Artificial Intelligence is taking the world by storm, as the old saying goes, not just automating redundant tasks, but reshaping business, society, and science. However, many large AI projects fail to deliver on their goals. AI algorithms have proven to encode societal biases, spread rumors on the internet, provide fake news, and make mistakes, such as with advice given by AI driven chatbots.
Writing in Wired, Jim Guszcza, the former US chief data scientist of Deloitte Consulting and research affiliate at Stanford University, asserts that intelligence is not in the computer. While AI can do many tasks perhaps better than humans, such as drive cars, interpret spoken language, recognize faces, write reports, solve calculus, and create program software, we may believe that computers understand what we are doing. “That’s an illusion,” he writes, and clarifies his point:
AI algorithms are “mind tools,” not artificial minds.
This misconception leads to unrealistic or even misleading expectations for what AI can do. Successful applications of AI require more than big data, big iron, and advanced math. Human-centered design is also crucial. AI applications must reflect realistic conceptions of user needs and human psychology. Paraphrasing the user-centered design pioneer Don Norman, AI needs to ‘accept human behavior the way it is, not the way we would wish it to be’.
“Smart technologies are unlikely to engender smart outcomes unless they are designed to promote smart adoption on the part of human end users. Just as adding more intelligent people to a team can result in its effectiveness being diminished, so can AI result in ‘artificial stupidity’ if poorly designed, implemented, or adapted to the human social context. Human, organizational, and societal factors are crucial.”
For those of you implementing AI to improve efficiency and cut costs, I recommend you reflect on the many times AI implementation went south and reflect on Jim Guszcza’s advice to keep humans at the center of AI Design.
Jim Guszccza has four aspects of human-centered AI Design:
Goal-relevance. Data science products and AI applications are most valuable when insightfully designed to satisfy the needs of human end users. For example, typing “area of Poland” into the search engine Bing returns the literal answer (120,728 square miles) along with the note: “About equal to the size of Nevada.” The numeric answer is the more accurate, but the intuitive answer conveys the approximate size of Poland to far more people.
Handoff. “The paradox of automation” Jim Guszcza writes, “is that the more reliant we become on technology, the less prepared we are to take control in the exceptional cases when the technology fails—conditions that often require more skill than those handled by the AI.” People who let their skills erode, or don’t seek learning experiences, can find they’ve become dependent on a machine that fails them when they need it most.
I can add another example. I watched a Waymo, driverless car, approach two traffic cops who were standing between the car lanes waving the Waymo car (which had cautiously come to a dead stop) to move forward in a construction area, where the car could not turn left or right. The car wanted to turn right, signaling a right turn, to find an alternative route from the middle lane. I was behind the car in the right lane. If the car made its turn, it may have led to an accident. The Waymo car froze for several seconds, which seemed like an eternity with a line of cars behind me building. One of the police offers walked towards the car, careful to stay to the left of the car, and began making exaggerated hand jesters for the car to move forward. After about 15 seconds of this, the car began to move forward cautiously. The dilemma was solved due to the human-being: the police officer intuitively knowing he had to repeat and exaggerate hand motions for the computer to understand his meaning.
Feedback loops. Automated algorithmic decisions can reflect and amplify undesirable patterns in the data they are trained on, because AI will do as it is trained, without a sense that it has gone off the rails by misinterpreting user behavior, or being unaware of social boundaries that humans instinctively work within.
I will add an example. In July, 2023, during the height of the Barbie movie excitement, Buzzfeed published an article featuring AI-generated images of what “Barbie would look like in every country in the world.” People on Twitter have been pointing out the blatant racism and endless cultural inaccuracies seen on many of the “dolls.” For example, the German barbie was wearing an uniform reminiscent of an SS Nazi General. Not a stereotype Germany wants. An AI-generated Barbie for the South Sudan held a gun on her left side. The Barbie for Vietnam was not wearing traditional Vietnamese clothing.
Another example includes the popular AI chatbots from OpenAI Inc., Google LLC, and Meta Platforms Inc. These platforms are prone to “hallucinations” when answering legal questions, posing special risks for people using the technology because they can’t afford a human lawyer, new research from Stanford University said.
The Stanford University study found that large language models used by AI chatbots hallucinate at least 75 percent of the time when answering questions about a court’s core ruling, the researchers found. They tested more than 200,000 legal questions on OpenAI’s ChatGPT 3.5, Google’s PaLM 2, and Meta’s Llama 2—all general-purpose models not built for specific legal use.
Psychological impact. Algorithms designed without a focus on their users’ human nature can impair user behavior. Social media platforms deliberately designed to be addictive have been linked with feelings of unhappiness and “fear of missing out” in compulsive users. Silicon Valley insiders increasingly worry about people’s minds being “hijacked” by addictive technologies, leading to problems with the work and home lives, health, and mental well-being of obsessive users. Addictive social media have also been blamed for polarizing society and undermining rational discussion.
Finally, Guszcza asserts, that the decision environment (the algorithm and the human decision-makers) must be well-designed. An algorithm’s end users should have a sufficiently detailed understanding of their tool to use it effectively. Guidelines and business rules should be established to convert predictions into prescriptions and to suggest when and how the end user might either override the algorithm or complement its recommendations with other information.
Everyone implementing AI needs to remember that the intelligence is not in the computer. That is an illusion. Smart technologies are unlikely to engender smart outcomes unless they are designed to promote smart adoption on the part of human users.
About Victor
Victor Assad is the CEO of Victor Assad Strategic Human Resources Consulting and Managing Partner of InnovationOne, LLC. He works with organizations to transform HR and recruiting, implement remote work, and develop extraordinary leaders, teams, and innovation cultures. He is the author of the highly acclaimed book, Hack Recruiting: The Best of Empirical Research, Method and Process, and Digitization. He is quoted in business journals such as The Wall Street Journal, Workforce Management, and CEO Magazine. Victor has partnered with The Conference Board and the US Department of Energy on innovation research. Subscribe to his weekly blogs at http://www.VictorHRConsultant.com
1 comment