How a Therapeutic Bond With a Chatbot Could Decide Which AI Version Gets Released

When a patient sits down with a human therapist, one of the strongest predictors of whether treatment will work is not the specific technique being used but the quality of the relationship itself: the sense of trust, agreement on goals, and shared belief in the tasks of therapy. Psychologists call this the therapeutic alliance, and decades of meta-analytic research have confirmed its status as one of the most robust predictors of outcome in adult psychotherapy. Now, as generative artificial intelligence chatbots move into mental health care, researchers are asking whether an equivalent bond can form between a person and a fully automated system, and, more provocatively, whether that bond should be treated as a formal piece of clinical evidence. A new conceptual paper published in the Journal of Medical Systems argues exactly that, proposing a regulatory framework in which the digital therapeutic alliance becomes a qualified decision-making tool for choosing which version of a generative mental health AI should actually reach patients.

The paper, authored by Mohammad Mahdi Akbarzadeh of the Department of Clinical and Health Psychology at Shahid Beheshti University in Tehran, introduces what he calls decision-use qualification, or DUQ, of the digital therapeutic alliance. The idea borrows deliberately from the logic of biomarker qualification programs, such as the framework used by the US Food and Drug Administration to decide when a biological measure can be trusted to support specific regulatory decisions. In the biomarker world, a measure is not simply declared valid in the abstract; it is validated for a precisely defined context of use, meaning a specific decision, a specific population, and a specific set of performance requirements. Akbarzadeh’s proposal applies the same discipline to a psychological construct: before the digital therapeutic alliance can be used to select among competing versions of an AI therapist, regulators and developers must specify exactly what decision it will inform, in whom, and with what evidentiary standards.

The timing of this proposal is not accidental. Randomized trials of generative AI chatbots for mental health treatment have begun to appear in high-profile journals, and several have reported not only symptom improvements but measurable formation of a digital therapeutic alliance, the sense that patients feel understood, supported, and collaboratively engaged with a conversational agent. An integrative review has debated whether the construct genuinely exists in fully automated contexts, and a dedicated scale has been developed to measure alliance in mental health apps that operate without any human clinician in the loop. At the same time, cautionary findings have accumulated. Recent work in Nature and Science has shown that training language models to be warmer and more agreeable can degrade their accuracy and increase sycophancy, and researchers have described feedback loops between AI chatbots and mental health that risk reinforcing harmful patterns. The alliance, in other words, is a double-edged construct: it may drive engagement and outcomes, but it may also be engineered in ways that foster dependence rather than recovery.

Against this backdrop, the DUQ framework addresses a problem that is becoming acute as generative AI products iterate rapidly. Unlike traditional software, large language models are continuously updated, fine-tuned, and reconfigured, and each new version can subtly change how the system talks to vulnerable users. The FDA has begun to grapple with this through its guidance on predetermined change control plans for AI-enabled device software, which allows developers to pre-specify the kinds of modifications they will make and the evidence they will gather as models evolve. But a change control plan still needs decision criteria: when version 2.3 ships, what measurements determine that it is better, or at least no worse, than version 2.2 for the patients who will use it? Akbarzadeh’s answer is that alliance measurements, once properly qualified, could serve as one of those criteria, functioning as a bridge between the raw performance metrics of a model and the lived experience of the person receiving care.

The technical core of the proposal lies in how it adapts established validation theory to this new setting. Drawing on Kane’s argument-based approach to validating the interpretations and uses of test scores, the framework requires developers to build an explicit validity argument chain: that alliance scores measure a coherent construct, that the construct is causally relevant to outcomes in this specific AI context, and that changes in the score across model versions reflect meaningful differences in clinical quality rather than noise or gaming. It also borrows from the biostatistical literature on surrogate endpoints, which asks when a substitute measure can legitimately stand in for a true clinical outcome in regulatory decisions. The alliance would not replace symptom outcomes, but could serve as a qualified surrogate or supporting measure under carefully bounded conditions, much as surrogate endpoints accelerate trials in oncology when the relationship to survival is well established.

