Watch all the Transform 2020 sessions on-demand here.
In the world of business and design, we have started using terms like “algorithm” and “machine learning” as magic calculations for problems we would like to gloss over. These terms often become blockers for deeper problem solving and can stall even the most worthwhile projects.
“We’ll figure it out with an algorithm” used generally is similar to the fantastic conversations my daughter and I had when she was six. She would come up with inventions for amazing things like solar power in a backpack and magic windows. If she wanted to skip over something, she included, “and then you put some potion on it.” This would quickly get her to the other side of the idea. Soon after that, she moved past solving problems with potion and is still coming up with inventive solutions for many things.
Describing an algorithm in this general way has become the latest version of putting some potion on it. Often, the real solution is not particularly complicated, but does require some deeper thought. As strategic designers, we can all get more involved in emerging artificial intelligence by breaking down the problem and getting more deeply involved in the discovery process.
Identifying the manual task
Often the first part of creating the “potion” is to find a task that can be done manually. If the goal is analyzing text in articles, for example, the first task should be taking the text from articles and finding the right words, word pairs, or sentences that would get to the result we are looking for — make sure it works before you automate it.
June 5th: The AI Audit in NYC
Join us next week in NYC to engage with top executive leaders, delving into strategies for auditing AI models to ensure fairness, optimal performance, and ethical compliance across diverse organizations. Secure your attendance for this exclusive invite-only event.
The Flame Index was a service I cofounded with a three-person team in 2010, which scanned thousands of news sites 24 hours per day and published a real-time ranking of companies generating the most negative press. It sounds like an impressive algorithm, but the starting point was the manual and time-consuming process of reading news articles and creating a list of words and word pairs that seemed to define the issues we were looking for. The spreadsheet analysis work helped us dig deeper into the problem we were hoping to solve. Terms like “bankruptcy,” “oil spill,” and “explosion” were prime candidates for how we were analyzing issues — they revealed how “on fire” each company had become in the media. Over the five-year life of the company, we manually identified more than 2,000 words and word pair terms for the ranking.
For each of the terms on the index, we assigned a score to rank negative news. When we knew what we were looking for, we automated the process of looking through articles, calculating a score for each article and for the company overall. We didn’t try to capture all of the words from the beginning, and we didn’t automate the process of finding new words. We continued with the manual process of adding words and adjusting scores over time. Identifying words and sentiment worked much better as a human curation process.
After six months of iterations, we launched an automated service, which seemed to accurately reflect what we were seeing in the news, and interestingly, it also reflected impact to stock trading volume. More of the details are captured in this New York Times Tech Talk Podcast on the Flame Index from August 2011.
This process revealed the need for several jobs as artificial intelligence progresses:
- The initial problem solving and manual development of the logic
- Ownership and maintenance of the analysis
- Refinement and evolution of the analysis over time as other services change
- The refinements become important problems to monitor, especially if people become dependent on the information provided by the automated analysis service
Analyzing words and working with people
Very often ingredients for the potion are found in word and sentence analysis — this is the magic of natural language processing. Part of that starts with finding which of the words are nouns, verbs, and adjectives and using that information to do the more interesting analysis of written intent. Automation increases the speed of making those connections.
This process lines up very well with Google’s Human-Centered Machine Learning philosophy, which focuses on how people might solve the problem manually before resorting to algorithms to solve a problem. This approach connects us directly to the outcomes that would have the largest impact on potential users of the services; it also keeps the focus on practical prototyping with people and gathering participant data as part of the process.
Identifying locations for retailers
At Fjord, we are working on a process that would help retailers find the right locations to meet potential customers. This started with weeks of the manual process of looking up connections in social media and creating words in spreadsheets and a wall of photos and terms, which start to approximate an automated process we would eventually like to implement.
Even before we get to the algorithm, the manual process has already resulted in recommendations for retailers and has revealed the repetitive tasks we would like to automate in order to speed up the analysis process.
If we start to look at “algorithm” as a sequence of tasks to solve a complex problem and not a mysterious potion, it becomes possible to apply our human intelligence to AI. Designing AI with human values and curation is critically important to avoid the unintended consequences of automation out of control.
John Jones is senior vice president design strategy at Fjord, a design and interactive studio from Accenture Interactive.
This story originally appeared on Medium. Copyright 2017.
