Human Behavior Bench tests whether AI can predict group choices
A new benchmark from Rehearsals, Google DeepMind and Kaggle evaluates whether AI can predict response distributions for specific U.S. cohorts across 1,199 tasks.
2 min read
Rehearsals, Google DeepMind and Kaggle have introduced Human Behavior Bench, a benchmark that asks AI systems to predict how specified groups of people will answer questions rather than merely imitate an average respondent.
The benchmark contains 1,199 evaluation tasks derived from 95,553 responses by 1,244 U.S. adults. Data was collected through voice interviews and forms between June and September 2026. Each task asks a model to predict the response distribution for a named audience or cohort, and scoring uses 100 × (1 − total variation distance), so closer distributional predictions receive higher scores.
The organizers report that Rehearsals’ Engram system scored 94.1, compared with 80.6 for Gemini 4 Argon, 79.5 for GPT-5.6 Sol and 79.3 for Claude Opus 5.5. These are organizer-reported results. Engram was evaluated in-house, while base models were run through Kaggle, so the leaderboard should not be treated as a perfectly controlled independent comparison.
The benchmark intentionally selects tasks where a cohort differs from the general population. That makes subgroup prediction central to the test, but it also means the score is not a broad measure of general intelligence or universal human understanding. Its U.S.-only sample and limited collection period further constrain generalization across countries, cultures and time.
Kaggle is expected to maintain the benchmark and update model results. Public submissions are planned but were not open at announcement, limiting immediate third-party verification. Human Behavior Bench is most useful as a focused test of population-level prediction—and as a prompt to examine what models miss about real groups.
Image credit: Photo by Vitaly Gariev on Unsplash, used under the Unsplash License.
Sources: Human Behavior Bench announcement and methodology · Image · Unsplash License
Conversation
0 approved commentsNo comments yet. Start a thoughtful conversation.
Join the conversation.
Sign in with ChatGPT to leave a comment or rate this article.