Behavioral and neural evidence for the underestimated attractiveness of faces synthesized using an artificial neural network
Behavioral and neural evidence for the underestimated attractiveness of faces synthesized using an artificial neural network
复制标题
DOI：
10.1101/2023.02.07.527403
复制
发表时间：
2023-02
期刊：
影响因子：
--
通讯作者：
S. Nishida
中科院分区：
文献类型：
--
作者：
S. Nishida
尽管人工智能（AI）最近取得了一些优势，但人类对人工智能的潜在厌恶尚未消除。如果这种厌恶降低了人类对人工智能合成的视觉信息的偏好，那么这种偏好应该仅仅因为人类相信信息是由人工智能合成的而独立于其外观而降低。为了验证这一假设，本研究设计了一个任务范式，在该范式中，naïve参与者在虚假的指令下，对使用人工神经网络合成的各种面孔的吸引力进行评级，其中一半是合成的，另一半是真实的。这个设计允许评估参与者的信念对他们的吸引力评分的影响，而不是面部外观的影响。此外，为了研究信念效应的神经基础，在该任务中使用功能磁共振成像收集了大脑对面孔的反应。研究发现，当参与者认为人脸是合成的时候，他们的评分就会下降。此外，信念改变了fMRI信号对右侧梭状回皮层面部吸引力的反应。这些行为和神经方面的发现支持了这样一种观点，即人类对视觉信息的偏好降低仅仅是因为相信这些信息是由人工智能合成的。
Despite recent advantages in artificial intelligence (AI), the potential human aversion to AI has not been dispelled yet. If such aversion degrades the human preference to AI-synthesized visual information, the preference should be reduced solely by the human belief that the information is synthesized by AI, independently of its appearance. To test this hypothesis, this study designed a task paradigm in which naïve participants rated the attractiveness of various faces synthesized using an artificial neural network, under the fake instruction that half of the faces were synthetic and the other half were real. This design allowed evaluating the effect of participants’ beliefs on their attractiveness ratings separately from the effect of facial appearance. In addition, to investigate the neural substrates of the belief effect, brain responses to faces were collected using fMRI during this task. It is found that participants’ ratings declined when the faces were believed to be synthetic. Furthermore, the belief changed the responsiveness of fMRI signals to facial attractiveness in the right fusiform cortex. These behavioral and neural findings support the notion that the human preference to visual information becomes lower solely due to the beliefs that the information is synthesized by AI.