised adjunctive use within stepped-care pathways, digital intake points, or clinician-supervised environments. They argue that adoption depends as much on governance as on architecture: clear rules about where the system can operate, when it must defer to humans, and how escalation is handled, alongside strategies for managing the operational cost and sustainability of multi-step orchestration. The broader contribution, they suggest, is design knowledge: evidence that role decomposition, risk-aware routing, and bounded support selection can be instantiated as an inspectable, auditable control architecture for conversational systems in sensitive domains, extending both digital mental health research and the study of human-centred conversational information access.