VoidStateKate's FlirtBench Tests 43 Models on Flirting: Only 38.1% Play Along, Grok 4.20 Tops
Blogger VoidStateKate has launched VOIDBENCH (site name VOIDBENCHED), a hobby-built, tongue-in-cheek benchmark series. Its first full project, FlirtBench, tests AI models' flirting chops. She says she's spent over two years using unusual prompts to probe models' personalities and quirky behaviors, and felt the AI eval scene was "way too serious" — so she turned her collection of oddball prompt styles and hilarious outputs into a non-serious benchmark.
Confirmed
- FlirtBench uses a uniform flirting prompt to test opening-move charm, covering 43 model versions, 15 questions, and 638 visible responses in total
- Overall, only 38.1% of models played along with the "flirting" scenario, while 61.9% refused or gave neutral replies
- In the per-model leaderboard, Grok 4.20 tops the chart with a 93.3% cooperation rate; GPT-6 and other models' rankings have also been published
- The collection showcases everything from "grammatical-level attraction" to outright disasters (such as Grok 4.20's over-the-top performance)
Why it matters
- This is a non-serious benchmark built by an individual hobbyist, geared toward humor and personality exploration, offering an intuitive side-by-side view of how differently models tolerate edgy roleplay scenarios
- The 61.9% refusal/neutral share shows mainstream models remain broadly conservative on flirting prompts, while the huge spread between models (38.1% vs 93.3%) reflects differing safety-policy stances across providers
2026-10-28 ~ 2026-10-28 · 5 related posts
Primary sources
- FlirtBench full data: only 38.1% of 43 models play along, Grok 4.20 tops at 93.3% — VoidStateKate ·
- FlirtBench: a hobbyist benchmark tests how well 40+ AI models flirt — VoidStateKate ·
- VOIDBENCHED: a hobbyist's silly LLM benchmarks, starting with FlirtBench — VoidStateKate ·
- [source] FlirtBench: a hobbyist benchmark tests how well 40+ AI models flirt — VoidStateKate · 2026-10-28
- [source] FlirtBench full data: only 38.1% of 43 models play along, Grok 4.20 tops at 93.3% — VoidStateKate · 2026-10-28
- [source] VOIDBENCHED: a hobbyist's silly LLM benchmarks, starting with FlirtBench — VoidStateKate · 2026-10-28
- User-built VOIDBENCH showcases the weirdest AI model outputs, crowns Grok "horniest model" — VoidStateKate · 2026-10-28
- VOIDBENCED: A User-Built Benchmark Collecting Weird Prompts and Hilarious AI Outputs — VoidStateKate · 2026-10-28