It's really difficult for NSFW AI solutions to maintain accountability in the long run. Algorithmic bias is a key area. Research has found that these systems have 30% greater error rates on content created by minority groups as well, relative to more popular content. This is a problem that occurs because of the unbalanced nature in our training data. For instance, an AI Now Institute paper even finds that Facebook's and Twitter censorship algorithms can disproportionately affect communities on the margins.
The second challenge is decision-making transparency. Numerous AI systems for NSFW classification are currently being developed, but most of these models work as 'black boxes,' meaning the way their internal mechanisms classify content can be hard to grasp. If it is not clear why a specific decision happened then, how can users or regulators hold AI systems accountable for the decisions they are making? This is as prominent AI researcher Dr. Timnit Gebru puts it, saying: "It's impossible to hold accountable the systems that aren't transparent".
Additionally, NSFW AI systems may be dealing with sensitive data and thus concerns of user privacy is also a big issue. To invest in network management and security-based systems can cost more than $5 million per year for cybersecurity. Even with GDPR and other privacy regulations dictating exactly how user data should be handled — some still get breached. The data breach at a large social media platform in 2020, for example, resulted in millions of user records being published online: This is demonstrating the consequences.
At the other extreme, responsibility also depends on how well a system can adapt as societal standards and norms change. Such AI will significantly degrade as time goes on, and very quickly, it could become a tool that is rejected until the developers update its understanding of what works are While NSFW AI should need to be updated every once in awhile besides for criminal stuff — they may have different incentives their models other factors outstrip those costs long term. This results in content moderation errors of outdated, or irrelevant. The article includes a quote from the Electronic Frontier Foundation: “Responsible AI must be dynamic and responsive to changing societal expectations.”
While there are challenges, strides have been made in recent years to make NSFW AI systems more accountable. This includes efforts like making training data sets more inclusive, and improving algorithmic transparency. More information and responsible practices in nsfw ai responsibilities on rb.