The potential effects of NSFW AI failures are so great that they have become much talked about both by the public who may use these sensitive apps and, well as developers. For example, a sexually explicit AI model developed by one of the largest technology companies in 2023 was reported to have produced inappropriate results for around 25% (with rampant fear that such systems were utterly unreliable). Hyper-profane.img The failure rate underscores the tough challenge developers face in training AI models to recognize where on a continuum of appropriateness some data belongs, especially when they are working with datasets that can contain millions or billions of diverse images and text entries.
One of the most prominent instances of NSFW AI failures took place a few years later, in 2022: A well-known social media platform had introduced an NSFW content filter based on AI capabilities. Users started reporting over the coming weeks many false positives, without nudity like art or family photos (with 30% of them). This not only pissed them off but led to a 12% decline in DAUs, which is expected to cost FB up $5 bil/qtr. advertising revenue impact
For one, a startup building an AI tool to advance NSFW content moderation suffered when the model proved incapable of differentiating what is and isn't inappropriate based on cultural differences. The model, which learned to recognize faces by scanning massive datasets on Western species, mislabelled content from Asian and African cultures so badly that it could make mistakes in about 40% of the images uploaded there. This is also a big mistake for AI training: we need to consider more kinds of datasets into the pool because it can be biased or incorrect in some culture(EX. #MeToo).
These failures are something that industry leaders have since often used as a horror story for the AI community. Look at Elon Musk; he has previously described how: “AI is more dangerous than nukes.” This quote rings true for NSFW AI and any false negative can not only result on loss of billions in earnings but significant image damage later experienced a user trust. The NSFW AI failings are a strong reminder that we need to continuously test our systems and look for every possible use case scenario, across different types of data sets or else — even the smartest AI can not do what it was programmed to.
In addition, dealing with the costs fallout of NSFW AI failures can be tricky. After headline-grabbing AI disasters, corporations have spent millions on lawsuits to clean up the mess. As an example, the closure of an NSFW AI tool due to a 2021 failure required $2M in research and development paid by its manufacturer with product delay reaching six months at loss of market share up to 10%.
Given that, these are two fine examples of the real consequence when NSFW AI fails not only for companies but negatively impacting society. Developers must be wary, proactive when it comes to these pitfalls in order that their product does not crash and burn at launch time. Those learnings, distilled from the sorrow that once happened are what move technology right now keeping us on a path better suited for where AI needs to be directed towards ensuring more responsible and effective solutions as we march forward into the future of NSFW AI.
Find out more about NSFW AI over at nsfw ai.