Early results show even the best commercial clones top out near 71% on the substantive-viewpoint axis, compared with above 90% on surface style — confirming what researchers have long suspected: style is easy, judgment is hard.
Research · March 22, 2026
Stanford HAI Releases First Open Benchmark for Personal AI Twins
The TWIN-100 benchmark scores how faithfully a personal AI clone reproduces a target individual's style, opinions, and decisions across 100 standardized prompts.