Content Designers: the unsung heroes of design

There are people that can make a sentence dance to the rhythm of its own syllables. They know the difference between ‘there’ and ‘their’ is anything but subtle, and as lovers of the almighty word, they design experiences with voice and tone. They are former English majors and writers who have found themselves at the helm of AI, a technology that will change the way we work, live, play, and perceive the world.

Professionally, these wordsmiths are known as content designers. While the designing of digital products may seem like it’s all about visual aspects like layout, color, icons, or buttons, another deeply important layer is at play: language. This is where content designers apply their trade, but just like all other UX disciplines, the AI era is prompting them to reinvent their expertise. Like never before, their linguistic knowledge is needed, but it’s being applied in a totally different way.

“One is the standard way; it’s the traditionally deterministic level. You could consider it the surface level that all designers are engaged with. The introduction of large language models (LLMs) creates a probabilistic UX,” said Jonathan Foster, who leads the content design team for Microsoft 365 and Copilot. From deterministic to probabilistic UX, the two levels couldn’t be anymore opposite.

In deterministic UX where outcomes are predictable and straightforward, like scheduling a meeting in Microsoft Outlook, content designers craft language that guides people through linear steps. “We’re often associated with voice and tone on the surface level, and that language is put in there as design material,” said Jonathan. Some of the very words you are reading at this moment were produced with their help. The importance of content designers may not be fully realized, but language is the bridge between cold, sterile technology and the people using it. The challenge here is ensuring clarity and simplicity while accounting for different user levels and changing contexts, but “what’s happening is the evolution from a purely deterministic UX to a probabilistic paradigm,” said Jonathan.

Particularly in AI-driven systems, probabilistic UX outcomes are less predictable because they are based on probabilistic models or machine learning. A customer’s prompt input affects their own UX because it’s instructing the LLM, which in turn creates the UX. On the backend of Copilot, content designers are writing the system (or meta) prompts that are a combination of Python code and the English language. A system prompt is a pre-written instruction or message that sets the context, tone, and boundaries for AI’s responses. “It’s the guardrails you’re giving to the LLM,” said Kate O’ Leary, who leads the content design team for Teams.

The evolution of content design: from peripheral to central

In the olden days, writers were often seen as peripheral to the design process. They were brought in towards the end, after the interface had already been designed, so they could fill in the blanks with copy. This approach treated language as an afterthought, focusing more on the visual and interactive elements of a product.

Today, LLMs power AI, a technology trained on the corpus of English language that makes up the Internet. Since AI is now language-based, content design will likely take on a central role in product development. “It’s all about the language, there’s no getting away from that,” said Jonathan. The LLMs tend to be very sensitive to the words in the metaprompts. Whether people are asking a voice assistant for directions or they’re receiving auto-repair shop recommendations from a chatbot, the design is probabilistic or non-deterministic. Design used to be the sole input into UX, making it deterministic and consistent for everyone who encountered it. Now, designers are just one of many inputs. The LLM, the work done within the LLM stack (including model tuning, prompt engineering or metaprompts, grounding data), and the customer’s prompts all contribute to the UX. This means we can no longer “determine” the UX; we must be comfortable with the probabilistic outcome. Designers can only nudge it in different directions, resulting in a unique UX for each person.

Now, content designers are doing a new kind of design thinking around how customers interact with Copilot. When it comes to system prompts, they are influencing the UX from inside the LLM stack. Rather than filling in, they collaborate with UX designers, product managers, and engineers to shape the person’s journey from the start, ensuring that the UX outcomes are in the service of people. Their system prompts guide the AI on how to behave and interact with people, ensuring consistency and alignment with specific objectives. Essentially, system prompts act as a framework that shapes the AI’s behavior throughout the interaction.

Designing for trust: picking the write words

When creating an online account, a designer might focus on making the “Create Account” button prominent and easy to click. But what about the label of the button itself? Imagine if it were labeled “Register Your Soul” instead of the straightforward “Create Account.” While it’s an exaggerated example, it highlights how wrong word choices can lead to confusion, frustration, or mistrust.

In a more complex scenario, some fans of Taylor Swift were once duped by a scam that imitated her concert ticket sales using poorly designed language (yes, this actually happened). The wrong choice of words in prompts or notifications can mislead people into thinking they’re on an official site when they’re being deceived. Misaligned content can result in real-world consequences —financial loss, wasted time, or even psychological harm. And what about health apps or online banking interfaces? A vague or poorly phrased instruction could create errors with serious consequences, from incorrect medication dosages to money transferred to the wrong account. Not to mention, how you set the tone to communicate with people in one country may be outside the cultural norms of another, even if they share the same language.

Designing trustworthy and safe AI experiences often means balancing nuanced language with the limitations of the technology itself. In Copilot’s case, the content designer’s job is not just to explain but to build trust. Trust is everything. That means giving responses that are logical to people, and addressing potential misunderstandings or biases that could arise from Copilot’s interpretations. But how do content designers do that?

No matter the product or feature, content designers will have to rigorously test and iterate on their prompts to see how Gen AI systems process and value the information given to them. When we first started working with Gen AI models, “There were three things that became very clear. One was that conditional statements worked really well. So, if a user does X, you should do Y,” said Isobel McCrum. She’s responsible for the language side of metaprompts within Copilot in Microsoft Word. Depending on the product, it’s helpful to include ideal output responses to set the product’s tone and voice. Since Generative AI systems are great at identifying patterns, they’ll duplicate your response in various contexts. Secondly, some AI models tend to remember the end of instructions but ignore the beginning. If there’s a hard rule you want to drive home, repeat it at the end. Lastly, clearly and precisely explain your response to get rid of ambiguity. Ultimately, it’s all about the words you use.

Staying humanity-centered, ethical, and inclusive

The more AI advances, the more it risks becoming detached from the human experience. At a time when this technology could potentially help find a cure for cancer, be used to monitor nuclear reactors, or optimize supply chains for getting products to consumers, it’s easy to fall into the trap of creating systems that are efficient but impersonal. Content designers must remain vigilant, ensuring that the language used in digital products keeps humans at the center.

This is particularly important for responsible AI. As AI systems become capable of making decisions that affect people’s lives —whether it’s in hiring, lending, or healthcare —it’s crucial that these systems precisely and fairly communicate with people. In a scenario where there’s an AI-driven medical alert system and a patient is experiencing a medical emergency, the system needs to provide immediate, actionable guidance to both the patient and responders. If the design is confusing or the information is poorly communicated, it could lead to delays in treatment or incorrect actions, potentially resulting in severe consequences. The language must be adaptable and transparent, providing people with the sources they need to understand where the data is from, and reminding people to cross reference their outputs, versus only relying on AI for accurate information.

But since people don’t always have time to double check their findings, content designers will more and more be tasked with crafting system prompts that do not reinforce harmful stereotypes or exclude vulnerable communities. In an era where misinformation can impact decision making or perception, grappling with the ethical implications of the technology they help shape is a great responsibility. As products become more powerful, content designers will contribute a deep sense of empathy and ensure that their work upholds ethical standards. We’re here to empower and serve the needs of our customers, upholding Microsoft’s core values and practices, and how we craft the content in products deeply matters.

Content designers are integral to that process. As the digital landscape continues to evolve —particularly with the rise of AI —their role will only become more important.