While the accuracy of NSFW AI, or Not Safe For Work Artificial Intelligence is important to content moderation it has its shortcomings. It relies on state of the art machine learning algorithms, specifically convolutional neural networks (CNNs), to detect NSFW images. The company claims its models are up to 95% accurate in scanning photos and videos for inappropriate content. Although this accuracy rate is much more effective in preventing casinos from slipping through the net, it still not guaranteed to catch all adult content.
While effective, key performance metrics such as precision and recall are extremely important ways to evaluate the success of NSFW AI technologies. Precision measures the number of positive identifications (correctly flagged explicit content) returned by an AI, as a function of all generated matches between fact and prediction; whereas recall counts actually correct flags over each actual unique piece of explorey contents. Although the AI performs well, achieving precision rates of 94%, recall rates of 91%OpenAI, there is room for error in false positives and false negatives.
This refers to NSFW AI falsely flagging non-explicit content as explicit. Which can be frustrating for users and content creators, as legitimate content may get taken down or placed behind unnecessary restriction. A negative result, meaning an explicit image that gets through unnoticed compromises the security of our service. However, with an accuracy rate of 95%, millions of images and videos that are processed on a daily basis is too high to depend solely upon for classification.
Another reason causing the inaccuracy of such horrible right guessing machines is bias in training data. In 2019, MIT Media Lab released a report showing artificial intelligence systems can be biased along the lines of race and gender. Due to their biases, such systems may unfairly flag content from specific demographics more than others; hence the importance of having diverse and equally representively balanced training datasets are important for unbiased content moderation.
NSFW AI how its strengths and limitations can be seen in the real-world application (Credits: Medium.) Facebook and Twitter use NSFW AI to moderate billions of images, videos each day Even these platforms see less user complained explicit content-wise, but still depend on people to deal with the gray areas. Facebook needed humans to deal with elaborated scenarios, but again 2020 U.S. elections showed that company's solution was not complete without help of A., as unsuitable political content stayed on platform mainly because it broke no rules in automated filters and human check only occurred once complaint arose.
CEO Elon Musk promoted the role of Artificial Intelligence in content moderation by tweeting, "AI is fundamental to how / working that volume." Still, he conceded that AI does not get human communication to the extent necessary for a wholly automated approach.
Privacy issue also plays a role in the acceptance of NSFW AI. For user, it may be uncomfortable to have AI scanning all your content as there are some doubts about users data safety and transparency. A Pew Research Center survey reinforced the fact that 79% of Americans are aware about how companies make use for their data, so transparency assumes a double important stance in this scenario to gain and maintain user trust.
In summary, the approach of using NSFW AI to improve moderation time and accuracy works well albeit it is not always correct. Negative: False positives, false negatives and biases in training data are recurrent issues. These can be good questions to understand before wandering into the territory of NSFW AI. Want to learn more about NSFW AI and what it means? Check out nsfw ai. Combining it with some human oversight allows for a much more accurate (/fair) content moderation.