Today we learn two things:
- We are almost all WEIRD (Western, Educated, Industrialized, Rich, Democratic).
- AI models like ChatGPT, which are trained by analysing the Internet (especially a predefined 44TB text dataset), mostly reflect this Western mindset, showing a strong Anglo-Saxon cultural bias: they think, evaluate and respond more like Americans or Northern Europeans than people from other cultures.
To create a truly global AI, we need more inclusive and multicultural data, evaluators, metrics and rules.
If you’re wondering whether you can access that 44TB dataset… of course! It’s open source. I’ll explain it in my next newsletter 2042, along with a full analysis of the study: https://scholar.harvard.edu/sites/scholar.harvard.edu/files/henrich/files/which_humans_09222023.pdf
