At December’s Lunch & Discussion, the Tang Institute hosted MIT postdoctoral associate Kevin Mills for a conversation about moral guardrails” in AI and related technology. These features are designed to prevent the use of the technology in ways that are judged to be immoral or ethically controversial. The question at hand was when, if ever, we should expect these guardrails to be imposed by technology developers. 

Anyone who has used the popular AI systems ChatGPT or DALL‑E is likely to have brushed up against these guardrails. ChatGPT won’t discuss politically sensitive issues or provide information that may insight violence or hate. For example, if you ask it to write a story about the covid vaccine causing major health issues, it will refuse. Or if you ask it to help you create an explosive device, it will likewise reject your request. DALL‑E — the AI image generator — will similarly refuse to generate violent, hateful, or adult images.

As Mills notes, most of us believe that there are good reasons to deploy certain guardrails like this. But doing so also imposes a cost: these features of technology do limit people from pursuing ends or projects that we deem them to be free to pursue. So when is the cost justified? And when is the curtailing of user freedom too steep?

To help disentangle the contours of that question, Mills noted that it is worthwhile to consider further distinctions between moral obligation, permissibility, and praiseworthiness.

Note that sometimes it is permissible for us to perform an action. However, we are not required to do it. If someone is raising their fist at me, it is acceptable for me to defend myself with the requisite force. But I am not required to do so. It is up to me. Other times, we are morally required to act in certain ways. Nobody has the right to go out of their way to inflict bodily harm to another for no reason whatsoever. And still yet, sometimes what we do — though not obligatory — is not only permissible, but also praiseworthy. For example, if there is a building on fire near me and I run to save the resident cats, some may praise me. But it was not expected for me to do this. It went above the usual duty of a regular citizen.

With these distinctions in mind, Mills discussed considerations that may make it required to implement moral guardrails, considerations that may make it permissible to implement moral guardrails, and considerations that may make it praiseworthy to do so. Perhaps of particular interest, Mills raised a question about whether these features can ever truly be praiseworthy.


Phillips Academy and the Tang Institute thank Dr. Mills for the fascinating discussion of a pressing issue in the ethics of technology. We look forward to continuing this year’s speaker series on the ethics of AI.



Kevin Mills is a postdoctoral associate in SERC (the Social and Ethical Responsibilities of Computing) at MIT. His research interests include privacy, online manipulation, misinformation, and the ethics of artificial intelligence.

Back to Top ↑

Be a part of our community!

Subscribe to our newsletter, Notes on Learning, for monthly updates.