Designing for Artificial Intelligence
When designing for an AI based system, one must rethink today’s paradigm of data entry, to a model of decision validation.
Artificial intelligence (or machine learning, cognitive computing, etc. ) requires product designers and managers to rethink how users interact with systems. While there is lots of buzz and hype with complete AI automation, in reality, AI still requires human interaction to teach and hone relevance algorithms.
When designing AI products, it’s important to understand the fundamentals of how AI shifts user experience. For most systems, the majority of the user experience is about getting data into the system. Once the information is in the system, now the user can do specific things with it. Most people spend 90% of their interactions with systems specifically in data entry, (e.g., buying, consuming or sharing) with 10% acting on the computed output.
Combining the data that is user-generated, either behavioral or manually entered, provides greater capabilities to identify patterns that reside within the data (big data). This approach resulted in visual charts and graphs, but fundamentally still required interpolation of what the data meant, both from the data science as well as the one trying to understand what the charts say.
The intent with AI is to leverage the data that exists and allow the system to automatically develop “answers” to user questions known or unknown in a time frame that is real-time. Also, it shifts the model from users trying to interpolate the data, to a model where the AI provides explanations and suggests specific actions based on the patterns in the data. As a result, the more information it can process, the more opportunity for the system to provide more relevance in its proposed actions.
But this only goes so far. In essence, humans have to interact with this data to continually hone how AI identifies, processes and derives accurate responses. Without this, the system can only do a baseline level of processing. Humans must be involved to guide the direction of what the AI needs to do, resulting in a rethinking of how the experience shifts from data entry to “teaching” the system to deliver valuable automation.
When I design for an AI-driven system, I break the design into four primary tenets; Tuning, Simulation, Exception, and Narrative. While each drives individual value to AI, as a whole they ultimately feed a closed loop of human input which drives the automation and desired output.
Tuning focuses the behaviors needed within the AI algorithm. All algorithms begin with a baseline, and continually learn and grow based on user input.
Evaluation is a model that allows a user to predict the resulting outcome of the AI results. Since an AI system will learn based on user input, this enables the user to understand the baseline direction in which the AI will learn and evolve.
Exceptions are ways in which the system will require human intervention outside of the norms of generalized tuning and simulation. Exceptions are clarifying questions (anomalies) that the AI asks the user, ultimately helping to hone the ways that the AI will solve similarly or like situations.
Narratives are the method in which AI communicates to the end user in a format in which all users can understand. Instead of a user having to interpolate data, the AI system provides a conversation to guide a user. Bots are a good example of how AI creates a narrative experience.
At its simplest terms, it is just like a school experience; tuning is the curriculum; evaluation is the testing, exceptions represent the dynamic question/answering, and the narrative is the teacher delivering the knowledge.
When designing for AI, one must rethink typical approaches of data entry, to one of decision tuning & validation. That in itself is a complete paradigm shift to how we have thought about experience design for the last 20+ years.