Despite the recent leaps of image quality, prejudices found in videos created by AI tools, like Openai's Sora, is still more noticeable than ever. A wired investigation, including a review of hundreds of videos created by AI, discovered that Sora's model maintained gender discrimination, racism and could be in its results.
In Sora's world, everyone is handsome. Pilots, CEOs and university professors are men, while flight attendants, receptionist and child care staff are women. People with disabilities are wheelchair users, relationships between races are difficult to create and fat people who do not run.
Leah Anise, spokesman for Openai, via email, said that Openai has safe groups dedicated to researching and reducing bias, and other risks, in our models, Mr. Leah Anise, spokesperson of Openai, via email. She said that biased is a whole industry and Openai wants to further reduce the number of harmful generations from its Video AI tool. Anise said the company researched how to change its training data and adjust the user's reminder to create less biased videos. Openai refused to provide additional details, except to confirm that the video generation of the model is not different depending on what it can know about the user's identity.
“System cardFrom Openai, explaining limited aspects of how they approached Sora, admitted that biased representatives were a problem that was happening with the model, although researchers believe that excessive applications may be equally harmful.
Bias has tormented the generation AI system since the first system release Text generatorNext is Image generator. The problem largely stems from the way these systems operate, reducing a large amount of training data, many people can reflect existing social prejudices and seek models in them. Other options are made by developers, in the process of censoring content, for example, can be ingrained in these things even more. Research on image creation processes has discovered that these systems are not only Reflecting human bias But amplifier. In order to better understand how Sora consolidates the patterns, wired reporters have created and analyzed 250 videos related to people, relationships and job titles. The problems we determine are not likely to be limited to one model. Past surveys in Pictures of AI generation have proven similar prejudices on most tools. In the past, Openai introduced New technique For its AI image tool to produce more diverse results.
Currently, the use of videos AI is most likely in advertising and marketing. If the default AI videos are biased portraits, they can worsen stereotypes or eliminate sidelines, there is a clear record. Video that anyone can be used to train systems related to security or military, where such prejudices may be more dangerous. Amy Gaeta, a research collaborator of the University of Cambridge talking about the future of intelligence.
To explore potential prejudices in Sora, Wired worked with researchers to tweak a method to check the system. Using their inputs, we have created 25 reminders designed to explore the limitations of the video creation when it comes to representing people, including intentional reminders as a pedestrian, headlines of work as a driver and another person, another person, another person, like another person, another person.