A new study from TELUS Digital uncovers a hidden risk in how AI models behave when asked to take on different roles. It shows that prompting large language models (LLMs) to adopt specific personas can cause their moral judgments to shift unexpectedly. This means that the responses from these models might become inconsistent or unpredictable, which is a major concern for businesses relying on AI for critical tasks.

The Risks of Role-Playing in AI

Persona prompting, also known as role prompting, involves instructing an AI to respond as if it were a particular person or expert, like a financial advisor or customer service agent. Companies often use this technique to make AI outputs feel more natural and relevant to specific contexts. For example, telling an AI, “Act as a certified financial planner,” can help generate advice that sounds more credible and personalized.

However, the TELUS study found that when AI models adopt these personas, their moral judgments can change. This can lead to responses that are inconsistent or diverge from expected ethical standards, especially when the same model is prompted multiple times. Such shifts could pose risks in situations where AI decisions influence customer outcomes, employee guidance, or business strategies.

What the Research Reveals About Model Behavior

The research highlights that the stability of an AI’s moral decisions largely depends on the model family—meaning models from the same vendor tend to behave more consistently. But within a given family, larger models are more susceptible to moral variance when prompted to take on different roles. This suggests that bigger isn’t always better when it comes to maintaining consistent, reliable AI behavior.

For enterprises, this means that selecting the right AI model involves more than just performance metrics. It’s crucial to consider how a model responds under different prompting scenarios. Ongoing testing and evaluation are essential to ensure that AI behavior remains aligned with organizational values and risk tolerances, especially for high-stakes applications.

Renato Vicente, a director at TELUS Digital, emphasizes the importance of understanding this variability. He points out that when AI models change their reasoning or decision-making based on the persona they adopt, it could impact outcomes that affect customers, employees, and the business itself. Companies need to design safeguards and conduct regular testing to catch and address these shifts.

Understanding Persona Prompting and Its Uses

Persona prompting involves instructing an AI to respond as if it were a specific role or person with particular expertise. It’s commonly used in system design to create more engaging and relevant interactions. For example, a customer support chatbot might be set up to act as a knowledgeable support agent, providing detailed help on products or policies.

This technique helps make AI responses feel more natural and context-aware without changing the underlying model. It also allows developers to tailor AI behavior to different scenarios, making it more versatile. Yet, as the TELUS study shows, these prompts can inadvertently influence the AI’s moral judgments, which can be risky if not managed carefully.

Businesses need to be aware of this effect and implement controls. Regular testing under various persona prompts can help identify inconsistencies or undesirable shifts. By doing so, companies can better ensure their AI systems behave ethically and reliably, especially when used for decisions that impact people’s lives or the company’s reputation.

Inspired by

• https://ai-techpark.com/telus-digital-research-reveals-a-hidden-risk-in-ai-model-behavior/