With the volume of online content rising exponentially, NSFW AI has become a vital technology for filtering out inappropriate work. At that scale, Facebook and YouTube (and Google Drive, Gmail) can no longer operate as simple user submission platforms; they need to let AI systems take the first stab at filtering out undesirable content — up over 95% of all removal actions on Facebook now come from AIs. NSFW AI had the power of scale — it can process thousands images per second and detect in a fraction of seconds what human moderators would need minutes or more.
But accuracy is still a fundamental concern. NSFW AI is based on deep learning models such as convolutional neural networks (CNNs) and transformer architectures, which routinely achieve accuracy rates ranging from 85% to 95%, depending on the dataset and complexity of the model. Although these percentages are high, there is still a danger of errors in cases that require the understanding of more nuanced content by an AI since this will have difficulty gauging context and lead to over-censorship or under-filtering.
And because of the cost-efficiency, NSFW AI is easier to be implemented. Traditional content moderation often requires human reviewers and can cost over $100,000 per year for just 1 full-time moderator. By comparison, AI is capable of cutting these current costs by as much as 70%, depending on volume and automation. That cost structure makes NSFW AI feasible, at least economically, for massive platforms that need to moderate content at high volumes on a constrained budget.
Unlike academic research, real-world use cases showcase how NSFW AI is truly changing the game. They told me that as user activity surged for platforms like Twitter in the early days of COVID-19, scaling content moderation was crucial. Over this period, Twitter said it escalated its use of AI to moderate content by 50%, demonstrating the speed with which these systems can be scaled in case incoming volumes go through existing manual capabilities.
But we cannot disregard the ethical questions around using NSFW AI. AI has not matured nuanced enough to understand which nuance is rightful, critics argue that hence AI may lead the content moderation with biases and removal of some justified posts by mistake. One such incident occurred on Instagram in 2021 when this AI began to flag posts related to breast cancer awareness, leading public complaints and an example of how current model AIs are poor at inferring context.
Of course, regulatory pressures also influence the future of NSFW AI. Around the world, governments are requiring strict rules around how content is policed; witness the European Union providing stiff fines to services that do not police against dangerous speech in its upcoming Digital Services Act. Moreover, given the scale at which NSFW AI can operate and the fact that these rules tend to be set either by large global platforms operating under a massive variety of legal jurisdictions with their own requirements as you go down or each nation enforcing its chosen values one way up (it is less relevant what those values are than it is who make them — meaning we could imagine very strict anti male nipple laws enforced over vast areas!
The outlook for NSFW AI and content filtering is optimistic, but it will likely be a balance of human handiwork along with evolving generations (or months) of improvements in transparency as well as ethics. By outsourcing human judgment to the existing workforce and computational scale of AI, content moderation reaches a middle ground. The keyword nsfw ai truly embodies this, it is a representation of the continuously changing role AI will play in any content filtering strategy.