Persona

A persona is an invented figure with clearly described traits. These usually include a name, an occupation, a tone of voice, and certain likes or dislikes. This figure doesn’t really exist; it’s a tool for thinking and planning. In advertising, personas are used to picture a target audience as a single person. In the computing world, a program is given a persona so that it always behaves consistently. The word comes from Latin and originally meant mask.

Why companies give their chatbots a character

A language program can, in principle, respond in any style. Without guidance, it wavers: sometimes formal, sometimes casual, sometimes rambling. For a company, that’s a problem. Customer service should leave the same impression on every inquiry. A defined persona provides this reliability.

There’s also a safety aspect. A persona defines not only how something is said, but also what isn’t discussed. A travel bot shouldn’t give legal advice or answer questions about competing products. Such boundaries are written into the role description. This is no substitute for real safeguards, but it helps in everyday use.

The providers of the major AI systems themselves also work with personas. The well-known assistants have a recognizable character: polite, cautious, explanatory. When a new model noticeably changed this character in 2025, many users complained loudly. They had grown accustomed to a particular way of answering. This shows how strongly a persona shapes the perception of a product.

How a role is technically defined

Most often, this happens via a hidden piece of text at the start of every conversation. This text is called a system prompt and is an instruction the user doesn’t see. It might say, for example: You are a customer service representative for Company X, you don’t address anyone informally, you answer in at most five sentences. The model reads this instruction along with every response. It acts like a stage direction in theater.

This method is cheap and can be changed within minutes. However, it’s not particularly stable. Users often deliberately try to break the role by talking the program into a role-play. Such attempts are called prompt injection. In long conversations, a persona also sometimes fades on its own, because the original instruction loses weight relative to the rest of the text.

Those who need more stability go a step further. In so-called fine-tuning, a finished model is retrained with many example conversations in the desired style. The role then sits deeper within the model and is harder to talk it out of. The price: this costs computing time, sample data, and money. For small applications, the effort is usually not worth it.

Between product design and the chat window

In everyday life, you encounter a persona every time a chat window opens on a company website and introduces itself with a first name. Companion apps like Character.ai also build their entire offering on personas. There, users choose a character or invent their own. These applications are controversial because some people blur the line between a role and a real counterpart.

The second area has nothing to do with AI and is older. In product development, teams describe typical customers as invented people: Lena, 34, commuter, little time, uses her phone on the train. Such profiles help with decisions about features and design. When you read the term in business news, this meaning is often the one intended.

An important distinction: a persona is not a separate model and not a separate program. It’s just a layer of instructions on top of a general system. Two chatbots with completely different characters can use the same model in the background. Understanding this also reveals why an assistant’s personality can change overnight with an update.