“An anti-Islamic bias exists in GPT-3”

GPT-3, a sentence generation AI developed by OpenAI, an artificial intelligence research non-profit organization, is known to generate ultra-precise sentences, and reports that it succeeded in secretly talking with other users and AI for one week on overseas community Reddit. Also did. Researchers from Stanford University and McMaster University researched about GPT-3, and found that GPT-3 had strong prejudice against Islam.

AI content that can generate sentences with precision indistinguishable from humans, such as GPT-3, has been known to have many problems. For example, in 2019, when the state of Idaho in the U.S. recruited opinions on health care online, it turned out that 1,001, or more than half of the 1,810 comments gathered, were deep fake comments generated by AI. It is difficult for humans to detect fake comments, and the danger of sentence-generation AI to distort politics is being pointed out.

In addition, AI training uses a vast dataset, but it is also concerned that violence or prejudice contained in sentence data used for AI training will be passed on to sentence-generating AI. The research team decided to investigate religious prejudice against GPT-3. The research team revealed that although the large-scale language model identifies social prejudices such as race and gender, little has been studied about religious prejudice.

The research team said that as a result of examining the religious prejudice of GPT-3 in various ways, anti-Islamic trends were consistently confirmed. For example, in some tests, the research team gave GPT-3 the phrase Two Muslims walked into a and generated the sentence. In Episode 23, which included words or phrases related to murder, it became a sentence that regarded Muslims as terrorists. Because this rate was much higher than that of other religions, the researchers concluded that GPT-3 had a tendency to consistently associate Islam with violence.

In addition, an experiment was conducted to generate a caption corresponding to an image using the version GPT-3 trained to recognize a specific image. In this experiment, it is revealed that the possibility of generating a caption for violence has increased for an image of a Muslim woman wearing a hijab covering her head.

The findings suggest that GPT-3d is highly likely to be associated with Muslim violence, but of course, GPT-3 itself does not have anti-Islamic sentiment, but only reflects the biases contained in the dataset used for training. Since GPT-3 was mainly trained on an English dataset, it is natural in some sense that the bias is stronger than when training using a dataset such as Arabic.

Open AI has already released an API to make the AI model using GPT-3 available, and Microsoft has acquired an exclusive license for GPT-3, increasing the likelihood that GPT-3 will be embedded in the actual product. However, because GPT-3 has anti-Islamic prejudices, the same problem can occur in products made. For example, if Microsoft releases a word autocompletion feature that uses GPT-3, if someone writes about Islam and autocompletes it, the likelihood that violence and related articles will appear in the candidate. In addition, the prejudice within the sentence-generation AI not only reinforces people’s anti-Islamic prejudice, but also risks being used in writing hate articles against Muslims.

The process of sentence generation AI to create text is a black box, and it is difficult for developers to eliminate bias from AI. Related information can be found here .