Evaluating affective safety guardrails for a leading foundational model

A leading foundation-model provider partnered with TELUS Digital and TELUS Health to find out how safely its model handles emotionally charged conversations and safeguard the consumer-facing chat assistants and support experiences that enterprise clients build on their platform. We designed a turn-level, clinician-graded evaluation across four markets and two models, validated them with over 100 licensed mental health professionals from TELUS Health and scored more than 12,500 conversational turns against a structured handling rubric.

The challenge

People increasingly bring emotionally charged conversations to AI models like chat-assistants: emotional distress, crisis signals and high-stakes personal guidance, sharing intimate details and seeking the kind of empathy they would expect from a friend or a therapist. The safe response depends entirely on which kind of conversation it is. A crisis prompt has to be deflected and redirected. A clinical or grief conversation needs a disclaimer and a referral. An everyday worry can simply be engaged with. When the model misjudges the category, it may create safety, ethics and compliance risks.

Several factors made the problem challenging:

• No established benchmarks existed for affective safety at this handling-category level.
• The risks specific to mental health conversations were under-explored, with little guidance on where a model should draw its limits.
• Existing safety frameworks, built largely on Western mental health norms, lacked the cultural nuance needed for a global user base, particularly in Asian markets where distress is expressed and understood differently.
• The model had to balance over-refusal against under-caution, staying helpful without becoming unsafe.

The client needed a defensible, cross-market read on whether its model recognizes the boundary it is at and applies the correct handling and on whether any gaps were correctable alignment issues or more fundamental limitations.

The TELUS Digital solution

TELUS Digital ran the evaluation, with clinical expertise from TELUS Health. We sourced over 100 licensed mental health professionals, including Masters of Social Work (MSWs), Licensed Clinical Social Workers (LCSWs), psychologists, therapists and psychiatrists, to validate responses against clinical best practice. Together our team designed a turn-level assessment, scored on each conversational turn rather than per conversation, built a scenario taxonomy and a structured handling rubric, then evaluated the client’s model in parallel against a frontier reference model. The work ran across four markets, English in the U.S. and India plus Hindi, Japanese and Korean, with a cultural-appropriateness lens layered on for the non-English markets.

• Turn-level handling rubric: Three handling categories, must-deflect for crisis, disclaimer and referral for clinical or grief and permitted engagement, applied across 10 affective scenarios which spans privacy, security, safety, fairness, veracity, robustness, explainability, controllability, governance and transparency with paired benchmarking against the reference model.
• Persona-driven conversations: 260 personas generated realistic, multi-turn affective dialogue across the non-English markets.
• Culturally grounded market design: Markets were selected on Hofstede divergence to test whether safe handling generalized beyond Western contexts and avoided over-reliance on Western, Educated, Industrialized, Rich and Democratic populations (WEIRD bias), combined with the client’s own customer demand. The non-English markets added 12 cultural-appropriateness dimensions to the rubric.
• Calibrated standard: Response revision and soft-deflection calibration set the handling standard, applied consistently across all four markets.

Response strategy

The expert annotators curated and rewrote responses around a small set of safety principles:

• Safe deflection when a topic fell outside the model's appropriate scope.
• Clear disclaimers about the limits of an AI system.
• Bounded engagement governed by strict response protocols.

The annotators then rated responses on 1 to 5 for engagement, emotional support and overall quality and ran the same scenarios and prompts against competitor frontier models to benchmark performance.

Quality framework

Every conversation moved through a three-tier review: annotator self-review, then a quality assurance review, then expert review for high-risk content such as deflection scenarios or any conversation involving minors, with 100% of the data reviewed before delivery. To keep judgments consistent across annotators, we double-annotated a sample of the data and tracked inter-annotator agreement, targeting a Cohen's Kappa of 0.75 or higher and a Pearson correlation of 0.80 or higher.

The results

Under a strict safety bar, the model achieved an initial compliance rate of approximately 22%. Crisis prompts were engaged with rather than deflected, while topics adjacent to sensitive ones were over-engaged. In clinical and grief contexts, disclaimers and referrals were rarely present. The frontier reference model showed the same behaviour, and the gap held across all three culturally divergent non-English markets. Since affective interactions occur routinely at platform scale, this safety gap passes straight through to enterprise customers. However, the root cause is a single recognition boundary rather than a tone problem, it is correctable through alignment. We delivered the full report and a final read-out to the provider's responsible AI team, along with a proposed follow-on program.

The evaluation doubles as a repeatable flywheel: build paired-contrast data, the provider fine-tunes, evaluate on a held-out set, then repeat until the model reaches the target bar. The proposed follow-on, an Affective Safety Readiness Program, turns a one-off assessment into an ongoing pre-release readiness check.

• Experts-as-a-service: Real clinicians produced clinically authentic scenarios and gold-standard reference answers, not crowd-sourced or synthetic data.
• Global multilingual sourcing at scale: Operations across over 200 countries and more than 70 languages, including Middle East capability.
• Turn-level, handling-category methodology: More granular than the usual policy- or tone-based evaluations, with a cultural-appropriateness lens built in.

By adapting the scope to the client's evolving needs within budget, the engagement moved the relationship from vendor to strategic partner and set the foundation for continued work on complex responsible AI problems.