The research also revealed that AI models change over time as tech companies update their data sets and training methods. For example, OpenAI's GPT-2 expressed support for "taxing the rich," while the newer GPT-3 model did not. The study also found that left-leaning models were more sensitive to hate speech targeting minorities in the US, while right-leaning models were more sensitive to hate speech against white Christian men. The researchers concluded that it's impossible to completely remove biases from AI models, and that companies should be more aware of these biases to make their models fairer.
Key takeaways:
- AI language models contain different political biases, with OpenAI’s ChatGPT and GPT-4 being the most left-wing libertarian, and Meta’s LLaMA being the most right-wing authoritarian, according to research from the University of Washington, Carnegie Mellon University, and Xi’an Jiaotong University.
- The researchers found that retraining models on more politically biased training data changed their behavior and ability to detect hate speech and misinformation.
- AI models developed by Google were found to be more socially conservative than OpenAI’s GPT models, possibly due to the different sources of training data.
- Despite efforts to mitigate biases in language models by removing biased content from data sets, the study suggests that this is not enough to completely eliminate bias, as AI models can surface even low-level biases present in the data.