Human-centered AI design is different. The speed, scale and scope of AI materially alter the standard design process, so we have tailored our sprints for AI development needs.
Human-centered AI design starts with considering how people are reflected in the data, because an AI system touches more humans than just the immediate users. Our AI design sprint techniques start to define proxies, approximations, alternative data, graph connections, missing data, data completeness, data skew, bias, representation and cleanliness. This helps non-technical people understand and anticipate the marketing and communication needs that come from data selection and processing choices.
There are many flavors of algorithm. The breed of AI gets set here – such as answering the question of whether to choose deep learning or another technique such as a support vector machine. These highly technical systems all deal with data differently, offering different strengths and weaknesses. Different algorithms will deal with transparency and trade offs in different ways so the voice-of-the-customer is vital.
Sometimes it may be necessary to trade off one need for another. For example, in a customer-facing process where there is a high need for transparency, it may make sense to disallow certain algorithms or model forms. While some predictive power may be lost, which has an economic consequence, the benefit of a simpler, more explainable system can increase trust in the result.
While there are many details and project requirements that come later, our Sprints set parameters for non-technical people to follow the progress of design later in the process.
There are important trade-offs to track when building AI models: selection of training data, detecting overfit and tuning of models. It’s important that machine learning engineers and data scientists can be creative, and there may not be a bright line between a creative decision and a policy decision. Non-technical leads need to be knowledgeable about model creation to be able to support the creative work of technical people.
Without strong ties between business and technical leads at this point in the process, everyone is working suboptimally. Our sprints build alignment between technical choices and business goals.
Predictions are getting cheaper. By definition, AI will be wrong at some point: false positives or false negatives, model drift, bias, unfamiliar situations, inappropriate context or precision-recall tradeoffs. This means that business leaders need to understand how these predictions are made and to plan for what to do when an AI makes mistakes.
Ultimately, accountability for the performance of an AI system is with humans.
AI learns from human reactions. Humans change their behavior as they react with an AI. Design to take account of a new intelligence and use AI to create a new relationship, whether it’s with a customer or an employee. What kind of relationship do you want to build? How should friction be eliminated? How should a user learn about the intelligence they are interacting with? What human qualities matter and when?
AI sprints plan for AI “in the wild” and the virtuous cycle of human-machine interaction.
Sonder Scheme AI sprints help technical and non-technical people come together and make better AI-enabled products, faster.