The context is the same but its an entirely new challenge for nsfw ai since content can have different meanings. Although present systems are accurate about 90% of the time when detecting explicit content, their accuracy drops to around 70% in differentiating between artistic expression and inappropriate material. This gap is most visible in user-generated content on social media platforms as it involves cultural, linguistic and contextual factors.
Yes, something like a's photo of a nude sculpture might be labeled nsfw in one context but not-an-art-case within another. Last year there was a significant event where an educator uploaded educational material sharing information about human anatomy that they taught and it censored due to its adult content, over 1 in four complaints related with the same issue involving educators and people from institutions. The situation illustrates exactly the sort of problem nsfw ai can have with differentiating between outright porn and educational or artistic content.
The machine learning models underlying nsfw ai are trained on massive datasets, sometimes numbering in the millions of labeled images; however even with infinite training no amount of AI is likely to completely understand context as well as a human. Performance can be increased by about 15% by using natural language processing (NLP) to understand the surrounding text although this still gives failure for complex and ambiguous cases.
Pioneers in AI, such as Andrew Ng have gone so far to point out scalability of context that the present day´s system is not capable off. Ng provided another example on his point: AI is really good at recognizing patters, not perfectly to relate all the nuanced nuances of human behaviour. That was during a conference in 2022 where he reminded again that context remains an open question for developers working with making sense out of the world via artificial intelligence.
In other words nsfw ai are both capable and not of understanding the context, being that it only understands clearly defined submissions but is completely lacking in nuanced cases. For much of the same reasons, nsfw ai improvement efforts in these areas focus on greater algorithmic leverage and incorporating more sophisticated contextual recognition.