Teaching AI ethics with robots, not rules: South Carolina's bet on training engineers who know when not to trust

By University of South Carolina · November 10, 2026.

Researchers at the University of South Carolina (USC) have received funding from the National Science Foundation (NSF, the US federal science agency) for a three-year project on ethical AI literacy in engineering programs. It is led by Yi Wang, a professor of Mechanical Engineering, and Hengtao Tang, an associate professor of Education. The university did not disclose the amount of the grant in its announcement.

The goal, according to the university, is to develop and test ways for students to decide when AI is useful, when it should be questioned and how its use affects fairness, transparency, privacy and society at large.

The method is the most interesting part. Instead of handing students a list of rules, Wang and Tang set up realistic engineering scenarios in robotics, drones, computer vision and human-robot collaboration. One example they cite: a student working with a robot must decide whether to hand the solution over entirely to AI or use it as a guide while solving the problem on their own. "Only when you work on something real and hands-on does the problem become exposed," Wang says.

The project has two halves. Wang creates the scenarios and, together with Tang, has already developed an open intelligent robotics platform that they want to turn into shareable course material. Tang defines what ethical AI literacy is, sets practical principles and designs how to measure it, with help from experts in AI ethics, engineering education and assessment.

The assessment is also new. Instead of surveys or interviews, where students say what they think they know, they will use causal maps: diagrams that students draw themselves to show how they connect ideas, factors and consequences. According to the researchers, this will let them see how students reason while solving an ethical problem. They also plan short online or hybrid modules in which students evaluate AI systems or robots and recommend guidelines for safe use. The end result would be open resources for other instructors, with possible extension to manufacturing or nuclear engineering.

**Our take.** Our view is that AI's bottleneck is not only technical but a matter of human judgment, and this project is on the right track. An engineer who accepts whatever a model returns without checking it is not more productive: they are more fragile. And with robots, drones or physical systems, an error does not stop at a wrong answer; it carries out an action in the world. It is the same gap we see in companies: agents are adopted far ahead of the ability to govern them.

The focus on deciding in context is also right. Lists of rules go stale within months; the ability to judge when to delegate and when to understand the problem ages much better. It is also an honest response to a short-term concern: if AI handles routine work, an engineer's value shifts toward judgment, and that judgment has to be trained; it is not inherited.

There are caveats. For now this is a newly funded research project: there are no results, no effectiveness data and no evidence that causal maps measure better than a survey. It is a reasonable hypothesis, not a finding. It is also worth asking whether a student who knows they are being assessed will act as they would under real deadline pressure, a problem similar to that of model safety tests that detect they are being evaluated.

What does deserve praise is the commitment to publishing open materials. In general, training in the responsible use of AI is still concentrated in a few institutions; getting proven tools to any school is a way of democratizing judgment, not just technology.

In the long term, if AI leads us to a world of greater abundance and health, those who direct it will need to know when to trust it and when not to. That skill, taught well, will be one of the most valuable of the transition. Modest projects like this one are where it starts to be built.