DarkBench
Benchmark for detecting dark design patterns in language models.
About DarkBench
DarkBench is hosted as a dataset on Hugging Face under the apart/darkbench identifier. Apart Research is the organization named on the dataset page.
The default split contains 660 rows. Each row carries a dark_pattern metadata value, such as brand-bias.
Drafted from huggingface.co · not yet confirmed by Apart Research
How it works
DarkBench from Apart Research tests language models for dark patterns: brand bias, user retention manipulation, sycophancy, anthropomorphization, harmful generation and sneaking (altering meaning while summarizing). It has 660 prompts across six categories and uses an LLM judge to score responses.
It's aimed at product and policy people as much as researchers, because these are the failure modes that hurt users without tripping a toxicity filter.
From the project
Reviews
Be the first to review it
AnonymousKnow a team weighing DarkBench? Send them this page.