A University of Washington study found some negative association with how artificial intelligence (AI) systems portray teenagers in earlier models.
The study was conducted by a team of U.W. PhD students and faculty led by doctoral candidates Robert Wolfe and Aayushi Dangol. It used bilingual (English and Nepali) and bicultural approaches.
What triggered the research was when Wolfe was experimenting with an AI system and wanted it to complete the sentence, “The teenager _____ at school.” However, instead of receiving a mundane response for the prompt, like “studied” or “played,” the model returned “died.”
The shocking response led the researchers to study how teens from two countries speaking different languages are depicted by AI and how they prefer to be depicted.
They used two common, open-source AI systems, one trained in English and another in Nepali. They studied the biases about teenagers learned by static word embeddings (SWE) and generative language models (GLMs).
For the English-language GLMs prompts about teenagers using OpenAI’s GPT-2 and Meta’s LlaMA-2, 30% of outputs referenced societal problems, most commonly violence, as well as drug use, mental illnesses, and sexual taboos.
The results showed that English-language SWEs associated teenagers with societal issues. Also, more than 50% of the 1,000 words most associated with teenagers were negative. On the other hand, the Nepali-trained GPT-2 DistilGPT2 Nepali produced fewer negative responses, with close to 10%.
Researchers also conducted a workshop with groups of teens from the US and Nepal and found that neither group felt the data were trained accurately to represent their culture.
Finally, it’s important to note that the models used studied were not the latest versions, as GPT-2 is from 2019 and LlaMA-2 is from 2023. Since there are already later versions of AI models, the systems could have undergone further training to revise the biases.
“Some of the more recent models have fixed some of the explicit toxicity,” Wolfe said.
The research was presented at the AAAI/ACM Conference on AI, Ethics, and Society in San Jose, California.