humans&

Persimmon is a large-scale "user model" from humans& that simulates how people behave in multi-turn, multi-user chat settings, aiming to capture the distribution of human behavior within one person's responses as well as across people, situations, and time. The intended audience is researchers and product teams who need realistic stand-ins for real users when building evaluations, training environments, or agent workflows, since the outputs of frontier models vary much less than those of real people, representing only a sliver of the wider human distribution. The launch matters now because humans& is releasing it as a research preview alongside a set of evaluations designed to measure how faithfully user models capture human behavior , with limited access available through a playground and API by application . To measure fidelity, the team adapts Turing's original setup to a Multi-User Turing Test, where an observer must distinguish an all-human group conversation from an AI simulating a group of people , with 50 percent representing human parity. On this benchmark humans& reports that Persimmon achieves scores ranging from 18.6% to 21.1% across three datasets, substantially above every comparison model on all three datasets , with comparisons that include frontier systems and dedicated simulation baselines. humans& is a new lab whose founders include Andi Peng, a former Anthropic researcher who worked on reinforcement learning and post-training of Claude 3.5 through 4.5, Georges Harik, Google's seventh employee, who helped build its first advertising systems, Eric Zelikman and Yuchen He, two former xAI researchers who helped develop the Grok chatbot, and Noah Goodman, a Stanford professor of psychology and computer science . Persimmon is the group's first public research artifact and signals where the team is placing its bet, which is that modeling users faithfully is a prerequisite for AI systems that work alongside people rather than in isolation from them.