During a snowstorm last winter, University of Delaware philosophy professor Thomas Powers fixed his own thermostat by typing the error codes into ChatGPT and following the instructions it provided. As he discovered on that cold, snowy night, AI can make DIY projects that much easier, potentially saving time and money.

But ChatGPT’s instructions could have easily been wrong. Instead of clearing an error code, Powers could have caused a larger problem, wasting time and money, resulting in an unheated home during a serious cold spell. 

Ever the philosopher, Powers wondered about the ethics of his solution: What is the impact of taking the repair job away from someone who relies on the work? What about the risk of being wrong? Is the environmental impact of using AI worth the result?

After 20 years in the Department of Philosophy and as director of the Center for Science, Ethics and Public Policy, Powers is uniquely suited to explore questions about the ethics of AI. His class on the ethics and impact of engineering (ECE/ELEG 491) is a breadth requirement for engineering majors. Powers also teaches a graduate course on ethics in data science and AI. 

He isn’t the only one asking those questions. As seen in recent issues of The Economist and The New York Times illustrate, AI companies are too. 

UDaily spoke with Powers about applying ethics to AI, what it means for the future and how philosophers are helping shape the conversation. 

What does “ethical AI” mean? 

We want AI systems that are helpful and useful, but also safe. We want systems that don’t produce biased or dangerous results. One approach to achieve that, which I’m hopeful about, is to create systems that are “safe by design” or “ethical by design.” In other words, instead of relying entirely on outside regulation to prevent harmful behavior, we try to design the systems themselves so that they behave in ways we consider ethical.

Is it possible to make an AI system truly “safe by design” if it eventually becomes capable of acting in ways its creators didn’t anticipate? 

That's an excellent observation, and a recent essay by Anthropic's CEO Dario Amodei seems to cast doubt on the efficacy of technical guardrails, given the propensities that large language models (LLMs) have displayed. 

There is a big difference between most technologies and AI. With the latter, especially current LLMs, it appears that there is some unpredictability in their behaviors. Thus, their apparent autonomy complicates the "safe by design" motto. AI autonomy introduces the need for "guardrails" that constrain what AIs can do. But even this may not be possible. 

Why are AI companies hiring philosophers, and what does this mean for the field of philosophy?

Some of the hardest questions in AI aren’t purely technical questions. To build systems that are safe, fair or aligned with human values, you first have to decide what those concepts actually mean and how they can be translated into something a computer system can use. The challenge is to educate people who can move between those worlds — who understand the technology but can also ask the deeper questions about what we should want that technology to do.

Are universities preparing future employees to ask those questions?

I think we're doing a good job, but it's an incredibly hard job. And that’s partly because these models, especially LLMs, have been coming out at such a rate and the improvements and with them the more wide-reaching abilities that it's really hard to get a handle on.

Another challenge is that faculty have very different perspectives. Some are very optimistic about AI and see tremendous possibilities for teaching and research. Others are deeply concerned, particularly in humanities fields such as English, history and philosophy, where research papers and independent writing have traditionally been central to teaching. For those faculty, AI has really upended the way they teach.

But I don’t think universities can simply tell students, “Never use AI.” There are helpful and legitimate ways to use tools such as ChatGPT, including as a starting point when students are learning about an unfamiliar subject. At the same time, AI is already forcing faculty to reconsider how they assess student learning.

How do your classes bridge that gap?

It all starts with the syllabus. You have to select readings that have some philosophical ideas or discussions in them, but you also have to have readings that are well informed about the most recent science and technology. 

One key issue we read about it is privacy. To better protect individual privacy, scientists can use mathematical techniques to introduce “noise” into data sets. This allows us to uncover population-level patterns in data while maintaining the anonymity of individuals. We also discuss the moral and legal implications of privacy. Authors like Shoshanna Zuboff and Cynthia Dwork write a lot about the intersection of AI technology and philosophical issues. 

How has AI changed philosophy?

Philosophers have long discussed the philosophy of action with the concept of agency. For decades we assumed that an agent must be human — someone who acts on their own behalf, or on the behalf of another person, with some kind of plan or attitude. 

This concept transfers to discussions about AI, and whether it can be considered an agent and be responsible for its actions. 

The next step is to ask harder questions: Are autonomous AI agents responsible, in some sense, for what they do, and does this suggest or imply that they are conscious?

I don’t think AI will become conscious any time soon, but I don’t think it is impossible, and other philosophers and scientists are broaching the question. 

What do you think the future holds for AI on campus? 

For quite some time, it seemed like AI wasn’t going to go anywhere, and then, particularly over roughly the last decade, things came together very quickly. 

Ultimately, I’m hopeful that faculty who are enthusiastic about AI and those who are skeptical of it can find some kind of rapprochement — or at least develop a set of policies that everyone can live with. But it isn’t going to be easy.