Ethical Strategies To Eliminate Implicit Bias In Artificial Intelligence ——Experimental Analysis Based On Female Attractiveness Test

Implicit bias is everywhere.It affects all aspects of our lives without being conscious.It usually manifests as a stereotype of a certain kind of thing.It is conceivable that artificial intelligence trained by humans undoubtedly repeats this prejudice.This article systematically expounds on the interpretation and causes of implicit prejudice,and studies how it affects moral judgment and evaluation.From the perspectives of consciousness and cybernetics,attribution theory,and deep self theories,it analyzes the moral responsibility of each theory for implicit attitude.On the basis of these theories,we analyze the experimental samples of the artificial intelligence recruitment system based on facial beauty prediction（FBP）used in the cover girl recruitment system.Artificial intelligence is widely used in recruitment systems,such as Amazon,Google,Microsoft,etc.This experiment system will also be widely used in the recruitment and selection of models and actors in advertising,movies,and other media.We put forward two hypotheses from the experiment.Hypothesis 1 is that the results of the Chinese and German annotations appraisal of the attractiveness of female images in the AsiaEurope data set are implicit biased.Hypothesis 2 is that artificial intelligence will copy human prejudices.Both hypotheses have been supported by our experiments.How to eliminate the implicit bias of artificial intelligence? We use a fair preprocessing method to achieve this goal.First,we use the Generative Adversarial Network（GAN）method to expand and diversify our facial beauty prediction system database to solve the problem of bias in training data.After repeated calculations of the weighted weights of the balanced samples,the best value of the weight ratio of the unbiased input preprocessing of the data is obtained,and an unbiased artificial intelligence network is explored in combination with the advanced convolutional neural network（CNN）algorithm.The accuracy of intelligent machine learning（the parameters Pearson correlation coefficient,average absolute error and root mean square error are examined here）are better than the published ranking leading artificial intelligence FBP systems,we provided success for other artificial intelligence to eliminate bias demonstration.Based on the experiment,this article analyzes that the biases of artificial intelligence may be due to the biases of the training data set,AI algorithms,and the implicit biases generated in the process of human-computer interaction.Furthermore,we proposed and explored the feasibility of Kant’s human-oriented Formula of Universal Law as an ethical principle for unbiased artificial intelligence design.In the last chapter of the article,this study proposed a variety of ways to eliminate artificial intelligence implicit biases by establishing moral algorithms,diversifying sample sets,improving women’s disadvantaged position in the field of artificial intelligence,and increasing ethical and moral training in the AI industry.