I’ve always felt slightly awkward in cocktail conversations. There’s a real art to pleasing a crowd without sounding bland — saying something easy to grasp, yet distinctive enough to reveal a bit of personality and invite others to continue the exchange.
In that sense, AI may be the perfect cocktail conversationalist. It is trained (mostly) on how people speak in public, not how they talk in private.
In this second micro-game in a series developed with Anthropic’s Claude Code, I wanted to explore this question: when given the opportunity to express an opinion, do we default to the most average, predictable responses (those that an AI could easily generate)? Or are we willing to express something else?
In the game, players pick a topic and construct an opinion — something they might plausibly say at a cocktail party — using autocomplete-style phrase fragments. Once they’ve finished, they can see how predictable their stance is, ranging from “crowd pleaser” (highly predictable) to “safe,” “conversational,” and more “fizzy” responses — the most peculiar or original.
Players can choose from pre-curated themes such as diversity, the rise of AI, and freedom, or introduce their own topics, generated in real time by the system.
At the time of writing, the AI behind the game isn’t entirely convincing. Breaking sentences into branching autocomplete options turns out to be difficult, and it often distracts the system from the core task of producing statements that span the spectrum from predictable to original.
On the plus side, some generated sentences are unexpectedly poetic. More importantly, the game reveals the subtle, almost insidious influence of autocomplete on our thinking — nudging us toward ideas that feel like our own, without quite being so.
Then again, in many social contexts, is that really the point? Human expression is often performative, as at cocktail parties, where we prefer not to reveal too much. In a previous experiment, “Rhetoric Fighter,” I explored what we prefer to hear — AI-generated rhetoric or more original human voices. Here, I turn to what we prefer to say: something truly our own, or something half-scripted.
It leaves me wondering whether the success of AI comes not from its ability to challenge some mythical, deeply original human voice, but from how closely it mirrors what we already expect — and perform — in public.
The game was developed using Claude Code and the API (Haiku 4.5).