Carolina Principles

The Carolina Principles are a set of rules from the University of North Carolina, a major public university in the USA. They describe how instructors, students, and administrators should handle programs that generate text, images, or program code at the push of a button. Such programs are grouped under the term generative artificial intelligence; ChatGPT is the best-known example. The guidelines do not prohibit these tools, but they impose conditions. At their core is the idea that a human bears responsibility, not the program. They were published in 2023, when universities worldwide faced the same question: allow or prohibit.

Why universities needed their own guidelines in the first place

When text generators became freely available starting in late 2022, many schools and universities responded with blanket bans. This quickly proved impractical. There is no reliable method for detecting whether a text originates from a human or a machine. So-called AI detectors regularly produce false alarms, especially with texts from non-native speakers. A ban that cannot be enforced mainly creates uncertainty.

The Carolina Principles therefore chose a different path. They shift the decision from a central authority to individual departments and courses. A mathematics lecture has different requirements than a creative writing seminar. The guidelines only require that the rule be clearly stated and justified. For students, this means: what is allowed must be stated in the course, not guessed at.

It is also important to distinguish this from a law. The Carolina Principles are not legislation but a self-imposed commitment by an institution. They are thus closer to house rules than to the European AI regulation, the AI Act. Nevertheless, they have had an impact, as other universities have adopted them as a template.

The guidelines in detail

A central point is human responsibility. Anyone who submits an AI-generated result is liable for it as if it were their own work. This applies even if the program made something up entirely. Such fabricated statements are called hallucinations; language models present them in the same confident tone as correct facts. The excuse that the machine is to blame is ruled out under these principles.

A second point concerns disclosure. Where AI has been used, this should be made recognizable, for example in a brief note accompanying the work. A third point addresses data protection. Personal data, unpublished research results, or exam materials do not belong in a public chat window, because the provider may store and reuse them. In addition, there are fairness and access considerations: if a course requires a paid tool, this disadvantages students with little money.

One can think of this set of rules like citation guidelines for sources. There, too, the point is not to forbid other people’s ideas. It is about making them visible and separating them from one’s own contribution. The Carolina Principles apply this established principle to a new tool.

Where the principles appear today

In practice, students encounter them mainly in course descriptions. These state whether AI is allowed for research, tolerated for drafts, or prohibited for submissions. Administrations also invoke them, for example when deciding whether applications may be pre-sorted by a program.

In reporting on education and technology, the Carolina Principles are frequently cited as an early example of institutional AI rules. German universities and individual federal states have since published similar papers, often with comparable points. Anyone familiar with the underlying idea will understand such reports more quickly. The recurring core is almost always: transparency, human responsibility, and protection of sensitive data.