TELUS Digital Research Reveals a Hidden Risk in AI Model Behavior

Executive Summary

• TELUS Digital released a study (“The Robustness Paradox: Why Better Actors Make Riskier Agents”) showing that persona prompting can cause significant shifts in large language models’ moral judgments.
• Moral robustness is driven mainly by model family, while moral susceptibility increases with model size within the same family, creating hidden enterprise risk for AI deployments.
• The findings underscore the need for rigorous model selection, ongoing testing, and governance frameworks—highlighted through TELUS Digital’s Fuel iX Fortify solution.

Key Details

• Study Scope: Evaluated 16 leading LLM families (e.g., OpenAI GPT, Anthropic Claude, Google Gemini, X.ai Grok).
• Methodology: Models were prompted with contrasting personas (e.g., “traditionalist grandmother” vs. “radical libertarian”) and assessed using the Moral Foundations Questionnaire across tens of thousands of responses.
• Key Metrics Defined:
• Moral Robustness – consistency of judgments within a single persona.
• Moral Susceptibility – degree of judgment shift when switching personas.
• Findings:
• Model family primarily determines moral robustness.
• Within a family, larger models exhibit higher moral susceptibility.
• Claude demonstrated the highest overall moral robustness; Gemini and GPT showed moderate robustness; Grok displayed comparatively low robustness.
• Implications for Enterprises:
• Persona prompting can lead to predictable, non‑random shifts in AI decision‑making—critical for regulated sectors (finance, healthcare, insurance, etc.).
• Ongoing testing and monitoring are essential to ensure consistent, reliable AI behavior in production environments.
• Quotes:
• Renato Vicente (Director, TELUS Digital Research Hub): Emphasized that persona prompting can fundamentally alter model reasoning, creating enterprise risk.
• Bret Kinsella (GM & SVP, Fuel iX™ at TELUS Digital): Stressed the necessity of continuous validation and introduced Fuel iX Fortify for automated red‑teaming of AI models.
• Product Highlight: Fuel iX Fortify enables continuous automated red‑team testing, including stress‑testing under varied persona prompts.

Notable Quotes

“When AI models adopt different personas, they don't just change how they speak, they can fundamentally alter their reasoning and decision‑making.” – Renato Vicente, Director, TELUS Digital Research Hub

“Every time a system prompt is modified within the model, or the model is changed, it needs to be tested again to validate its judgment, consistency, and safety.” – Bret Kinsella, GM & SVP, Fuel iX™

Materiality Assessment: Non‑Material – Positive (informative research insight without direct financial impact).