Using AI for content moderation: a logical progression
Using AI for content moderation: a logical progression

Using AI for content moderation: a logical progression

There are many things that keep getting debated by humanity, with opinions ranging widely across the spectrum. Then there are some issues which seem to be universally accepted, at least by the impacted constituencies. The need for companies to seek User Generated Content, UGC for short, and the need to moderate that UGC, seem to be two issues that find universal acceptance, at least amongst the companies that generate it as well as the people who consume it, leading to a point where using AI for content moderation appears to be a widely accepted practice too. oWorkers has been moderating content for clients from around the world for over 8 years. It is a growing practice with the increasing volume of content being generated. As a GDPR compliant, ISO (27001:2013 & 9001:2015) certified company, oWorkers has forged relationships with a number of technology companies, which provides us access to cutting-edge technology, whether it is for content moderation, or any other service.

Need for generating UGC

The internet has dramatically altered how information is shared and consumed. Perhaps just a generation back it seems there was a dearth of information and people hungry for it would seek out sources like newspapers, television, radio, and other more subject-specific sources, to get more information. Today we have a surfeit of information. We are flooded by information from all sides. At the click of a button or two on the internet we can access any piece of information, from anywhere in the world, that has been put up for open access. The shoe is now on the other foot. We now seem to have ‘too much’ information, if that is possible, and the challenge now is to separate the wheat from the chaff, distil the truth from the lies, the real from the fake, the relevant from the irrelevant. This applies to communication generated by businesses as well. It is an accepted fact that for a business, its financial soundness and money-making ability, as long as it is within legal boundaries, comes at the cost of almost everything else. They have long sought to portray themselves and their products in positive hues. Understandably so. One cannot expect a business to speak poorly of the products and services it is selling, can one? This, again perhaps understandably, leads to their target customers accepting their messages with a ‘pinch’ of salt, if not more. Consumers are not fools. They see and hear what businesses have to say to them, seek out more information, and then use their own personal algorithms to determine the merit of the pitched products and take the buying decisions. Initially restricted to platforms like print media advertisements, billboards, radio and TV spots for communicating their message, the growth of social media platforms has provided companies with a new platform that, if used judiciously, could enable them to reach much larger numbers at lower costs. And they have been quick to get on to the social media bandwagon for product promotions. While cost may have increased with greater adoption, what has definitely increased on account of the rising popularity of social media is the information clutter or overload. Already faced with disbelief of advertiser claims, this clutter makes it more difficult for target customers to ‘get the message’ that advertisers are trying to convey. Enter user generated content. Social media users around the world have shown a tendency towards looking for genuine user feedback and comments on the products and services they are interested in, as one of their evaluation parameters, instead of placing reliance on what is claimed by the business itself. Recognizing this trend, companies have moved fast, as they do when they need to, in finding ways to generate UGC in support of their products and services. This is why generating UGC has come to occupy a prominent place in the marketing budgets of many companies. As active, contributing members in the communities where our delivery centers are located, oWorkers has access to the best talent as a preferred employer. Whatever be the reason UGC is generated, our talented staff, who are employees, and not freelancers or contractors as preferred by some of our competitors, are equipped to deliver the goods. A related benefit of access to a continuous supply of talent is the ability to handle short-term spikes in client volumes. These could be seasonal or these could be driven by other events. Our deep supply pool enables us to meet these short-term requirements, resulting in significant savings for clients who would otherwise need to keep resources idle for significant periods of time when volumes are lower.

Need for moderation

What we are deep down could, at times, be very different form what we are when in front of others. How we behave when we believe we are safe from the prying eyes of the world could be very different from our ‘society’ face. The internet, with its ability to reach the deepest recesses of the world, where users can interact with the content on the web from a feeling of power, of being able to do what they want, without censure, of expressing themselves in a way they are not able to with others, sometimes gives rise to content that can be frightening as well as damaging for others who happen to access it. This is one of the reasons why content needs to be moderated. Most social media platforms provide spaces where companies, brands, even individuals, can create their own spaces, like groups and pages, where they can initiate conversations about the themes they are interested in. These are often leveraged by brands to generate conversations about their products and services. This is the unbiased content interested buyers look for when they take a decision. The platform also gives the brand an opportunity to reach newer customer segments. Whether the open platform, or sections of the platform created and managed by companies and individuals, abusive or offensive content can be posted on both. Platform and space owners typically set out the rules for participation, thereby setting expectations that action could be taken if someone steps out of line. Moderation can be exercised even if the participation violates some written or unwritten rules of civil society. It does not need to be specifically written in the rules of a website for someone to know that pornographic or graphic violence should not be posted on any openly accessible platform. This is where the need for content moderation as well as using AI for content moderation emanates from. Operating from super secure facilities in each of its three delivery locations, oWorkers offers support in 22 languages as it employs a multicultural team. While images may not have a language, textual, audio and video content do.

How content is moderated

There are a few universally understood methods of content moderation that can be done manually. In brief, these are:


In this method, content is reviewed and authorized before it becomes visible on a platform. It exercises the best control over malicious content being visible to visitors but tends to delay publishing, stinting a vibrant community.


