One another game suggestion: The Turing Test, reversed.

Using LLM API, make people:

1. Ask one thing and spot a human answering them;
2. Answer one question and make another human believe they are an LLM.

Only these two rounds (or a text box + checkbox) per a session with a set delay, from a random user or a robot. The goal is counter to the current discourse of noticing LLMs being not humans, but the opposite – people being rewarded for acting like machines and spotting real people in the web poisoned by generative models. I don't know if scores are needed, just complimenting DMs maybe, because mastering tactics to break LLMs destroys the fun.

Besides only English input for the simplicity and a stop-list of words, it needs a balancing system that mixes fake and real inputs to cause no LLM-only hours across all timezones. To make it more interesting, there should be some additional prompts to LLM, like 'answer like an old lady' or 'answer like a nerd' to make LLM seem more humane.

Theming it after spies, zombies, whatever may help.