Measuring the social, ethical, and political values of different AI models

I continue to enjoy reading my hard copy of The Economist every week. Occasionally, I post examples of insightful graphics presented in articles.

Here are some graphics that struck me recently. They are taken from a fascinating article. AI models’ values are very different from most people’s

The first figure shows a comparison of the answers given to the World Values Survey by different Large Language Models and people from different regions (cultures) of the world. The vertical axis goes from traditional values at the bottom to secular values at the top. The horizontal axis goes from survival (or communal) values on the left to self-expression on the right.

Note how the different cultures do not overlap. Furthermore, the LLMs are distant from almost everyone. Just a few are close to the English speaking world. 

The second graph compares answers to the VOTER survey of political questions. The horizontal and vertical axes correspond to Social and Economic issues respectively. Positive and negative values correspond to conservative and liberal respectively. Answers from LLMs are compared to a sample of Trump and Biden voters from the 2020 USA Presidential election. Note the clear political polarisation. Furthermore, the AI models are all liberal.

I offer no comments on whether any of this is good or bad. I do suggest two implications. First, extensive use of AI will tend to shift peoples values in a particular direction, just like media (music, TV, movies, art) does. Second, given the conflict of values with most people these biases will offend or concern many communities and increase the backlash against AI companies and its widespread adoption. Consequently, some will modify the training of their models so they align more with the values of vocal communities. Whether or not that is seen as a good thing, will depend on whether you think some of those values should be affirmed.

What do you think?