Introduction
Conversational User Interfaces (CUIs) are nothing new: as of 2022 about 30% of Canadians owned a smart speaker [1] and about 67% of consumers worldwide had used a chatbot within the last year. [2] However, the recent boom of AI technology has rapidly accelerated the capabilities of conversational design - the proliferation of Large Language Models (LLMs) means that AI-enhanced conversational interfaces can interact in a way that’s more natural and human-like than ever before.
On one hand this means there is a massive opportunity in terms of how LLMs can continue to enhance the user experience of these platforms through more empathetic and ‘realistic’ interactions. On the other hand, the non-deterministic nature of AI opens a host of new challenges and considerations that need to be addressed.
Content generated from AI isn’t alive or conscious – it can only replicate language and patterns to the best of its understanding. It’s up to us to teach it how to behave in ways that are meaningful, useful, and relevant to users.
As designers (a term I will use in reference to any person who has a hand in creating these conversational experiences, from business analysts to developers!) our job is to combine our understanding of technology, psychology, and language to harness the power of AI in a way that augments users’ experiences.
But how do we do that?
The Power of Conversation
Before we can dive into how to apply principles of human-centered design to conversational AI, we need to take a step back and understand why the medium of conversation is so powerful.
At its core, communication is the way in which we exchange information with others. Every day you communicate across dozens of different mediums without thinking twice: you make small talk with your partner during your morning routine. You wave to your neighbor as you head to work. At your favourite coffee shop, you go through the familiar rhythm of asking the barista how their weekend was. You like a friends’ Instagram story while waiting for your drink and save a meme to show a coworker once you get into the office. The list goes on and on as we rack up hundreds of these micro (and macro!) communication interactions a day.
Without realizing, we all abide by a social contract that makes these moments of communication work: we assume that our interactions with each other are collaborative – we co-create the contexts of our conversations and subsequently develop our response based off what our conversation partner has to say.
How does this tie into conversational interfaces? By allowing users to fall back on the subconscious patterns of communication they utilize every single day, we completely reduce the cognitive load it would take for them to learn new processes, platforms, or technologies - effectively streamlining the way we empower users to meet their needs or accomplish their goals.
So how can we effectively use this knowledge and apply principles of good user experience design to conversational AI?
Center the User
It seems obvious, but we need to keep users at the center of our focus. It can be easy to get lost in the capabilities of AI, to see the mountains but miss the molehills.
A study conducted in 2024 found that only 31% of Canadians trust AI – a whopping 19 points lower than the global average. [3] We need to be mindful of the effect this has on users’ perceptions of experiences with AI and do our best to proactively mitigate anything that may further erode a users’ trust.
With this in mind, we can consciously construct experiences that actively build trust with users, by following four simple guidelines:
Align to the Real World
Conversing with something that we know is not human but behaves like a human can be a weird experience if you think about it for too long. However, research shows that when interacting with technology, people respond to the persona on the other side as they would another human. This means they expect to follow the existing mental model of human-to-human conversation and the social contracts we abide by when speaking with one another.
At this point, we understand the importance of mimicking natural conversation – but this also includes things like adapting to context. Adapting to context can take on a few different meanings: how does the CUI respond to different emotional contexts? For example, if a user is dealing with a frustrating situation, is there a conversation path we can design to help react in an empathetic way? If the user wants to refer to something that took place earlier in the conversation, or in an earlier conversation (a common pattern among humans) can the CUI refer to – and even make changes to – that information if needed?
Create a Persona
We’ve established that people speak to technology as they would another human – meaning they will also assign it human characteristics if those traits are not defined. Rather than leaving that up to the user, it’s better to take the opportunity to define your CUI’s persona. Are they peppy? Informative? Chatty? Blunt? Calm? Enthusiastic?
It’s not our goal to make the user think they’re talking to a human, but consistency in tone and voice creates a more realistic exchange. Imagine having a conversation with someone who changes how they talk every few sentences – it would be jarring! By creating a persona that is consistent in its vocabulary and approach, we can continue to offer a sense of familiarity and normality that builds trust between the CUI and the user.
This persona can even evolve over time based on a users’ interactions; for example, while some may prefer a more empathetic conversational tone, others may respond better to more logical, straight forward conversation. The AI persona prompt (sometimes called a system prompt) being used to define the persona can be tweaked and tuned as needed. Matching the tone of the conversation is another way we can convey empathy and align to users’ mental models of communication, while still maintaining trust and familiarity.
Keep it Simple
Make sure you’re providing your user what they need to move forward – and only what they need. Now is not the time to be wowing users with additional information. We want to keep things as simple as possible - that means no complicated words or technical jargon. Be as literal as possible and avoid using idioms or turns of phrase - you never know if someone is a native English speaker or is familiar with your language choice!
We can provide instructions to the AI to put important information first or last, or to chunk out information in a way that reduces cognitive load. For example, this could look like breaking information in blocks of 3 +/- 1, or applying a ‘one breath’ rule - if you can’t say the sentence with one breath of air, it’s too long for your CUI.
Develop Error Strategies
While we want our CUIs to mimic human behaviour as closely as possible, it serves no purpose to pretend that they are humans. Acknowledging and preparing for this up front can alleviate potential friction down the line when it comes to known issues that can occur with LLMs such as:
- Hallucinations: Generating false or nonsensical information.
- Lack of common sense: Difficulty understanding queries that require world knowledge or reasoning.
- Sensitivity to input phrasing: Producing different responses to slightly rephrased queries.
- Verbosity: Providing overly lengthy or irrelevant information.
- Bias: Reflecting biases present in the training data.
To create truly effective and user-centric conversational AI, it’s crucial to address these limitations and make interactions more intuitive. Here are some key strategies:
- Incorporate structured knowledge: Integrating external knowledge bases or databases using RAG (Retrieval Augmented Generation – you can read more about that in our second blog post of this series) can ground the LLM’s responses in facts, reducing hallucinations and improving accuracy.
- Fine-tuning: Training the LLM on domain-specific data enhances its understanding of particular topics and helps mitigate bias. It could also involve recognizing patterns in user corrections, identifying common misunderstandings, and adjusting behavior to minimize future errors
- Intuitive feedback mechanisms: Allow users to easily highlight and correct inaccuracies or provide feedback directly within the conversation. This could involve clickable elements to flag problematic responses or a “this is incorrect” button that prompts the AI to reconsider its output.
- Natural language error correction: Develop AI agents capable of understanding and responding to natural language corrections. For example, if a user says, “No, I meant X,” the AI should be able to interpret this as a correction and adjust its response accordingly.
Conclusion
The rapid growth of LLMs and AI models means we have extremely powerful tools at our disposal that can help us create more engaging and empowering user interactions than ever before - however, it’s important to remember that these tools are ultimately mimicking our communications patterns, not understanding them.
As designers and builders of these conversational experiences, it’s up to us to configure these tools in ways that are built upon the tenets of good UX. We hope this blog has given you some tools that can help shape your conversational experiences to behave in ways that are meaningful, useful, and ultimately build trust with your userbase.
Sources
[1] https://www150.statcan.gc.ca/n1/en/daily-quotidien/230720/dq230720b-eng.pdf?st=MOWcvFNU
.png)
