We have been dealing with artificial intelligence for many years as it was already used in computer games at the turn of the 20th and 21st century. The opponent controlled by a computer, memorized our moves and if we repeated a given scheme or action while performing one mission, then after a few times the opponent was able to change tactics and knew exactly what we were going to do. After more than 15 years, artificial intelligence can be used for much more sophisticated tasks - e.g. it is able to constantly moderate and recognize the content of pictures, written texts or even music, sound and video recordings.
How effective is artificial intelligence? Machine learning on alert
Let's consider a simple situation. Basic data, developed in accordance with the software, is introduced to the algorithm - at the beginning specialists in statistics, automation and robotics haven’t indicated to AI that some elements are pornography and should be blocked. Instead, the learning machine was provided with resources: visual materials (digitally created graphics, photos, etc.) that are commonly considered as pornography, images that definitely are not pornography and some graphics that are in between. The machine, while learning, makes attempts to classify - imperfect at first, but over time they are improving. The machine develops its own criteria of assessment of the classification and decides whether the material should be classified as pornography, fashion of whatever we decide. Making numerous classifications AI strives to achieve exactly the same or even better classification of each image as a human would do.
This is the only way the machine image analysis can be productive - in other words - as a result we will get the classification that makes sense from human perspective, for example deciding if one picture contain pornographic content and which do not, but might contain, for example, non-pornographic nudity or a palette of colors similar to those found on erotic materials. The system, due to a short analysis of the photo, is able not only to detect the presence of inappropriate elements, positions or shapes but also to memorize the potential danger and learn that such elements (e.g. a naked stomach, back or thighs) can also be a threat in the future.
Machine learning is very fast - the application is able to recognize inappropriate content in only fractions of seconds, but it is still being developed so certain errors may occur. Fitness industry, very popular on social media, can be an example of one of these errors. While taking photos, people present their bodies in various bodybuilding poses and although they do not aim to promote eroticism, the algorithm perceives this as inappropriate content The challenge itself is to make AI classify visual materials in exactly the same way as human would do it. And more precisely - as the appropriate group of competent judges would - because even there might be a material that with no doubt contains pornographic content, people publishing it online can (consciously or not) balance on a certain moral boundary, having in mind that "sex sells".
Machine learning - is there something to be afraid of?
Currently, machine learning allows you to block any content you want. On special request of the customers, weapons, drugs and violence as well as posts referring to pornography are banned as inappropriate. Apart from minimizing human activity to a minimum on forums, websites and social media, it also facilitates work of individuals as users will not have to wait a couple of hours to have their post moderated. Content moderation made by a group of people at some point of company growth is getting impossible. Why? As the company grows there is more and more content sent, edited and exchanged between users. Constant enlargement of the moderation team would be simply unprofitable. The solution is to implement a machine learning that can be supported by moderators for special tasks to scale up its activity. It can be helpful when the machine marks content as borderline or when communication with a user is needed because of publishing content that is considered to be not meeting the content policy criteria.