Why AI Needs a Dose of Design Thinking - Deloitte/WSJ.com
Artificial intelligence technologies could reshape economies and societies, but more powerful algorithms do not automatically yield improved business or societal outcomes. Human-centered design thinking can help organizations get the most out of cognitive technologies.
Today’s artificial intelligence (AI) revolution has been made possible by the big data revolution. The machine learning algorithms researchers have been developing for decades, when cleverly applied to today’s web-scale data sets, can yield surprisingly good forms of intelligence. For instance, the United States Postal Service has long used neural network models to automatically read handwritten zip code digits. Today’s deep learning neural networks can be trained on millions of electronic photographs to identify faces, and similar algorithms may increasingly be used to navigate automobiles and identify tumors in X-rays. The IBM Watson information retrieval system could triumph on the game show “Jeopardy!” partly because most human knowledge is now stored electronically.
But current AI technologies are a collection of big data-driven point solutions, and algorithms are reliable only to the extent that the data used to train them is complete and appropriate. An algorithm trained to pilot a car would be useless at translating a document, and vice versa. A self-driving car that performs reliably in Palo Alto, California, might fail in Pondicherry, India. One-off or unforeseen events that humans can navigate using common sense can lead algorithms to yield nonsensical outputs. Finally, there is growing recognition that algorithms trained on data can reflect—and possibly amplify—unintended societal biases.
While algorithms can automate many routine tasks, the narrow nature of data-driven AI implies that many other tasks will require human involvement. In such cases, algorithms should be viewed as cognitive tools capable of augmenting human capabilities and integrated into systems designed to go with the grain of human—and organizational—psychology. To explore this emerging discipline, known as cognitive design thinking, Deloitte Consulting Innovation hosted a salon at the Cambridge Innovation Center in Boston, Massachusetts.
“An AI revolution is underway right now, but I believe it needs to be complemented with a design revolution,” says Jim Guszcza, chief data scientist at Deloitte Consulting LLP. While some prognosticators imply computer intelligence may eventually supersede human intelligence, AI algorithms lack common sense, he says. “We don’t want to ascribe to AI algorithms more intelligence than is really there. They may be smarter than humans at certain tasks, but more generally we need to make sure algorithms are designed to help us, not do an end run around our common sense.”
Collaboration between humans and machine can create a form of collective intelligence when the strengths of one counterbalance the weaknesses of the other. “Getting there requires design thinking,” Guszcza says.
Although cognitive design thinking is in its early stages in many enterprises, the implications are evident, says Rajeev Ronanki, a principal at Deloitte Consulting and leader of its cognitive and health care innovation practices. Eschewing versus embracing design thinking can mean the difference between failure and success. For example, a legacy company that believes photography hinges on printing photographs could falter compared to an internet startup that realizes many customers would prefer to share images online without making prints, and embraces technology that learns faces and automatically generates albums to enhance their experience.
To make design thinking meaningful for consumers, companies can benefit from carefully selecting use cases and the information they feed into AI technologies, Ronanki says. “Go after rich domains of data, and ask consumers to grant permission to use their data to teach the platform what’s best for them,” he says. For instance, the Nora Intelligent Assistant, powered by Deloitte’s cognitive technologies and platforms, collects personal health data and uses machine learning algorithms to generate health-related insights made readily available to patients. These insights can both empower users and be useful to health plan administrators in understanding their populations, he says.
In determining which available data is likely to generate desired results, enterprises can start by focusing on their individual problems and business cases, create cognitive centers of excellence, adopt common platforms to digest and analyze data, enforce strong data governance practices, and crowdsource ideas from employees and customers alike, Ronanki says.
In assessing what constitutes proper algorithmic design, organizations may confront ethical quandaries that expose them to potential risk. Salon attendees watched a video of Joy Buolamwini, a graduate researcher at the MIT Media Lab and founder of the Algorithmic Justice League, discussing how unintended algorithmic bias can lead to exclusionary and even discriminatory practices. For example, facial recognition software trained on insufficiently diverse data sets may be largely incapable of recognizing individuals with different skin tones. This could cause problems in predictive policing, and even lead to misidentification of crime suspects.
“If the training data sets aren’t really that diverse, any face that deviates too much from the established norm will be harder to detect,” Buolamwini says. Accordingly, across many fields, “we can start thinking about how we create more inclusive code and employ inclusive coding practices.”
Karthik Dinakar, co-founder and CTO at startup Pienso, underscores the importance of understanding processes that generate data used to train cognitive technologies. “Philosophy is very important to design thinking. It’s critical to understand your own attitude and mindset prior to approaching a problem,” he says. Machine learning can help companies model the generative processes underlying data sets, and Dinakar intends Pienso to democratize machine learning for domain experts who lack technical, data science, or modeling backgrounds. For example, “Predicting heart diseases with algorithms in some cases may not be ideal, but we can encourage cardiologists to use algorithms to improve health care decision-making,” he says.
CIOs can introduce cognitive design thinking to their organizations by first determining how it can address problems that conventional technologies alone cannot solve, Ronanki says. “The technology works with the right use cases, data, and people, but demonstrating value is not always simple,” he says. “However, once CIOs have proof points that show the value of cognitive design thinking, they can scale them up over time.”
Technologists should not develop algorithms in a silo, Guszcza adds. “CIOs benefit from working with business stakeholders to identify sources of value,” he says. It is also important to involve end users in the design and conception of algorithms used to automate or augment cognitive tasks. “Make sure people understand the premise of the model so they can pragmatically balance algorithm results with other information,” he says.
For more discussion, see “Cognitive Collaboration: Why humans and computers think better together” in Deloitte Review.