Q&A: AI ethics expert sees philosophy as critical to AI debate

Cansu Canca is a 2024 Mozilla Rise25 honouree. Mozilla’s Rise25 awards celebrate the people leading the next wave of AI - using philanthropy, collective power, and the principles of open source to make sure the future of AI is responsible, trustworthy, inclusive and centred around human dignity. 

Context is a partner of Mozilla Rise25.

RICHMOND, Virginia - As artificial intelligence rapidly changes the way people live, work and even think, it is the daily job of Cansu Canca to ponder the big philosophical and ethical questions surrounding the expanding deployment of AI.

To Canca's thinking, philosophy helps strengthen and steer her quest.

"In some sense I think if I was not a philosopher, I couldn't have approached this core question of AI ethics," she said.

Some organizations, she says, can look for ways to simply check a box on ethics in AI.

"If you are not coming from an analytical philosophy mindset," she said, what you often mistakenly "end up looking for is a checklist: 'Did we comply with X, Y, and Z? .... Now just go forward with the innovation – that is the exciting part.'"

"For us, ethics in innovation is exciting in itself."

Canca is the founder of the Cambridge, Massachusetts-based AI Ethics Lab, one of the first labs dedicated to crafting systematic programs for ethics in AI.

Canca is also Director of Responsible AI Practice at Northeastern University's Institute for Experiential AI and a research associate professor at Northeastern.

She has a Ph.D. in philosophy, specializing in applied ethics, has served as an ethics advisor to Fortune 500 companies and works with the World Economic Forum's AI Governance Alliance on guidelines and best practices.

She spoke with Context about how philosophy is central to ongoing debates over AI and ethics, where gaps persist in the ethical deployment of AI and potential risks moving forward.

This interview has been edited for length and clarity.

You obviously have a significant background in philosophy. How has that colored your experience with AI and ethics?

Once we start talking about AI ethics, you are necessarily talking about philosophy.

As soon as you want to go beyond simply asking 'Is this fair?' you have to start thinking about 'What does fairness mean in this context?' and that is a deeply philosophical question.

There are also many other philosophical questions that are arising with new AI systems, like large language models (LLMs)– what is the relationship between language and mental models? What is the relationship with language and capabilities like reasoning? What is necessary and sufficient to have human-level or beyond human-level capabilities – perhaps not mind or consciousness, but just language.

And this draws in philosophy of mind, philosophy of language.

So I would say I'm even more glad to be a philosopher at this point.

In the past you've touched on bias in AI – have you seen any meaningful improvements (or) attempts at improvements at such issues over the last five years or so?

I think there is definitely improvement, especially when you think about the famous cases. Both the developers and deployers, for example, now know that facial recognition is a sensitive technology, and we need to be careful. It's already well exposed in its weaknesses.

The same, though, is not true when you are dealing with less-showcased use cases. Unfair bias is a critical concern for most AI systems that deal with humans and society. For example, when you're creating a recommendation or ranking systems for which opportunities to present in education, in finance, in jobs - basically anything that matters for people's life choices, it becomes crucial what they get to see and what is judged to be 'not suitable' for them.

I said this so many times already, but I think it still is important to keep in mind - we are viewing the world through an AI-mediated structure.

Through your social media or your job search or your education search – things are ranked and structured and recommended for you. This makes sense because the information is so abundant that you cannot deal with it if it is not structured.

Structuring makes sense. But ... there's always a value judgment within that structuring.

And when we are dealing with unjust biases in this context, what we are really dealing with is shaping and limiting one's world because we/the AI judges that they just don't need to know the rest of the world and its opportunities.

The system basically decides and in effect chooses for you, looking at your gender, your socio-economic background, maybe you don't need to know about this high-paying job, this great educational opportunity, this good credit card.

And that is a decision that is as relevant, as important, if not more (so) in many cases, than the biases in facial recognition. It's still very hard to make sure that organizations pay sufficient attention.

To what extent has 'ethics washing' improved in practice when it comes to AI?

The industry is doing better on that front, but it's absolutely not even close to doing great.

You see this also in the trends of when they are hiring and when they are firing. The incentive still seems to come directly from the regulatory and market perception.

Most companies still don't see it as an integral part but rather almost like a part of communications and PR – which is absolutely not the right attitude to have.

What do you see as some of the major potential pitfalls and risks for companies that mean to use or leverage AI well?

I think one way to think about this is instead of focusing on the thing, which is AI, we should be focusing on, what is your worry? What is your concern?

It doesn't really matter whether this is an LLM or this is a predictive model or this is some unreliable search engine that you're using – we should probably just ask you what we are concerned about.

The AI part should inform how we formulate the questions, but if we just focus on this big, bad AI that we are chasing, it's hard to explain what ... is really the risk that we are concerned about and trying to avoid.

When we are integrating something like this into our way of operating, we also need to be very careful about what other side effects and unexpected changes we might see and prepare for.

If you think about organizations using different LLMs for different tasks and connecting those LLMs, what is potentially getting lost in communication?

If my AI is reading your email and responding to your email, where is my thinking going into this? (If) I'm just reading the AI summaries, am I getting the nuances of what you really meant, which I would get with your facial expression, with your conversation?

So it's almost like we are creating a web of AI systems interacting with each other and we have to be mindful (of) where the human interaction is necessary and desired.

Because it's very easy to sort of lay back and watch tasks being done and, while doing so, losing a lot of creativity and value in the process.

(Reporting by David Sherfinski; Editing by Ellen Wulfhorst.)