AA-PersonalityBench: open weights models lag proprietary on truthfulness
The gap pattern suggests open weights models optimise toward punchy, scannable output at the cost of factual care and structure, a trade-off worth weighing when picking models for enterprise or research workflows. Teams choosing between open weights deployments and proprietary APIs can use the 30-trait percentiles as a filter beyond raw intelligence scores. The benchmark also gives prompt engineers concrete levers for compensating where open models fall short.
Artificial Analysis introduced AA-PersonalityBench, which percentile-ranks models from 1 to 100 on 30 personality traits. Comparing the 10 highest-scoring open weights and 10 proprietary models by the Artificial Analysis Intelligence Index, open weights models averaged 32 points lower on Truthful, the largest gap of any trait. Open weights models also ranked lower on Pragmatic (79 vs. 49), Natural Sounding (77 vs. 52), Evidence-Based (76 vs. 47), Structured (73 vs. 45) and Literate (70 vs. 47). They ranked higher on Slop (30 vs. 49), Bottom-Line Summary (47 vs. 62) and Scannable (41 vs. 53).
In AA-PersonalityBench, the top open weights models on average rank lower than the top proprietary models on truthfulness, pragmatism and structure, and higher on slop and summarizing the bottom line
We compared the 10 open weights and 10 proprietary models measured with the highest Artificial Analysis Intelligence Index scores. Each model is percentile ranked from 1 to 100 on each of 30 personality traits. On average, the top open weights models rank 32 points lower than the top proprietary models on Truthful, the largest gap of any trait.
Average rank, top proprietary vs. top open weights:
➤ Open weights rank lower: Truthful (79 vs. 47), Pragmatic (79 vs. 49), Natural Sounding (77 vs. 52), Evidence-Based (76 vs. 47), Structured (73 vs. 45) and Literate (70 vs. 47)
➤ Open weights rank higher: Slop (30 vs. 49), Bottom-Line Summary (47 vs. 62) and Scannable (41 vs. 53)