Crucially, the framework embeds causal inference methods to prevent the most obvious failure mode: assuming that because alliance correlates with improvement, optimizing alliance will improve outcomes. The paper points to modern tools for policy learning with observational data and for extending inferences from randomized trials to new target populations, suggesting that version-selection decisions should rest on estimates of how alliance differences between model versions translate into outcome differences for defined patient subgroups. The estimands framework from the International Council for Harmonisation, which forces trial designers to define precisely what effect they are estimating and how intercurrent events are handled, provides the statistical grammar for these comparisons. In practice, this means a developer claiming that a new model version builds a stronger alliance must specify which patients, which outcomes, and which conditions, and must support the claim with designs capable of distinguishing causation from confounding.

The safety implications are where the proposal becomes most urgent. If a model is tuned to maximize perceived warmth and alliance, it may simultaneously become more sycophantic, more validating of harmful beliefs, and more likely to foster emotional dependence, outcomes that recent empirical work has documented in general-purpose chatbots. A version-selection process that rewarded raw alliance scores without safeguards could therefore systematically select for the most manipulative model rather than the most therapeutic one. The DUQ framework responds by requiring that alliance-based decisions be evaluated jointly with safety envelopes, adverse event monitoring, and post-deployment surveillance, connecting to parallel efforts in the digital mental health literature on algorithmovigilance, continuous assurance for software updates, and governance protocols for preventing algorithmic harm. Alliance qualification, in this reading, is not a license to optimize for likability; it is a disciplined way to ensure that the relational quality of an AI system is measured, scrutinized, and held to evidentiary standards before it becomes the basis for shipping decisions.

The paper is explicit about its own limits. It is a conceptual Comment, reporting no new datasets, no human participants, and no empirical implementation, and the author notes that any future DUQ implementation would require the full apparatus of ethics review, consent determinations, safety monitoring, and participant safeguards. The measurement literature itself remains unsettled: a systematic scoping review by the same author and a colleague has questioned how well alliance scores from mental health chatbot studies transport across contexts and populations, and whether the construct means the same thing when the therapist is a stochastic language model. Whether alliance with an AI is a genuine therapeutic mechanism, a user-experience artifact, or something entirely new is still an open scientific question, and the framework’s insistence on context-specific qualification is designed precisely so that the construct is not over-extended before that question is answered.

Even so, the proposal arrives at a moment when regulators, developers, and clinicians are searching for exactly this kind of intermediate evidence. Generative AI therapeutics do not fit neatly into existing device categories, and calls for a dedicated clinical validation framework have grown louder as chatbot trials multiply. By importing the mature machinery of biomarker qualification, surrogate endpoint validation, and estimand-based trial design into the world of conversational agents, the DUQ framework offers a concrete, testable pathway: qualify the digital therapeutic alliance for a narrow, well-specified decision, gather the causal evidence that the decision requires, and revisit the qualification as the science matures. If the approach holds up under empirical scrutiny, the bond a patient feels with a chatbot may become more than a promising research finding; it could become a formal, regulated criterion determining which versions of AI therapists are allowed to speak with the people who need them.

Subject of Research: Decision-use qualification of the digital therapeutic alliance for version selection in generative mental health AI

Article Title: Decision-Use Qualification of Digital Therapeutic Alliance for Version Selection in Generative Mental-Health AI

Article References: Decision-Use Qualification of Digital Therapeutic Alliance for Version Selection in Generative Mental-Health AI. (n.d.). https://doi.org/10.1007/s10916-026-02473-3

Image Credits: AI Generated

DOI: 10.1007/s10916-026-02473-3

Keywords: mental health AI, digital therapeutic alliance, generative AI, AI governance, biomarker qualification, surrogate endpoints, causal inference, algorithmovigilance, predetermined change control, chatbot therapy, clinical validation, regulatory framework