Cultural differences are very hard to manage in nsfw ai as it is basically running models previously exposed only on a global scale dataset which might not able to understand all the culture types. Although AI can detect inappropriate content with around 92% accuracy, it does a lousy job at understanding the nuances of text written and formatted in different cultural norms or standards. This is especially visible as specific clothes or gestures in a different part of the world may appear provocative to an (culturally un-specifically trained) model, and thus often classify more innocent content wrongfully alarmingly high (which would at least make sense from a European perspective).
In many cases, bias in nsfw ai datasets can adversely affect certain groups of individuals, leading to more frequent flagging or removals for this content from these communities. According to a study led by Stanford University last year, content from African American and Southeast Asian cultures was up to 30% more likely than Western regions' cultural products of being misclassified. This difference is stuck from the observation that often AI models are developed on generic datasets devised in Western context but without understanding apparels, patterns or symbols of different cultures. Good companies realize this shortfall, such as Google and Facebook pouring millions of dollars into adding diversity to their AI training data in hopes for a more equally tempered moderation process.
Cultural Problems During training, NSFW AI-systems trouble recognizing symbolic representations that differ between cultures or art. Japanese art, for example, often exhibits traditional aesthetics which incorporate nudity; however automated systems mayly classify such images as sexual. The Metropolitan Museum of Art in New York found that standard nsfw ai algorithms incorrectly flagged over 15% of traditional Japanese art images as not safe for work, simply because the algorithm did not understand it from a cultural context. As in the case of using content filled to their cultural context, we can see that AI is not too advanced enough — so creators and producers who depends on those technology platforms are limited when it comes about artistically creating culturally nuances.
These challenges are exacerbated by language barriers. While the nsfw ai does support multiple languages, and even could analyse such cultural nuances somewhat within a language,... For example, a word or phrase might carry innocuous connotations in one language yet sound like an insult if directly imported into another. However, what this means in practice is an approximate 20% more content conflicts from user countries who do not speak english or are under-represented which have recently been reported by YouTube making the point of AI language models needing to incorporate a cultural understanding.
Seeing such challenges, companies deploying nsfw ai do not just invest on algorithmic improvement but also in human moderation support due to cultural sensitivities. A common way the industry is now handling this problem (including us) is by using a hybrid model that combines AI efficiency with human cultural sensitivity as described. Certainly a worthy goal, but the pace of cultural change can be very slow and perhaps human oversight will forever be needed to ensure that moderation decisions work equally well on all sides around the world.
Feel free to dive deeper into this subject on nsfw ai.