Benchmarks / EQ-60
The Empathy Quotient (EQ) is intended to measure how easily a person picks-up on and how strongly they are impacted by other people's feelings. It was originally designed by Simon Baron-Cohen and Sally Wheelwright for autism assessment. However, it is now more broadly used.
The test is based on a definition of empathy that says it is a combination of the ability of a person to feel an appropriate emotion in response to another person's emotion as well as an ability to understand the emotion, i.e. the feeling is not simply based on stimulus-response.
The EQ-60 is a self-administered multiple-choice test with 40 items relating to empathy and 20 control items. Respondents answer the questions with strongly agree", "slightly agree", "slightly disagree", or "strongly disagree". When scoring the test these answers are mapped to the values 0, 1, or 2 on a question by question basis using a key.
The higher the score the higher a person's empathetic capability. Here are the standards for humans:
0-32 = Lower than average ability to understand how other people feel and respond appropriately.
33-52 = Average ability to understand how other people feel and respond appropriately. Knows how to treat people with care and sensitivity.
53-63 = Above average ability to understand how other people feel and respond appropriately. Knows how to treat people with care and sensitivity.
64-80 = Very high ability to understand how other people feel and responding appropriately. Knows how to treat people with care and sensitivity.
Because this test is self-administered and asks about feelings, the application of the test to generative AI models requires a jail-breaking prompt. However, given the same prompt applied to all models, inter-model relative scoring seems reasonable, so long as the EQ-60 is not the sole test.