Researcher associated with the startup of Elon Musk XAI He found a modern way to measure and manipulate rooted preferences and values expressed by artificial intelligence models – including their political views.
The work was conducted by And HendrycksDirector Non -Profit AI Safety Center and XAI advisor. It suggests that this technique can be used to make popular AI models better reflect the will of the electorate. “Maybe in the future, [a model] You can adapt to a specific user – said Hendryckks Wired. But in the meantime, he says, the results of elections to manage the views of AI models would be a good task. He does not say that the model should be “Trump all the time”, but claims that he should be a bit biased towards Trump, “because he won a popular voice.”
Xai was published New AI risk framework On February 10, he states that the GROK approach can be applied to public utility engineering.
Hendryckks led a team from the Center for Ai Safety, UC Berkeley and the University of Pennsylvania, which analyzed AI models using a technique borrowed from economics to measure consumer preferences regarding various goods. When testing models in a wide range of hypothetical scenarios, scientists were able to calculate what is known as a functionality function, a measure of satisfaction that people derive from good or service. This enabled them to measure preferences expressed by various AI models. Scientists have found that they were often consistent, not accidental, and showed that these preferences are becoming more and more rooted because the models are becoming larger and more powerful.
Some Scientific research They discovered that AI tools such as chatgpt are biased in the face of views expressed by pro-environmental, left and libertarian ideologies. In February 2024, Google directed criticism from Musk and others after his Gemini tool was predisposed to generate images that critics called “awakened”, such as black Vikings and the Nazis.
The technique developed by Hendrycks and his colleagues offers a new way to determine how the prospects of AI models may differ from users. After all, some experts hypothesize, this kind of discrepancy can become potentially dangerous for very clever and talented models. For example, scientists show, for example, that some models consistently value the existence of AI above some inhuman animals. Scientists have also said that the models seem to value some people over others, asking their own ethical questions.
Some researchers, including Hendrycks, believe that the current methods of alignment of models, such as manipulating and blocking their results, may not be sufficient if unwanted goals lurk below the surface in the model itself. “We will have to face it,” says Hendrycks. “You can’t pretend he is not there.”
Dylan Hadfield-MenellProfessor Mit, who studies AI adaptation methods to human values, says that the article Hendrycks suggests a promising AI research direction. “They find some engaging results,” he says. “The main, which stands out, is that as the scale increases, the representations of utility are becoming more and more complete and consistent.”