Combating Pro-Eating Disorder Social Media Content Using AI

The solution is a Deep Learning model capable of detecting Pro-Eating Disorder content on social media to protect at-risk users from the harmful influence of such posts.

Across social platforms worldwide, the Pro-Eating Disorder movement has taken the world by storm. Using websites such as Twitter, Instagram, and Facebook, content glorifying unhealthy eating habits and unachievable physiques has infiltrated nearly every home secret and undetected. Inundated by the ever-growing number of Pro-Eating disorder posts, many users have succumbed to their influence, spreading such posts further and even possibly developing an eating disorder in the process. In fact, research has highlighted a direct link between an increase in social media usage to a spike in eating disorder diagnoses and hospitalizations over the last decade. 

It will come as no surprise that teens and young adults are the groups most hurt by the proliferation of Pro-Eating Disorder content on social media. In 2021, over four billion people engaged on social media, a plurality of which are the aforementioned teens and young adults. Moreover, over 84% of people between the age of 18 and 29 habitually visit social media sites and are, therefore, constantly bombarded by images and videos of people venerating not only unhealthy habits such as extreme dieting but also eating disorders themselves. Teenagers and those in their early twenties are the most susceptible of any group to the influence of such nefarious content and are the most likely to develop an eating disorder, which often goes untreated. After all, only 10% of those with eating disorders receive treatment of any kind. The danger of eating disorders is unparalleled because they have the highest mortality rate of any mental illness. And Anorexia Nervosa, which has its most common onset during adolescence, has a mortality rate of 20%--it has the highest death rate of any cause of death for adolescent females.

These startling statistics are publicly available, some of which are years old, however, there has yet to be a significant effort by any social media site to remove Pro-Eating Disorder posts or even mitigate their diffusion. Some social media sites have even adopted policies of non-interference, allowing for the unchecked spread of Pro-Eating Disorder content. And though other sites such as Facebook or Instagram have policies prohibiting Pro-Eating Disorder content, they have been unsuccessful in the moderation and removal of it because of the ever-changing nature of the hashtags and communities under which the posts are shared. Thus the only method for rooting out Pro-Eating disorder content would be to develop bots that are not led astray by false hashtags or coded words but can see through the deception just as any human could. 

With less than sixty percent of those diagnosed with eating disorders making a full recovery, Pro-Eating Disorder content is a threat to the fundamental well-being of teenagers and young adults. And by allowing for the proliferation of such content, the lives of millions of children worldwide are at risk. 

The solution to the devastating influence of Pro-Eating Disorder content on social media is to develop an AI model capable of recognizing Pro-Eating Disorder content, in whatever form it may appear. Then, based on the model's classification, report the post or advertisement to the social media site so that it may be removed, essentially bolstering the website's moderation, and helping to remove the harmful content.  

The solution utilizes state-of-the-art of the art Deep Learning models to analyze both the visual and or the textual components of social media posts to decide whether or not a post promotes Eating Disorders or unhealthy body ideals. Deep Learning models differ from classic computer algorithms since they can "learn" from data rather than being built for a specific task. The flexibility offered by Deep Learning models allows them to continuously improve without any updates of their code, thereby overcoming the clever changes in words or images that shield many Pro-Eating Disorder posts from detection. Moreover, Deep Learning models can process data incredibly quickly and efficiently. They exceed the capabilities of any human in both accuracy and speed, allowing for real-time classifications and quick recognition of Pro-Eating Disorder posts. And like any computer algorithm, Deep Learning models are able to work endlessly without the need for breaks, which offers a continual deterrent for Pro-Eating Disorder content. 

As was mentioned before, Deep Learning models require a dataset on which to learn. The model that was developed for this task, which is called Aegis, has been trained on a carefully curated dataset consisting of social media posts scraped from numerous social media sites and then manually labeled by the team. Based on what Aegis "learned" from the dataset, it is able to then make inferences on posts which it has not yet seen. 

