Google opens LaMDA to the public through AI Test Kitchen

Google launched AI Test Kitchen, an app for the public to try limited demos built on its LaMDA conversational model, starting with a gradual rollout to users in the US who signed up for a waitlist. The app shipped on Android first, with an iOS version to follow in the coming weeks. Three demos were offered: “Imagine It,” which generated descriptive riffs on a named place; “List It,” which broke a stated goal into a list of subtasks; and “Talk About It (Dogs Edition),” a conversation constrained to stay on the topic of dogs.

Google described the demos as experimental rather than finished products, and built in several safety layers — output filtering and improvements to the underlying model’s handling of unsafe or biased content — while acknowledging in the announcement that LaMDA could still produce inaccurate, biased or inappropriate responses. User ratings within the app fed back into identifying problem outputs for future correction.

The launch arrived into an unusually charged context. Two months earlier, Google engineer Blake Lemoine had publicly claimed that LaMDA had become sentient, based on transcripts of its conversational responses; Google rejected the claim and fired him in July. AI Test Kitchen was the first time Google let members of the public interact with LaMDA directly rather than read transcripts curated by Google or by Lemoine, at a moment when the underlying question — what a fluent conversational model’s outputs do and do not imply about what is happening inside it — was already a matter of public argument rather than a settled research question.