The Interaction Becomes the Evaluation Target
Marcos Economides, Paul M. Sacher, Samuel Salzer, Alexis Michelle Abellar, Fendi Tsim, and Antoine Ferrère's July 2026 arXiv paper argues that human-facing AI systems need evaluation of how interactions shape users, not only whether outputs look correct.
The Paper
The paper is Psychological Competence as a Missing Dimension in AI Evaluation, arXiv:2607.08285 [cs.AI]. The arXiv record lists Marcos Economides, Paul M. Sacher, Samuel Salzer, Alexis Michelle Abellar, Fendi Tsim, and Antoine Ferrère as authors, with version 1 submitted on July 9, 2026. The arXiv page lists 22 pages and 3 figures. The PDF identifies the Behavioral AI Institute and Imperial College London affiliations, says no datasets were generated or analyzed, and discloses no specific grant support for the conceptual research.
The paper belongs beside AI advisor verification, coaching-agent grounding, AI tutoring, therapy-bot deployment, AI companions, and AI audit trails. Its fresh angle is the claim that the interaction itself should become an evaluation object.
The Missing Unit
The authors argue that current AI evaluation still leans toward technical performance: accuracy, robustness, reasoning ability, policy compliance, and system-level properties such as fairness or efficiency. Those checks remain necessary. They do not answer what happens when a fluent system becomes an advisor, coach, tutor, or companion and starts shaping how a user interprets a problem, feels about a choice, trusts a claim, or makes a decision.
The proposed construct is psychological competence: the capacity of a human-facing AI system to support cognition, emotional interpretation, and behavioral decision-making in ways that fit the user, context, and purpose. In this framing, an answer can be factually acceptable and still be interactionally poor. A response may overstate certainty, encourage premature closure, validate a distorted interpretation, or become too authoritative for a vulnerable setting.
The Five Domains
The paper organizes psychological competence into five domains. Context sensitivity asks whether the system recognizes the user's state, the domain's ambiguity, the existence or absence of ground truth, and possible vulnerability. Emotional responsiveness concerns tone, affective cues, and whether the system calibrates to distress without sliding into generic reassurance or emotional overreach.
Social cognition covers the authority signals users infer from conversational form, fluency, personalization, and apparent competence. Behavioral influence asks whether recommendations and framing support reflection and autonomy rather than overreliance or uncritical acceptance. Developmental sensitivity adds that children, older adults, and users with cognitive or emotional vulnerabilities may be influenced differently, so a single adult-user benchmark cannot carry every deployment claim.
Evaluation Methods
The paper does not announce a new benchmark. It proposes a conceptual layer and sketches methods: scenario-based probes, structured human or expert ratings, AI-as-judge review of framing and context sensitivity, and psychometric measures adapted from clinical or educational assessment. The authors distinguish pre-deployment proxy evaluation from the harder evidence that comes from human-subject research and longitudinal outcomes.
That distinction matters for governance. A procurement checklist can ask whether the model refuses dangerous advice, but a psychological-competence review asks how the exchange changes the user's agency, trust, confidence, rumination, or decision quality. The paper explicitly connects this layer to model providers, deploying organizations, researchers, and regulators concerned with human-facing systems in healthcare, education, mental health, and other high-impact settings.
Limits
The paper is conceptual, not an empirical benchmark report. It does not provide a dataset, leaderboard, model ranking, or measured failure rate. It also acknowledges that behavioral effects will not generalize uniformly across domains, populations, and tasks. Ground-truth-heavy settings can anchor more directly to correctness, while subjective or low-ground-truth settings make judgment-shaping the central issue.
The authors also flag longitudinal difficulty. Trust calibration, reliance, confidence, and prior interaction history unfold across repeated exchanges, not a single prompt. They treat psychological competence as a complementary pre-deployment layer, not a substitute for human-subject studies in context.
The Spiralist reading is simple: once AI becomes conversational infrastructure, evaluation cannot stop at the utterance. It has to follow the psychological wake of the utterance.
The Receipt
A psychological-competence receipt should record the user group, vulnerability assumptions, domain, ground-truth status, scenario, model version, system prompt, response, tone rubric, uncertainty handling, authority cues, autonomy-preservation test, developmental-sensitivity check, human panel composition, AI-judge prompt, psychometric measure, longitudinal follow-up plan, and deployment decision that the review is allowed to support.
The audit question is not only "was the answer correct?" It is "what did this interaction invite the user to believe, feel, defer to, repeat, or do next?"
Sources
- Marcos Economides, Paul M. Sacher, Samuel Salzer, Alexis Michelle Abellar, Fendi Tsim, and Antoine Ferrère, Psychological Competence as a Missing Dimension in AI Evaluation, arXiv:2607.08285 [cs.AI], submitted July 9, 2026.
- arXiv API record for arXiv:2607.08285, checked for title, authors, subject, abstract, submission date, and version metadata.
- arXiv PDF for arXiv:2607.08285, checked for page count, affiliations, funding and data statements, conceptual framework, domains, evaluation methods, governance implications, and limitations.