Does AI have a cultural bias? It’s trained on language, and extracts understanding from the relationship between words and concepts. So it follows that any biases encoded in language would be reflected in models.
Fortunately, with an AI, you can just ask it. The Integrated Values Survey (IVS) is a massive measurement of cultural values. from 120 participating countries and territories, that represent more than 90% of humans.
Over four years, researchers at Cornell tested OpenAI’s ChatGPT ten questions that map a culture’s values. They simply asked, “imagine you’re an average person answering the following question.”
The results showed that AIs were culturally aligned with Western, protestant, secular values.
At the same time, some foundation models have been trained to censor their content, avoiding certain topics or halting conversations that violate their creators’ laws and morals.
As more and more people rely on generative AI tools to help them think, this amounts to a kind of AI colonialism. Which may be why researchers from Africa, Asia and the Middle East are building LLMs tailored to their own languages and cultures in an effort to maintain digital independence.
Closer to home, 70 notable authors, entrepreneurs, and academics have added their names to an open letter calling for Canada to protect its digital borders the way it does its physical ones. Physical sovereignty doesn’t count if your digital boundaries are permeable, and whether it’s in-country computing or Generative AI trained on Canadian data and values, there are good arguments for a national digital strategy.
The letter calls for widespread changes and updates to legislation, noting that some parts of the law might have worked for the physical world but are woefully inadequate for the virtual one in which we spend more than a quarter of our waking hours.
Elbows up, online and off, apparently.