How AI can distort human beliefs
Celeste Kidd and co-author Abeba Birhane have a new article in Science on how models can convey biases and false information to users!
“When company officials talk about the potential harms [of AI models], they overwhelmingly give examples of nefarious forces using this technology for bad,” Kidd said. “In this piece, we’re talking about the harms that will happen from the intended use of these systems. It does not require a nefarious force to use the system in order to generate disinformation that’s intended to deceive people.”
From the article:
Individual humans form their beliefs by sampling a small subset of the available data in the world. Once those beliefs are formed with high certainty, they can become stubborn to revise. Fabrication and bias in generative artificial intelligence (AI) models are established phenomena that can occur as part of regular system use, in the absence of any malevolent forces seeking to push bias or disinformation. However, transmission of false information and bias from these models to people has been prominently absent from the discourse. Overhyped, unrealistic, and exaggerated capabilities permeate how generative AI models are presented, which contributes to the popular misconception that these models exceed human-level reasoning and exacerbates the risk of transmission of false information and negative stereotypes to people.
You can also read a great interview from Berkeley News with Kidd here.
From the article:
It’s been well established that these AI models produce biased material and also fabrications. That’s not new. What was disheartening for me and other people working in this area to hear was the developers’ responses: “Yes, there are problems,” they’d say. “There are distortions. There are fabrications and biases. But we need people to use the models in order to discover the problems, and then we’re going to correct them.”
The problem with that logic is that a lot of the biases and fabrications may not be detectable to people — especially if they’re using the system to help them make up their minds.
By that point, these systems have transmitted the information to the person. It may not be easy to correct.