According to renowned consulting firm Norman Nielsen Group, Generative AI depicts the third major paradigm in how humans interact with user interfaces. Evolving from batch processing and later, command-based interaction, people now express their intent, or desired outcome, rather than the exact operation the computer should run. This shift from defining the HOW to defining the WHAT is probably the most exciting development in Interaction Design over the last 60 years.
However, with new technology come new challenges, and this is especially true for Generative AI and Large Language Models, or LLMs. The latter is the piece of AI that can process and understand natural language, as it has been trained on vast amounts of text available on the internet and curated for this very purpose.
What is Conversational AI?
Conversational AI is the engine that powers chatbots. Its main purpose is to understand what you are saying and come up with a useful response. You can literally ask anything, and everything, and in many cases the responses are close to how a human would reply. This capability is probably the most “magical” aspect about this type of AI, as it allows you to have a real conversation, even if you use poor or improper language.
An important concept to bear in mind regarding AI-powered chatbots is the so-called Context. It can be imagined as the short-term memory of the LLM and is essentially your chat history. Each new prompt will look at what you asked previously and what the LLM replied. This will usually improve and enrich the responses you get unless you switch topics. You can use this to narrow down your initial request or ask for follow-up information on a response. Be advised, though, that context length is limited, depending on your subscription. Like short-term memory, the LLM can retain only a certain amount of information at any time.
A groundbreaking aspect of accessibility is that the LLM comprehends you, regardless of the language you use to formulate prompts. Given that the majority of LLMs are trained on English text, responses in English are typically considered to be most accurate. However, the AI can understand a vast range of languages, enabling live translations and helping non-English speakers all over the world.
Modality Matters
While text is still a very common input method in the context of Conversational AI, voice commands have been around for decades, especially in mobile applications and assistants, such as Siri and Alexa (Cortana, anyone?). The speech-to-text technique captures spoken input and converts it into text output, seamlessly integrating it into the chat interface as if it were typed. The better the voice recognition, the more productivity gains are to be expected. This is also an important aspect when thinking about accessibility or to simply have an alternative input method when your hands are busy.
But it doesn’t stop there. More recent models are capable of understanding and processing image or video data. This lets you upload a screenshot and have the AI interpret what it sees, for example. Applications like this can augment user experience drastically, by unlocking an additional layer of contextual information and machine-grade analysis. Product designers need to be aware of these additional input methods and their applications, as they have significant influence over the overall user experience.
Common Challenges with AI Chatbots
One major challenge faced by prompt-based AIs is how well the users articulate their requests, as crafting text can prove more difficult than comprehending it. AI chatbots depend on highly articulated commands to return good results in a few prompts. To minimize friction, users must initially be informed about the capabilities and limitations. While you may eventually reach your desired outcome, establishing a starting point is essential for conversational AI. Examples to assist users in onboarding include video or text tutorials, suggested actions, or a readily available help menu within the chat interface itself.
A second challenge with AI is that users need to know a thing or two about how to ask an LLM to achieve their goals. The model’s lack of intuitiveness and self-explanation has paved the way for an entirely new discipline, the so-called Prompt Engineering, with the goal to "make the LLM behave the way we want".
Finally, let’s consider the aspect of cost. Every conversation involves the transmission of words to and from the LLM. Off-the-shelf AI vendors typically charge per word (or token). This raises the question of whether you can monetize the AI in your product or offset your expenses in other ways. Fortunately, integrating AI into your product isn't limited to chatbots alone. Continue reading to explore common interaction patterns and elements that offer diverse avenues for AI integration.
Beyond Chatbots: User Interaction Patterns with AI
One of the most notable interaction patterns in AI involves static elements, for example, buttons within the UI, designed to trigger AI-assisted operations upon clicking. Often represented by the form of a wand decorated with sparkles, these elements evoke a sense of magic about to unfold. A typical application for this component is to extract the information from the UI and perform relevant actions. For instance, in IT Service Management (ITSM), it could summarize ticket contents or generate solution recommendations.
Another interaction method involves AI-driven recommendations and contextual information, presented without explicit user requests. There are two flavors: static and dynamic. Static elements enhance the overall user experience by offering additional information and context at predetermined locations within the UI. Dynamic elements utilize sophisticated techniques to determine where and when to appear, providing trailed advice to users. These recommendations often include a call-to-action, prompting users to dig deeper into presented information or to take follow-up steps.
Discoverability and Guidance
A last consideration is striking the right balance between discoverability, user guidance, and automation. While an intuitive interface can facilitate a users’ understanding and discovery of available AI features, it is also important to consider proper user guidance through AI-driven interactions. Moreover, balancing keeping the user in control with AI automation can help foster trust and ensures user empowerment in decision-making processes. It is important that AI-driven interactions remain responsive to a users’ needs, preferences, and comfort-levels with semi or fully automated actions.
In conclusion, the evolution from traditional user interfaces to Generative AI marks a revolutionary step and true paradigm shift. A successful integration of AI requires careful consideration to enrich user interactions effectively. Functionalities must be easily reachable and comprehensible to the user, regardless of the operational intricacies or modalities.
