Keeping NSFW AI is a non-stop process of monitoring, updating and perfecting the algorithms (both to catch more/explicit content ). The first step of this process is to retrain AI models on new data with consistency. In 2023, this also allowed a big tech co to claim their NSFW AI improved its precision by using these same set of images as follows: Their (NSFW) model received a strong +15pts accuracy after having integrated the additional ~50k new porn data samples into her training. This is a good amount of increase and it implies an effect on the performance due to data refreshment itself.
A regular rotation of data scientists and engineers is focused on continuing to improve the algorithms necessary for keeping NSFW AI up-to-date. This team makes changes to the AI model parameters and thresholds in order address new types of explicit content. The price of this maintenance can also be high and generally costs at least $100,000 to over $500,000 a year for large-scale platforms. But this is a necessary investment in making sure the AI works well in production.
Speaking of shifts in threats, industry leaders suggest illustrative references to answer this through adversarial testing — where problematic inputs are specifically designed and used for the AI robustness tests. A prime example occurred in 2022, when Google's AI research division performed adversarial tests that found chinks in the armor of their NSFW AI and increased filtering accuracy by 20 percent after making adjustments. The stress test phase helps train the AI to deal with even more sophisticated methods of trying to get around content filters.
AI is just like a child, as mentioned by of the famous AI expert Andrew Ng “It needs continuous learning to adapt and grow”. It directly applies to the idea of NSFW AI hygiene. The model needs to be updated, fed with new data and tested continuously so that it can evolve over time better respond to different types of emerging threats.
Although in practice, tech companies tend to schedule maintenance cycles once a quarter where the AI model is retrained and then tested against new datasets. This could include billions of simulated interactions to ensure that the AI works well in all situtations. This process is evaluated by how quickly updates can be deployed and the corresponding gains seen in improved performance of the AI. For example, a major social media service provider slashed its deployment of updates from 48 hours to 12 — quadrupling their ability to respond quickly.
Keeping NSFW AI is not a one-time activity, but instead an ongoing effort with huge resource and skill requirements. It requires a mix of regular data updates, adversarial testing, and continuous checking to maintain the effectiveness of system. The advantages are straightforward: a greater degree of precision in terms of content identification, so more user-protected and always one step ahead those trying to game the AI. For a deeper dive into NSFW AI, head to nsfw ai.