The neutrality of NSFW AI is an issue that is challenging to dispel, simply because these systems are built upon data and models laden with the natural biases of their source data training sets conjured up by human designers. Ideally, AI models should be free of biases however due to societal and cultural bias, the end-result can affect content filtering & moderation.
Basically, NSFW AI works on top of huge datasets (images, text and videos) leveraged by machine learning techniques such deep learning to produce a predictive model. These models learn to read explicit content and are able to filter out any inappropriate material by a set of predefined rules. Artificial intelligence scans millions of posts on Facebook and Instagram every day, automatically tagging and deleting content as inappropriate (NSFW). Facebook claims its AI systems proactively detect and catalog over 99.5 percent of all adult nudity/sexual content removed from the platform.warning to cease harassmentingtakedowns due to complaints [ ] — With help community Standards Report; Adult Nudity & Sexual Activity
This can be easily done, but the real challenge lies in training. The AI is only objective in the dataset it learnsEDIoreferrer The NSFW AI would suffer the same biases if these are present in the training data; eg, cultural or societal influences. A set of content that was believed to be offensive in one culture can also find acceptability within another, and an AI with a dataset trained predominantly on data from only place may simply apply stricter norms across the board leading false positives or unnecessary censorship. OpenAI recently stressed the importance of fighting biases in AI: “neo training data can commit societal discrimination to group into neo AI models which might make unfair infirmities.
A recent controversy on Instagram saw its AI systems come under fire for unfairly flagging posts from plus-size women, including many Black users in 2020. Hundreds of men also accused the platform of bias, saying that due to AI more posts with likes curvy body forms you add can only make it not explicit while a post white slendery-liked women isn't.flagged. And, crucially, exposed the bias of NSFW AI — which was making determinations from content by its training data rather than against some neutral standard.
The issues disturb into imaginative and enlightening substance, as well. It is also further compounded by the difficulty AI systems have in distinguishing between artistic nudity and actual erotica. For example, in 2018 Facebook's AI erroneously detected a well-known painting: "The Origin of the World" by Gustave Courbet. The system was unable to interpret the artwork within its broader artistic context, thus revealing a high-level weakness of AI in delineating fine distinctions. Both are examples of the fact that NSFW AI can not often be truly objective simply because it does not have in human capacity to really understand culture, history or any artistry.
Over-censorship is another issue. Because some AI systems, such as those used in identifying content that is not safe for work (NSFW), have to be sure they have positively identified something before flagging it. According to an investigation by MIT Technology Review last year, NSFW AI can wrongly identify up to 15 percent of "safe" photos as explicit. In this way content moderation results hlss and as a consequence some useful information, such as health tips or art works getting blocked.
The creators of the NSFW AI are in a process to tackle these issues. Improvements in AI alleviation,a context-aware algorithms are being trained to allow the AI systems and get a better understanding of it while moderating. Insurer B, on the other hand might respond to more nuanced inputs and so become less prone to piling up false positives over time as it makes decisions based on understanding of those clues. The major challenge is that one would never be able to reach the highest level of objectivity because we are humans.
Although AI can process more data at higher speeds than the human brain, it is not always able to discern context in a way that seems natural. Like Elon Musk has famously stated, "AI is a fundamental risk to the existence of human civilization," concern which can be widespread in terms of AI's effectiveness when it comes to subjective judgment areas like NSFW content moderation.
But after all, NSFW AI is hard because training data is far from ideal and it still faces cultural biases, not to mention context. Strides are being made toward this mission but creating entirely bias-free AI systems is still a ways away. You can head over to nsfw ai if you want more in depth journey into how these systems are maturing, and where they fall shorts