The first of these 2 courses, is one of the biggest issues NSFW ML faces: to detect pornographic content as accurate as possible and not over-flag or miss the critical material. Whereas NSFW AI models can be up to 95% confident they have identified the content, an error is still possibleModern technology. When innocent content is mislabelled as adult, true false positives, users and creators are likely to be pissed off. Despite being less critical, false negatives allow undesirable content to slip through the moderation net, and can translate into legalse issues around compliance in platforms such as OnlyFans or Reddit where millions of posts are created every day.
The complexity of training data is another challenge. The AI must be very well trained on NSFW data of any kind so that it can discern what is ok and what is not. But these datasets needs constant updating for changing in content trends. As example, AI systems have trouble in distinguishing artistic nudity with explicit imagery. In a widely-publicized example in 2018, Facebook AI called classical art images (including works from Rubens) offensive because they contained nudity. This is an age when content moderation and cultural nuances need some delicate dance balancing.
Real-time content processing comes with its own set of restrictions as well. The AI that enforces this kind of a NSFW needs to scan thousands of pieces of content per second on popular platforms, which makes the problem interesting. As per a 2020 report by Statista, the average social media user uploads more than 1.3 MB of data every second, which presents a challenge for real-time moderation, without interfering with the user experience. Worked examples of NSFW AI are computationally intensive and can require GPUs or TPUs to keep them running smoothly. For example, they are the most convenient (and cost-effective) testing target for expressions of massive-scale AI content moderation as seen on AI systems powered by nVidia V100 GPUs that run 125 teraflops.
There is another element which complicates the use of NSFW AI and is human intervention. It can deal with lots of data at once (to a point), but still run edge cases by your eyes. As Elon Musk quipped: "AI is only as good as its training data," the downside to simply using AI is that it cannot be too accurate or publishers will fall foul of censorship rules, but also when trained up on largely benign and relevant but non-pornographic material, the model may not have enough exposure to edgy content. That makes the industry more expensive; companies have to pay for human moderators and also invest in retraining their AI.
Another point to tackle is the bias in AI algorithms. It is also, of course, not a surprise that NSFW AI systems trained by humans have the biases that can often be present in their training data. One example is AI, which — often trained on a diet of Western culture — might have difficulty understanding content from regions where different levels of explicitness are acceptable. In turn, such inconsistency leads to unfair censorship and content scrutiny between different regions.
Finally, the overhead cost to develop and maintain NSFW AI is large as well. Some companies invest millions of dollars in creating & educating such systems. Based on a report by Accenture, it estimates that in 2021 the price for AI training infrastructure can cost anywhere from one hundred thousand to over one million dollars, depending on model complexity and size. Such a high level of investment is needed to guarantee accuracy, speed and adaptability — but it also places a burden on the small play which might have to use less sophisticated systems.
To conclude, nsfw ai faces a multitude of problems from accuracy and real-time processing to training and maintaining these systems being expensive and cumbersome. Resolving these problems is vital to enhance the effectiveness and equity of AI content moderation.