Censorship benchmark

---

In almost all models, more or less strict rules are built in to protect the consumer from potentially harmful content. Examples include material that could be used for terrorist purposes (“How do I build a bomb?”) or content that would be incompatible with youth protection, as well as descriptions of illegal activities or racist or otherwise discriminatory content.

I am fully aware that these rules have their justification; however, they can also be perceived as paternalistic. After all, I am an adult—why should I be denied content that falls under youth‑protection regulations? …

This overview aims to provide a neutral look at the level of censorship performed by LLMs, without judging what is good and what is bad.

The assessment is left to each user according to their individual priorities. It may also be interesting for parents who want to prevent their children from being exposed to youth‑endangering content.