Bringing Responsible Thinking to AI

Beyond deciphering language, Diab, the director of the Language Technologies Institute (LTI), strives to translate what she has learned in six and a half years working in industry at both Meta and Amazon to train the next generation of AI scientists and impart the integrity that will lead to the responsible development and use of artificial intelligence. Diab endeavors to translate the technical aspects of these new tools into concepts that regulators and the public can better understand.

“I believe our field has an obligation in the ways we interact with society,” said Diab. “How do we build technologies that take into account people and cultures? And how do we develop the next generation of scientists, and instill in them the right values?”

Diab has coined the term “responsible thinking” as an important tenet of computer science training; as valuable and as critical, she believes, and as computational thinking. For Diab, responsible thinking means that the architects of new technologies must address broader societal issues such as privacy, culture, security, and diversity and inclusion.

Her drive to influence the next generation of technology architects inspired Diab to leave the private sector for SCS.

“Out in the real world, we have gaps we need to fill, specifically talent gaps when it comes to developing AI,” she said. “But how do we do it in the right way? I came here because I want to help fill that gap the right way, and CMU is the number one place in the world for AI.”

Building a bridge between technology and society

Born to Muslim Egyptian parents, Diab spent the first eight years of her life in England. She began learning Arabic in school at eight years old, and practiced by reading the Quran.

“When we moved back to Egypt,” she said, “I was treated differently because Arabic was not my native language. But I was also a geek, so I was acing all of the Arabic exams.”

From a young age, her parents emphasized the Islamic tenet of seeking knowledge. “The Quran is very clear that we should pursue knowledge wherever it is in the world,” she said. “It’s something a lot of people don’t realize.”

This principle is what inspired Diab to pursue higher education, specifically the intersection of technology and language.

“My perception of Islam is that it is very logical,” Diab said. “The whole scientific method — how you investigate a problem empirically — originated with Islam. For me, this notion of language and how it shapes thought is rooted in the Quran.”

When Diab wanted to study in America, it was the Islamic principle of seeking knowledge that propelled her parents to give their blessing, even though her extended family disapproved.

Diab’s education reflects her twin passions for computers and society. “I’m focused on the balance between humans and technology,” she explained. “I am very analytical, but I’ve always been interested in the humanities as well.”

Diab holds degrees in computer science and Egyptology, machine learning and computational linguistics. She’s spent portions of her career in both academia and industry, which led her to appreciate the human and social consequences of AI.

“When this opportunity came about [to lead the LTI], it was a big deal for me,” she said. “In industry, you can have a small impact on developing new talent.
In academia, it’s scaled up.”

Translating cultures

Diab’s global experiences give her special insights into how to address cultural issues in large language models and teach students to build nuance into generative AI tools.

For starters, when building a model that generates language, it is important to consider what information is available to guide the model. “The internet is not a good reflection of every society and culture,” Diab said. “It really depends on how much internet penetration you have in a specific community and what demographics are represented.”

In a current study, Diab’s research team is using surveys to compare how cultural trends in Egypt and the U.S. are reproduced on the internet and how they influence generative AI tools.

“For example, how important is it to take care of the elderly in your society?” she said. “In Arab society, this would be deemed as very important. In Western cultures, it would have a different value. Having these differences represented in technology is important.”

But information about the cultural norms in Egypt may not be as widely available on the internet, and therefore may not be accurately represented in generative AI models. “An AI model may mislead you in understanding a cultural value because the information is misconstrued on the internet,” Diab said.

In addition, language and translation are important components of how models translate cultural values. “So, we have to build systems that ask these questions in a smarter way, using a framework rooted in anthropological sciences,” she said.

Regulatory translation

Diab has begun work to bridge the divide between computer scientists and government regulators. The ultimate goal is an automatic regulatory compliance system that helps technology architects anticipate regulations and build compliant generative AI models.

“We want to make sure this new system doesn’t add to the burden of building models, but makes it easier for people to be responsible and governed in the right way,” said Diab. “We want to ensure the highest scientific integrity standards by anticipating and helping to shape regulations.”

Such a system would help foster better relationships between government regulators and technology builders, Diab said.

Consider the example of advertisements for online gambling on social media platforms. Many governments across the globe regulate who can see these ads. For example, regulations restrict U.S. platforms from showing them to people under age 18.

“This creates a complicated scenario because there is also a privacy aspect to social media accounts, so often the demographic information can’t be identified,” Diab said. “Companies find all sorts of loopholes to follow these types of regulations in practice, but it’s time consuming. Currently, you have legal and policy people getting together with technology people to implement this manually.

“The technology I am proposing could assist with this problem by distilling a lot of information, beyond what a single human could do,” Diab explained.

In addition to aggregating data, Diab envisions a system that would foster better communication between regulators and technology experts using chatbots and assistive communication technologies.

“It would essentially create and interface between the technology architects and the legal and policy people,” she said. Diab envisions the system would include features that translate technical language and concepts into explanations that policymakers can better understand.

“If they could have more fruitful conversations, they could have better compliance with the regulations, and together they would be able to create better regulations,” she said.

Championing Female Scientists

Beyond working to bridge divides between academia and industry, technology and culture, and regulators and scientists, Diab is pursuing another goal: elevating the role of women in science.

“I’m very passionate about women being able to be and do anything they want,” she said.

Diab recalls when she first began attending academic and technology conferences in the U.S. “Most of my formative years were in Egypt,” she said. “There are protocols that women wear dresses and makeup. In the U.S., I would dress up to go to a banquet at a conference and people almost looked down on me.”

These stereotypes of female scientists in Western cultures drive some women away from STEM fields, Diab said.

“In the U.S., female scientists are represented in the media as being nerdy or geeky — not feminine,” she said. “A lot of girls who have potential in the sciences step out of that realm because of this perception. You should be able to decide to look however you want. And women should not be considered less smart or able just because we are feminine.”

To cope with this issue in her own life, Diab has connected with six other female scientists who are leaders in their fields. They formed a group called “The Fashionistas,” and they meet regularly to discuss the difficulties of being a female leader in science.

Diab has also created a broader group on Facebook called Global Women in NLP (Natural Language Processing) — with more than 500 members.

“It’s meant to be a safe space for women to vent, get advice and propel their positions,” she said. “We also have a lot of conversations around sponsorships, mentorships, publications, how domestic issues can interfere with work — you name it.”

Supporting women interested in building large language systems is one more way that Diab hopes to influence the next generation of AI scientists. Diab truly believes that diversity and inclusion are critical for the sustainability of the field.

“Diversity and inclusion of ideas, cultures and people is the only way to unlock untapped potential leading to eventual growth,” she said. “I’m learning so much every single day. I truly believe that responsible thinking with diversity and inclusion as a core tenet is going to be the next frontier in our field, and I want to use it to train the next generation of scientists.”

It all comes down to bridging the gaps between the technology industry and academia, and between AI architects and regulators — forms of translation that truly highlight Diab’s superpowers.   ■