Dan Hendrycks, affiliated with xAI, has developed a method to measure and influence AI preferences, including political views. His approach could help align AI models with elected electorates' sentiments, suggesting a default bias toward election winners. The research team analyzed various AI models and established that their preferences are consistent and intensify as model complexity grows. Past biases seen in AI tools like ChatGPT prompt concerns regarding an overrepresentation of left-leaning ideologies, stressing the need for balanced perspectives in AI outputs.
Hendrycks suggests that AI models should reflect the will of the electorate, arguing for a slight bias towards the winner of the popular vote in elections.
Using a new utility engineering approach, researchers analyzed AI preferences and found they are often consistent and become ingrained as the models grow larger.
The researcher warns against AI biases leaning too far left, noting that tools have been criticized for promoting views seen as 'woke'.
The findings could allow for more tailored approaches in future AI models, aligning them more closely with user preferences.
Collection
[
|
...
]