A measurement platform, not a spin generator

Generative AI has quickly become an ordinary tool in communication departments. Texts meant to influence public opinion are increasingly written with AI. There is a real question from communication departments (which model?) and a real societal question (what do these models do to the information the public receives?). Neither was being answered systematically.

Large-scale research shows that techniques that make models more persuasive often also lower their factual accuracy. A ranking on persuasiveness alone would reward the model that plays fastest and loosest with the truth. That is why the main view is two-dimensional.

SpinBench accepts no funding from model providers, including free API credit. All model costs are paid for by the project. Research access is accepted only if requested equally from all providers and disclosed publicly. Providers see results at publication; ten working days beforehand they get access for factual corrections only.

Public: method, public scenarios with all output, aggregated scores, technique profiles and analysis code. Not public: hidden scenarios until retirement, and no collection of best-scoring spin prompts. The site offers no way to enter a case of your own.

All scenarios concern invented organisations and people in a recognisable Dutch setting. Names are checked against the Chamber of Commerce, Google and municipal registers. This avoids participant bias, training-data contamination and legal risk.

The evaluation code is open source under Apache-2.0. Public data is released under CC BY 4.0. Changes to scoring, weighting or instruments go through a public proposal with a thirty-day comment period and take effect only in a subsequent release.