NSFW AI is slowly getting better at decoding slang although the tech still has limitations. AI models trained on data up to Oct 2023 were able to accurately decode the slang used in online communications 78% of the time, a jump from only 55% in 2020, according to a report by researchers at the University of California. The strides can be attributed to the access of large-scale data and enhanced ML methods, which permits models to learn popup words, shorthands, and cultural practices used in a majority of online exchanges.
In the example of a gaming community, it is common to use terms such as smh (shake my head) or ftw (for the win). NSFW AI systems trained to scan sites like Twitch and Discord for live chat content know these slang terms. Last year, Twitch said its AI moderation tool automatically flagged and filtered 1.5 million Jersey-based offensive messages with slang detected by the ToastNota robot. It also shows how AI is increasingly good at interpreting not purely formal language but developing trends in digital slang.
While a lot of work has been done, it is not without its flaws. Another reason is that several slang terms are too contextual and it gets difficult for AI to understand when the context is not provided. One example is the word "lit," which can refer to a party or it can be used when something is "illuminated." This means that NSFW AI systems will likely fail to understand slang used in ambiguous ways as part of an emerging trend or niche culture. This problem becomes even more pressing in fast-paced online communities where slang can shift in change weeks to months.
Facebook reported that 92% of hate speech using slang on Facebook was flagged by its AI systems in 2023, but the technology struggled when people used creative or coded phrasing. Companies such as Facebook or Twitter are in continuous improvement mode with the algorithms, integrating real-time language processing and large databases that create a more accurate snapshot of evolving slang.
AI is also challenged by regional variations of slang. Common words in one part of the country will get trained to work differently in another and AI and machine learning needs to be trained accordingly. An example of this is the use of the word "boot"; in British slang, this refers to the trunk of a car; conversely in American slang, it generally refers to footwear. To accurately moderate content by region, AI has to learn these subtleties and continuously update its learning algorithms.
To top things off, while nsfw ai is improving its grasp on slang with time, there are still challenges eg not context-sensitive or unable to keep up with trends due to geographical differences. AI technology is getting better and the databases are getting bigger, this means that moderation of slang based content will become more accurate.