5 ways to make AI more trustworthy

A Waymo self-driving taxi. (Credit: Adobe Stock)

Self-driving taxis are sweeping the country and will likely start service in Colorado . How many of us will be lining up to take a ride?

That depends on our level of trust, says Amir Behzadan, a professor in the Department of Civil, Environmental and Architectural Engineering, and a fellow in the Institute of Behavioral Science (IBS) at ̽����Ƶ.

He and his team of researchers in the at ̽����Ƶ are unearthing new insights into how the artificial intelligence (AI) technology we might encounter in daily life can earn our confidence. They’ve for developing trustworthy AI tools that benefit people and society.

And in the journal AI and Ethics, Behzadan and his Ph.D. student Armita Dabiri drew on that framework to create a conceptual AI tool that incorporates the elements of trustworthiness.

“As a human, when you make yourself vulnerable to potential harm, assuming others have positive intentions, you’re trusting them,” said Behzadan. “And now you can bring that concept from human-human relationships to human-technology relationships.”

How trust forms

Behzadan studies the building blocks of human trust in AI systems that are used in the built environment, from self-driving cars and smart home security systems to mobile public transportation apps and systems that help people collaborate on group projects. He says trust has a critical impact on whether people will adopt and rely on them or not.

Trust is deeply embedded in human civilization, according to Behzadan. Since ancient times, trust has helped people cooperate, share knowledge and resources, form communal bonds and divvy up labor. Early humans began forming communities and trusting those within their inner circles.

Mistrust arose as a survival instinct, making people more cautious when interacting with people outside of their group. Over time, cross-group trade encouraged different groups to interact and become interdependent, but it didn’t eliminate mistrust.

We can see echoes of this trust-mistrust dynamic in modern attitudes toward AI, says Behzadan, especially if it’s developed by corporations, governments or others we might consider “outsiders.” So what does trustworthy AI look like? Here are five main takeaways from Behzadan’s framework.

1. It knows its users.

Amir Behzadan

Many factors affect whether—and how much—we trust new AI technology. Each of us has our own individual inclination toward trust, which is influenced by our ences, value system, cultural beliefs, and even the way our brains are wired.

“Our understanding of trust is really different from one person to the next,” said Behzadan. “Even if you have a very trustworthy system or person, our reaction to that system or person can be very different. You may trust them, and I may not.”

He said it’s important for developers to consider who the users are of an AI tool. What social or cultural norms do they follow? What might their preferences be? How technologically literate are they?

For instance, Amazon Alexa, Google Assistant and other voice assistants offer simpler language, larger text displays on devices and a longer response time for older adults and people who aren’t as technologically savvy, Behzadan said.

2. It’s reliable, ethical and transparent.

Technical trustworthiness generally refers to how well an AI tool works, how safe and secure it is, and how easy it is for users to understand how it works and how their data is used.

An optimally trustworthy tool must do its job accurately and consistently, Behzadan said. If it does fail, it should not harm people, property or the environment. It must also provide security against unauthorized access, protect users’ privacy and be able to adapt and keep working amid unexpected changes. It should also be free from harmful bias and should not discriminate between different users.

Transparency is also key. Behzadan says some AI technologies, such as sophisticated tools used for credit scoring or loan approval, operate like a “black box” that doesn’t allow us to see how our data is used or where it goes once it’s in the system. If the system can share how it’s using data and users can see how it makes decisions, he said, more people might be willing to share their data.

In many settings, like medical diagnosis, the most trustworthy AI tools should complement human expertise and be transparent about their reasoning with expert clinicians, according to Behzadan.

AI developers should not only try to develop trustworthy, ethical tools, but also find ways to measure and improve their tools’ trustworthiness once it is launched for the intended users.

3. It takes context into account.

There are countless uses for AI tools, but a particular tool should be sensitive to the context of the problem it’s trying to solve.

In the newest study, Behzadan and co-researcher Dabiri created a hypothetical scenario where a project team of engineers, urban planners, historic preservationists and government officials had been tasked with repairing and maintaining a historical building in downtown Denver. Such work can be complex and involve competing priorities, like cost effectiveness, energy savings, historical integrity and safety.

The researchers proposed a conceptual AI assistive tool called PreservAI that could be designed to balance competing interests, incorporate stakeholder input, analyze different outcomes and trade-offs, and collaborate helpfully with humans rather than replacing their expertise.

Ideally, AI tools should incorporate as much contextual information as possible so they can work reliably.

4. It’s easy to use and asks users how it’s doing.

The AI tool should not only do its job efficiently, but also provide a good user experience, keeping errors to a minimum, engaging users and building in ways to address potential frustrations, Behzadan said.

Another key ingredient for building trust? Actually allowing people to use AI systems and challenge AI outcomes.

“Even if you have the most trustworthy system, if you don't let people interact with it, they are not going to trust it. If very few people have really tested it, you can't expect an entire society to trust it and use it,” he said.

Finally, stakeholders should be able to provide feedback on how well the tool is working. That feedback can be helpful in improving the tool and making it more trustworthy for future users.

5. When trust is lost, it adapts to rebuild it.

Our trust in new technology can change over time. One person might generally trust new technology and be excited to ride in a self-driving taxi, but if they read news stories about the taxis getting in crashes, they might start to lose trust.

That trust can later be rebuilt, said Behzadan, although users can remain skeptical about the tool.

For instance, he said, the “Tay” chatbot by Microsoft failed within hours of its launch in 2016 because it picked up harmful language from social media and began to post offensive tweets. The incident caused public outrage. But later that same year, Microsoft released a new chatbot, “Zo,” with stronger content filtering and other guardrails. Although some users criticized Zo as a “censored” chatbot, its improved design helped more people trust it.

There’s no way to completely eliminate the risk that comes with trusting AI, Behzadan said. AI systems rely on people being willing to share data—the less data the system has, the less reliable it is. But there’s always a risk of data being misused or AI not working the way it’s supposed to.

When we’re willing to use AI systems and share our data with them, though, the systems become better at their jobs and more trustworthy. And while no system is perfect, Behzadan feels the benefits outweigh the downsides.

"When people trust AI systems enough��to share their data and engage with them meaningfully, those systems can improve significantly, becoming more accurate, fair, and useful,” he said.

“Trust is not just a benefit to the technology; it is a pathway for people to gain more personalized and effective support from AI in return.”