Because humans are by our nature biased, our data – and our code – will necessarily be as well, said Ayanna Howard, dean of The Ohio State University College of Engineering. But there is hope: Sometimes we can leverage human bias to beneficial ends. The trick is that we need to build our systems so that, when we identify bad outcomes from bias, we can fix them rapidly.
“We need to think about using AI intelligently … in order to be very strategic as to how it disrupts,” Howard said in her opening plenary talk, “How the Computing Research Community Can Manage the AI Disruption Through Inclusion,” at the PEARC22 conference on July 12.
The Association for Computing Machinery (ACM) Practice and Experience in Advanced Research Computing (PEARC) Conference Series is a community-driven effort built on the successes of the past, with the aim to grow and increase inclusivity by involving additional local, regional, national and international cyberinfrastructure and research computing partners spanning academia, government, and industry. ACM PEARC22, now taking place in Boston, is exploring current practice and experience in advanced research computing, including workforce development, training, diversity, applications and software, and systems and software.
Bias: central to the human condition
Bias has helped the human species survive, Howard explained, and so is not always a negative. Her own work has focused on human interaction with artificial intelligence and how the bias of both user and designer affects that interaction.
AIs have a huge potential to improve outcomes, she stressed with an example from her previous position as chair of the Georgia Institute of Technology School of Interactive Computing in the College of Computing. She gave one example of a sister college in Atlanta where, with a high number of nontraditional students, commuting students and those from stressed communities, the school had traditionally experienced a high rate of attrition.
An AI system built on data concerning common issues faced by students – for example, offering information concerning financial aid when students made general queries about money – helped decrease that attrition. In a better-known example from the College of Computing at Georgia Tech, an email AI “teaching assistant,” “Jill,” was able to answer most students’ questions so effectively that they didn’t know it wasn’t human. While the anonymity of the AI was somewhat controversial – Howard recommends making users aware when they’re interacting with an agent — it also decreased workload on the actual TAs while helping the students.
The issue, Howard explained, is to identify and mitigate the unintended consequences of people surrendering their autonomy when interacting with an AI, and understanding how these consequences affect their willingness to continue the interaction.
“As humans interact with these agents, they are sometimes prone to making their own mistakes because they trust the agent,” she said. But “… when people are corrected after a mistake, they actually take it personally.”
Effects of bias at design, user end
Much of Howard’s work has studied how bias at the user’s end affects interactions with intelligent systems, and how to navigate these biases to obtain the best outcomes.
In one study regarding deception by AIs, subjects were asked whether it was “appropriate” for an AI to lie to a child, an adult or an elder “for their own good.” Generally, subjects believed that lying to adults was acceptable, but not to children or elders. Still, when asked whether such lying “made sense,” the proportion of subjects who answered “yes” was far higher, reflecting that the answer for an actual AI in use is often not a simple yes or no.
In another phenomenon that underlines the unintuitive results of user bias, subjects proved more uncomfortable with humans that conflicted with sex-associated career biases – for example, a woman engineer versus male nurse. Interestingly, this “trust gap” didn’t appear to apply to AIs with clear genders, though when an AI was designed to be gender-neutral, users often “read in” a male gender.
The negative effects of bias on the design end can be more straightforward. In one study, the New York Times English reflected in language datasets seeded models trained on those datasets with toxicity for users employing African American English. In another, facial recognition systems’ ability to identify emotion — important for uses as disparate as identifying students undergoing emotional crises and automating test proctoring – suffered a loss in accuracy for users over age 55. Of course, facial-recognition bias against people of color has been the subject of reports in the trade and lay press.
“Every single collection of data that we are amassing has bias,” Howard said. “There is no way we can remove one hundred percent of bias, because that’s what we are — we are human.”
Reducing, controlling bias
So what does all this mean in terms of the design of intelligent systems?
“There is hope,” she said. At the design end, coders can be trained to be more aware of these issues. “What happens if we can train the coders; what happens if we can think about the cognitive bias?”
Still, given the bias likely to be present in any dataset, systems must be designed to be fixed after the fact as well. Howard’s group focuses on developing “hierarchical” models.
“There are ways that you can layer on information … to actually fix [problems] as you find them,” she said. By training a “minimodel” to filter out bias, she explained, one can overlay it atop a model producing problematic output without having to re-train the underlying model.
At the user end, bias can actually be leveraged to improve outcomes. The bias underlying gender identification of an agent, for example, needn’t be a problem. Presenting an AI with a neutral gender allows users to assign a gender they are comfortable with, posing an opportunity to prevent the issue from interfering with users’ experience.
“Basically, humans treat robots as robots,” Howard said. “… If we try to [assign gender] as neutral, people will be given autonomy; they will have the ability to make their own decision.”
Generally, Howard’s work suggests that taking all these biases into consideration – and paying attention to users’ perception of autonomy – can enable those who design AIs to avoid the effects of bias when possible and repair them when not.
Given the growing ubiquity of these systems in everyday life, this is an important goal, she noted. “As we continue to push this [technology], continue to basically change the whole ecosystem, it also means that we are at the point where we have to think about this strategically, so we don’t mess it up.”
Robot conveying human emotions. Courtesy: Ayanna Howard