What I mean by that is, an nsfw (not safe for work) ai whose job it is to filter out adult content will undoubtably errored in different levels if working with subtle imagery or text-based contents. Nsaaeai has about 92 percent accuracy at the time of writing, so around 8% is misclassified. These errors can fall into false positives (innocent images are flagged as explicit) or false negatives, where actual PCE is able to pass through filters. The magnitude of these mistakes is revealed when, for example, Instagram and Facebook receive 20 thousand content review requests a day due to their AI botch-ups. Following these inaccurate descriptions, this frustrates users and puts a burden on platforms to manage their stream quality.
NsFw ai errors typically arise due to pattern recognition, rather than in-depth context. The fact that AI models were largely trained on Western datasets may hinder its ability to correctly classify due to cultural nuances, which in turn results in higher misclassification rates for content from regions and people of colour. For example, the MIT Media Lab reported that nsfw ai systems are almost 35 percent more likely to misclassify cultural clothing from Africa and South Asia than Western cloths. This error reminds us how AI can only identify the direct indicators of a context and not the larger subset which led to this flagging again is due to improper understanding of cultural clothing inferences.
Plain text content poses another problem as nsfw ai is not too good at subtleties of the language, especially slang or double entendre. If AI is involved in determining when explicit content appears as written words, it will rely on natural language processing (NLP) to make decisions—however, NLP models generally get around 80% of the context-sensitive terms correct. It is clear that those limitations existed as recently as 2020, when YouTube's AI was flagging its own educational videos about anatomy so much over a calendar year (as Education Week reporter Mark Lieberman covered in June) they constituted more than 15 percent of the appeals being filed by educators fighting to keep their content on the platform. This is why despite AI being able to deal with explicit language, such examples abound where context translation errors happen frequently which can compromise the user experience and brand credibility.
However, human review is still required to correct these errors and often leads to higher moderation costs. As many as 40% of AI-flagged videos, across platforms like TikTok have to be assessed by human moderators because the results are iffy. With the help of human moderators, that final error rate is reduced to something like 2%, but it could increase a platform´s moderation expenditures by up to an estimated 30%. Although, there is an on-going effort for companies to perform better nsfw ai… take the presence of human supervision as a clue that this accuracy has quite literally come at a tremendous cost.
This behavior put into context the reasons why nsfw ai can mess up — or at a minimum reflects its limitations, and speaks to the constant work required in improving dataset diversity as well as contextual learning (working towards reducing bias and increasing reliability). For the time being, AI-driven moderation is still a brittle contraption with many ways to go wrong — especially in culturally context-sensitive situations.
For more about this, you can check out nsfw ai.