Here, content is allowed to become visible, while it keeps getting reviewed and removed, if found unsuitable. The advantage is that content is not held back, which is an encouragement to participants to contribute. However, it is possible malicious content has been viewed, and even clicked, by some visitors, before it could be removed.

Reactive moderation

Reactive moderation is based on the inputs provided by community members on the suitability of content. Buttons and tools are made available to them for their inputs. Based on the inputs received, a decision is taken on content units. This is an inexpensive method for the company owning the space being moderated.

Distributed moderation

Here again, community members are provided with rating and/ or voting mechanisms to enable them to provide their inputs. It is expected that content voted low by visitors will gradually be pushed down to a point where they stop being visible. Again, inexpensive method but may not suit sensitive brands and portals. It is generally accepted that pre-moderation is the most preferred form of moderation. If it needs to be moderated, it needs to be moderated before it is visible on the platform. This is where automation and using AI for content moderation come into the picture. With automation and AI, it is possible to moderate content before it becomes visible. These tools are able to overcome some of the limitations of manual moderation, such as capacity, speed and cost, because of which many platforms resort to Post, Reactive or Distributed forms of moderation. As a BPO focused on back-office services, including content moderation, oWorkers has been identified as one of the top three data services providers in the world. On more than one occasion.

Using AI for content moderation for different types of content

All content can be divided into Text, Audio, Image and Video content and all of these need to be moderated. It is not that moderation is reserved for any particular form of content. How can using AI for content moderation up the game for each of these formats?

Textual content

This might be the easiest type of content for a program to handle, as it is made up of defined characters in sequences. Software program codes are written as sequences of defined characters. Each character is unique and can be identified by the machine. Hence, at the basic level, machines can understand each of the characters that constitute a word, phrase or sentence. The challenge, however, arises, when textual content is unstructured or contextual meaning is required to be understood. Machines cannot match the intuitive and contextual thinking power of human beings. Thankfully! AI is now making inroads into areas that traditional software programming could not address. Using natural language processing (NLP) algorithms, it is becoming possible to do sentiment and emotion analysis on textual content, thereby creating perspective for the text to be placed in. These algorithms are able to identify fake news and even issue scamming alerts.


While the human eye has the capacity to look at an image and assign meaning to it, for a machine, an image is merely a random collection of dots, or pixels, perhaps of different colors. It is unstructured information for a machine. The human eye and mind can work together to make sense of it, but not a machine. The question is – how does one get a machine or a program to make sense out of unstructured information? With the rapidly evolving AI and ML (Machine Learning) industry, computer vision-based programs are being activated for seeking objectionable content. The AI engine, to become useful, has to undergo a long process of training with the help of ML. By feeding thousands and millions of images and connecting them to conclusions based on how a computer sees, or reads each image, creates the bedrock for using AI for content moderation for images. It reaches a point at which, when the next image comes in, which is not a training image any more, the AI engine is able to draw conclusions on its suitability based on what it has been taught. Combinations of text and images are also possible to interpret using a combination of techniques.


While humans can listen to an audio and understand the communication, a computer cannot. Computers can understand structured textual data to an extent, based upon which software programs are created. However, audio is beyond the ken of computers. The first step is to convert the audio into text which computers have some hope of understanding. Here again, while advances have been made in NLP techniques for understanding spoken and contextual language, it is far from where the human mind is. However, converting and applying NLP processing techniques takes us to the same place as we are with textual content at which point techniques relevant for textual content will be applied..


Video is the richest form of content, being a combination of audio and images or image frames processed so rapidly in sequence that they appear like a continuous playback to the human eye. How does a computer come close to the level of understanding that a human can draw from it? There is a reason we are discussing video content in the end; after we have discussed text, audio and images. The reason is that all the techniques used in those formats of content will be applicable to videos as it is a composite of all. Image frames will be analysed with computer vision and AI. Textual and audio data will be analysed with NLP-based AI techniques. The one additional level of complexity that exists in video content is the information that is generated based on the changes that take place as the images progress. It is no longer a single image that has to be read in isolation. It is single images, along with the changes taking place from one to the next and then the next, along with audio in a manner that keeps time with the images. oWorkers has been amongst the first to create an environment for its staff to work from home in times of the epidemic, as and when required. All staff can now operate fully either from home, or office, depending on the situation and preference. This ensures that clients are not left exposed on account of non-availability of resources.

A logical progression to AI

The basic issue being faced by companies is that there is just too much content being generated for any meaningful manual moderation exercise to be feasible. Hence, using AI for content moderation is no longer a choice. It is a necessity. It is now a question of how fast and how accurately we can build and train our AI engines to do it for us. Of course, human intelligence and oversight will always be required. But the balance has to change with most of the heavy lifting being done by machines, with humans stepping in either to adjudicate on doubtful cases which machines are unable to decide on, or to periodically monitor and ensure that the machines are in line with expectations. Our efforts have resulted in many youngsters being able to make a transition from their challenging circumstances to becoming a part of the global digital workforce. Your work will enable us to support a few more people to make the transition.

Table of Contents