So what are the limitations of NSFW AI? Yes, it absolutely does. The major constraint that surrounds NSFW AI models, while considering their analysis is accuracy. In 2023 the marked types of content were discovered to be around 15% false positives or negatives, significantly compromising user experience and content-moderation efficiency. But the challenge is cost-effective deployment, including finding better ways to train these models (Facebook and Google spend tens of millions a year just making sure their prone-to-failure neural networks can even filter potentially rude stuff)
Ata technical level, NSFW AI is difficult because of a number of the failure modes it can have_image resolution_pipeline context_complex cultural nuances … A few imaging, for instance intensely pixelated versions or those with filters applied utilized in imagery might go undetected and cause innocent content to be falsely flagged. Although the field has progressed, these models still fall short of possessing the sophistication necessary to achieve a more nuanced form of contextual sophistication — limiting their utility in real-world applications.
Companies including Tumblr have previously stolen the headlines for coming a cropper with over-zealous AI like this. When they implemented an automatic NSFW filter in 2018, platform activity plummeted by 33%, as countless well-posted posts flew under the radar to be flagged, causing user uproar. Not only to it erase user engagement type data (read: ad revenue — otherwise known as cash) from the platform, but it also emphasizes the grave failures of these systems.
AI ethics luminaries, like Elon Musk, have sounded the alarms that AI is not useful at all, in many sectors of the economy… and NSFW AI makers think they know better. He's noted that "AI does not understand human intention and context, even with highly advanced algorithms." This may indicate that using AI alone to moderate sensitive content is unsustainable without very heavy human intervention.
At the application level, developers often mention trade-offs between precision and recall. Often, increasing one will decrease the other — ie. improving the AI’s function to detect NSFW content better but then in turn produces more false positives. This has put a strain on the budget since money spent in buying highly sophisticated multi-layered AI solutions is no longer an option for smaller companies.
These two examples drive home the fact that nsfw ai, as widely consumed as it is, has serious limitations. For such companies it is days of unsupervised learning without the need to add every new data point to maintain a good model and hence good performance, as soon they role out a change or start tuning something based on larger scale rollouts user feedback comes in and has its own tastes. You can read more about this at nsfw ai.