AI personas: useful prompt tool or troublesome avatar?

At our recent AI Knowledge Café, what began as a conversation about AI personas quickly became a discussion about something much bigger.

The word persona turned out to mean different things to different people. For some, it was a prompt technique: asking Artificial Intelligence (AI) to respond as a librarian, communications manager, finance director or expert patient reviewer. For others, it meant the user personas long used in service design and digital transformation. Some participants were thinking about avatars and chatbots, while others were discussing AI companions, simulated patients and autonomous agents. 

What became clear is that we are often using the same word to describe very different things.

One useful distinction emerged. User personas represent people. AI personas shape behaviour. The first should be grounded in genuine user research. The second is really a set of instructions that influences how an AI responds. Seen in that light, a persona is less like creating a digital person and more like creating a framework for a conversation. 

That distinction matters because AI can feel surprisingly human. Participants reflected on how confident language, structured responses and conversational warmth can create an impression of expertise and trustworthiness. Yet a polished response is not necessarily an accurate one. Several colleagues suggested that terms such as response framework or role frame may be more helpful than persona, precisely because they remind us that we are shaping a tool rather than creating a person. 

There is a relationship between tone and persona.

A number of participants described using prompts such as “act as a librarian” or “act as a director”. However, the role itself is only part of what shapes the response. Tone often has a much bigger influence.

Take a simple request such as describing a new library service.

An executive-focused response might say:

“The service expands access to evidence resources, supports organisational knowledge mobilisation, and contributes to digital transformation objectives.”

A supportive educator might write:

“If you’re short on time, the service makes it easier to access evidence and resources wherever you are, helping you find information when you need it most.”

Both responses describe the same service. The facts have not changed. What changes is the persona that the audience experiences. One feels strategic and outcome-focused. The other feels supportive, practical and user-centred. Neither is better. They simply meet different needs. The discussion highlighted that tone is often the mechanism through which a persona is created. 

This led to a wider conversation about how we design AI agents. Effective personas are not really about pretending to be someone. They are about combining purpose, audience, expertise and tone to create consistent, useful outputs.

For example, a Knowledge and Library Services agent might not simply be instructed to “act as a librarian”. Instead, it might be given a clear purpose: helping NHS staff find, understand and apply evidence. Its tone might be professional, supportive and evidence-based. Its skills could include literature searching, evidence summaries, critical appraisal and health literacy adaptation. It might be instructed to explain uncertainty, distinguish facts from recommendations, adapt language for different audiences and provide sources. Equally important are the boundaries: it should not diagnose, provide clinical advice or replace professional judgement. 

Seen this way, a persona becomes less about personality and more about transparency. Users can understand what the agent is for, how it behaves and where its limits sit.

Much of the discussion focused on where this approach could be useful. Participants described AI helping people think, remember, organise, translate, plan and reduce cognitive load. Examples included adapting evidence summaries for expert patients, supporting neuroinclusive communication, generating board-level briefing questions, helping students revise and testing services from different user perspectives. 

For those of us working in Knowledge and Library Services, this felt familiar. Much of our work is about helping people navigate information, understand evidence and make informed decisions. In many respects, personas simply provide another way of adapting information to meet the needs of a particular audience. The goal is not to change the evidence, but to make it easier for people to engage with it. 

Alongside the opportunities, there was also considerable caution. The group repeatedly returned to the question of trust. What happens when people begin to experience AI systems as companions rather than tools? Concerns were raised about vulnerable users, emotional attachment and over-reliance on systems that may appear understanding without actually understanding. Participants agreed that there remains an important boundary between support and substitution. AI may assist with thinking, remembering and organising, but it should not replace professional judgement, accountability or human relationships. 

The conversation also moved into agentic AI: systems capable of planning and taking actions as well as generating responses. Here the discussion shifted from prompting techniques to governance. If future systems are making decisions, accessing services, sending information or acting on behalf of users, questions about monitoring, oversight and accountability become increasingly important. As several participants observed, the challenge is no longer simply whether AI can do something, but whether it should. 

Perhaps that is the real lesson from the discussion. As AI becomes more capable, more personalised and more embedded in everyday life, the technical questions remain important. But the most interesting questions are still human ones. How do we build trust without creating dependency? How do we make information more accessible without losing critical thinking? And how do we balance the digital and human worlds?

The answer may be that the best personas are not artificial people at all. They are carefully designed frameworks that bring together purpose, tone, expertise, examples and boundaries. In other words, they help us decide not who the AI is, but how it should help.

This was generated from notes using Copilot in the style of Northern Lights blog.  Content was reviewed, amended and examples added.

Susan Smith, Knowledge & Library Service Manager, Mid Cheshire Hospitals NHS Foundation Trust

Feature image from Magnific by pikisuperstar