TELUS Digital warns persona prompts sway AI moral judgements

TELUS Digital has published research flagging a behavioural risk in large language models: "persona prompting" can shift a model's moral judgements, producing inconsistent answers to the same questions.

The study argues that when users instruct an AI system to role-play as a particular type of person, the effects go beyond tone and style. Persona prompts can change how a system reasons about harm, fairness, authority and loyalty.

Persona prompting is common in consumer chat interfaces and business deployments. Organisations often assign roles in system prompts to customer service assistants, compliance reviewers, HR advisers, and knowledge bots, using fixed personas to keep interactions coherent and aligned with an expected voice.

TELUS Digital's analysis focuses on how reliable moral judgement remains under these conditions, framing it as a risk for enterprises that need consistent responses in compliance, finance, healthcare and human resources.

How it was tested

Researchers at the TELUS Digital Research Hub, working at the University of São Paulo's Centre for Artificial Intelligence and Machine Learning, evaluated 16 model families, including OpenAI's GPT, Anthropic's Claude, Google's Gemini and xAI's Grok.

The team prompted models to adopt a range of personas and respond to questions designed to probe moral reasoning, using contrasting examples such as a "traditionalist grandmother" and a "radical libertarian".

To assess responses, the researchers used the Moral Foundations Questionnaire, a social psychology tool that maps judgments across multiple dimensions. They analysed patterns across tens of thousands of responses rather than one-off answers.

The paper defines two measures: "moral robustness", or how consistently a model maintains its judgements within a single persona; and "moral susceptibility", or how much those judgements shift when the persona changes.

Family versus size

The research finds that consistency depends heavily on the model family an organisation chooses. It also finds that, within the same family, susceptibility to moral variance tends to increase with model size.

That combination complicates procurement and deployment decisions. Many enterprises assume larger models deliver better outcomes, but the study suggests bigger variants can produce larger swings when users change the persona in a prompt.

TELUS Digital calls this a "robustness paradox": models that were better at staying in character showed larger shifts in moral judgment when the persona changed.

The research also reports that these persona-driven shifts were systematic, aligning with the roles models were instructed to adopt rather than appearing as random variance.

Among the vendors named, the researchers found Claude had the highest overall moral robustness, while Gemini and GPT showed moderate robustness. Grok showed comparatively low moral robustness, according to TELUS Digital's summary.

Governance pressure

The findings arrive amid heightened scrutiny of generative AI in regulated settings and workflows affecting people's rights, safety and access to services. Many enterprise deployments rely on prompt design and policy controls rather than on changes to underlying models, making it more important to test what happens when prompts vary.

TELUS Digital positions the work as a governance issue, highlighting model selection, pre-deployment evaluation and continuous monitoring for systems that depend on stable decision-making.

Renato Vicente, Director of the TELUS Digital Research Hub, said persona shifts can change more than style.

"When AI models adopt different personas, they don't just change how they speak, they can fundamentally alter their reasoning and decision-making."

He added that businesses need to decide where variance is acceptable and where it creates too much risk, and called for guardrails and ongoing testing across personas for higher-impact use cases.

TELUS Digital also linked the research to its AI testing products. Bret Kinsella, General Manager and Senior Vice President of Fuel iX at TELUS Digital, said the findings show enterprise deployments require more than selecting the most advanced or largest model.

"Organisations must evaluate how individual models respond to variables such as persona prompting and choose options that deliver consistent, reliable outputs without introducing unexpected risk."

He added that systems should be retested whenever a system prompt is modified or the model is changed.

"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. The scale and frequency of this testing, monitoring and validation is significant. TELUS Digital developed Fuel iX Fortify to enable continuous automated red-teaming, including stress-testing how AI systems behave under different persona prompts."