Kidd, C., & Birhane, A. (2023, June 23).
Science, 380(6651), 1222-1223.
doi:10.1126/science. adi0248
Here is an excerpt:
Three core tenets of human psychology can help build a bridge of understanding about what is at stake when discussing regulation and policy options. These ideas in psychology can connect to machine learning but also those in political science, education, communication, and the other fields that are considering the impact of bias and misinformation on population-level beliefs.
People form stronger, longer-lasting beliefs when they receive information from agents that they judge to be confident and knowledgeable, starting in early childhood. For example, children learned better when they learned from an agent who asserted their knowledgeability in the domain as compared with one who did not (5). That very young children track agents’ knowledgeability and use it to inform their beliefs and exploratory behavior supports the theory that this ability reflects an evolved capacity central to our species’ knowledge development.
Although humans sometimes communicate false or biased information, the rate of human errors would be an inappropriate baseline for judging AI because of fundamental differences in the types of exchanges between generative AI and people versus people and people. For example, people regularly communicate uncertainty through phrases such as “I think,” response delays, corrections, and speech disfluencies. By contrast, generative models unilaterally generate confident, fluent responses with no uncertainty representations nor the ability to communicate their absence. This lack of uncertainty signals in generative models could cause greater distortion compared with human inputs.
Further, people assign agency and intentionality readily. In a classic study, people read intentionality into the movements of simple animated geometric shapes (6). Likewise, people commonly read intentionality— and humanlike intelligence or emergent sentience—into generative models even though these attributes are unsubstantiated (7). This readiness to perceive generative models as knowledgeable, intentional agents implies a readiness to adopt the information that they provide more rapidly and with greater certainty. This tendency may be further strengthened because models support multimodal interactions that allow users to ask models to perform actions like “see,” “draw,” and “speak” that are associated with intentional agents. The potential influence of models’ problematic outputs on human beliefs thus exceeds what is typically observed for the influence of other forms of algorithmic content suggestion such as search. These issues are exacerbated by financial and liability interests incentivizing companies to anthropomorphize generative models as intelligent, sentient, empathetic, or even childlike.
Here is a summary of solutions that can be used to address the problem of AI-induced belief distortion. These solutions include:
Transparency: AI models should be transparent about their biases and limitations. This will help people to understand the limitations of AI models and to be more critical of the information that they generate.
Education: People should be educated about the potential for AI models to distort beliefs. This will help people to be more aware of the risks of using AI models and to be more critical of the information that they generate.
Regulation: Governments could regulate the use of AI models to ensure that they are not used to spread misinformation or to reinforce existing biases.
