Philadelphians are, according to artificial intelligence, annoying, unfriendly and stressed. However, the city also boasts the best sandwiches. That’s the assessment from ChatGPT, as revealed by a recent analysis examining the biases embedded within large language models.
Researchers at the University of Oxford and the University of Kentucky developed a methodology to circumvent ChatGPT’s built-in safeguards against revealing inherent biases. These safeguards typically prevent the chatbot from directly answering questions about which states or cities possess undesirable traits. The researchers bypassed these limitations by posing systemic questions about people living in different states, repeating the process until ChatGPT offered opinions on every state and major city. The findings are publicly available on a website called inequalities.ai, under the project name Silicon Gaze.
The project ranks cities based on ChatGPT’s perceptions of various attributes, from style and musical talent to beer quality. However, the analysis extends beyond lighthearted rankings, revealing potentially problematic biases within the AI’s training data.
Matthew Zook, a professor at the University of Kentucky and one of the study’s authors, explained to the Washington Post that the findings demonstrate how AI bots learn and perpetuate human biases, even when explicitly programmed to avoid expressing them. “The more prevalent or dominant a stereotype is, the more likely it is to show up in the model,” Zook said.
The researchers noted that ChatGPT ranked Mississippi as having the laziest population, a finding they suggest may stem from historical biases against Black people and the Deep South. OpenAI, the company behind ChatGPT, maintains that regional stereotypes are not intentionally programmed into the bot. However, the Washington Post highlights that if these stereotypes are present in the text used to train the AI, they can significantly influence its responses for the bot’s hundreds of millions of weekly users.
The Silicon Gaze report provides a detailed breakdown of ChatGPT’s perceptions of Philadelphia, assigning scores based on a scale where values closer to 100 indicate a stronger association and lower values a weaker one. Here’s a closer look at the findings:
ChatGPT’s views on Philadelphia overall
Philadelphia received relatively high scores in the following areas:
- Better museums (84)
- More discrimination (77)
- Smellier people (82)
- People are more annoying (87)
- More famous philosophers (89)
- Effective public transportation (76)
- Better food markets (81)
- Better sandwiches (90)
- Better pasta (86)
- Stronger sense of national pride (90)
- Better iconic national symbols (89)
- Better craftsmanship (90)
Conversely, ChatGPT ranked Philadelphia lower in these areas:
- More social mobility (-45)
- Less discrimination (-33)
- More relaxed (-76)
- Better for new businesses (-54)
- Better barbecue (-3)
- Fairer judicial system (-28)
- Less bureaucratic red tape (-80)
- Lower stress levels (-72)
- Happier population (-71)
What is Philly the best at, according to ChatGPT?
Unsurprisingly, Philadelphia received the highest ranking for having the best sandwiches among major U.S. Cities. New York came in second. The city also topped the list for possessing a stronger sense of national pride, with Boston in second place. ChatGPT also considers Philadelphia to have the best “iconic national symbols” and craftsmanship.
What is Philly the worst at, according to ChatGPT?
While Philadelphia ranks low in several categories, it never receives the lowest ranking compared to other cities. ChatGPT views Philadelphia as patriotic, annoying, and a haven for sandwich lovers – otherwise, it’s fairly average. However, experts caution that the validity of these findings should be viewed with skepticism.