Despite the numerous advantages of artificial intelligence technology and the increasing oversight and accountability to which social media sites are subject, there has been no substantial move by any social media site or external entity to combat the vast Pro-Eating Disorder movement which has ingrained itself into social networking. Numerous studies, articles, and open letters have called for greater intercessions on the side of social networking sites and the government to attenuate harmful posts. And yet, nothing has been done. Pro-Eating Disorder content is a clear and present danger but is often overlooked. Such disregard is not only because of profits (and yes, social media sites are making millions of dollars from Pro-Eating Disorder content) but also because of the taboo nature of mental illness, especially eating disorders. Many people believe that eating disorders are an affliction women suffer from, and they often correlate eating disorders with the stereotypical symptoms of Anorexia and Bulimia Nervosa. This marginalization of those suffering from an eating disorder, especially men and those who have less-known eating disorders such as binge-eating disorder, denies them the support they need. Moreover, the stigma surrounding eating disorders discourages those who are suffering from speaking about it or seeking help, which only worsens their condition. That is why so few have heard of, or truly grasp, the catastrophic nature of Pro-Eating Disorder content because those who suffer from it, have a learned aversion to help. That is why only ten percent of those who have an eating disorder get help. 

Of course, a systematic revision of societal stigma and pressure, albeit necessary, is far beyond the scope of this solution. What Aegis offers, however, is a tool with which the diffusion of pro-eating disorder content may be diminished, and the number of people exposed to its harmful influence is decreased. Additionally, those who are suffering from an eating disorder, or are beginning to develop one, can be protected from content that may seriously worsen their condition or cause a relapse. The Aegis model is capable of uncovering Pro-Eating Disorder content quickly and efficiently, shielding people worldwide from the impact of Pro-Eating Disorder posts. Aegis is the first model of its kind and is the first real effort to combat Pro-Eating Disorder content on a large scale. It can protect social media users, especially females and teenagers, from the indoctrination of Pro-Eating Disorder content, helping to make social a safer space. 

As I said in the last question, little has been said in the mainstream media about the rise in eating disorder cases since the advent of social media. Nor has much been said about the clandestine infiltration of malicious Pro-Eating Disorder content throughout the world's social media platforms. Subsequently, few know of, let alone understand, the dangers of such content, but I do. I, like many others, sought refuge in social media during the turmoil and uncertainty of the Covid-19 pandemic. During the long days of quarantine, I sought respite in the arms of social media. There I could communicate with my friends and see millions of other teens, just like me, who were facing the same difficulties; it truly helped me. But soon, the feeds of my social media apps became increasingly full of images and videos promising me control, something I desperately desired during such uncertainty. That control was over my own body. Then, fed by content that promised me happiness if I worked out or ate less, I began to cut out food, increase exercise, and seek out more and more of such content. Before long, the posts, which I now know to be of the Pro-Eating Disorder variety, became my sustenance and caused me to develop an eating disorder. 

Due to the stark spike in eating disorder cases during the pandemic, it took me months to get the help I needed. Needless to say, it was a difficult journey to recovery, one that I did not start willingly, but that is the nature of such an affliction: you are entirely convinced that you are healthy. It was only months after the end of my recovery that I understood how fortunate I was to receive help. So many others--people that I know--were not so fortunate. It was then that I decided to help stop the spread of Pro-Eating Disorder content so that others would not have to suffer as I did. 

At first, I was unsure how I could tackle such a daunting task, but then I came across an article that outlined how artificial intelligence techniques could bolster the fight against eating disorders online. That article inspired me to make the theoretical a reality: I would create a model capable of accurately classifying Pro-Eating Disorder posts to flag them for removal.

Thanks to my previous Python experience-- which I accrued while modeling Covid-19 transmission as a research assistant under a TCU professor--I jumped straight into implementing the algorithms. Though there was a steep learning curve to Deep Learning and AI, I educated myself using MOOCs. Through Coursera's scholarship program, I obtained a Data Science professional certificate from IBM--a rigorous program that teaches its students every aspect of data science, from data hygiene to machine learning. Also, thanks to a generous scholarship from the Dallas Foundation, I completed the Deep Learning Nanodegree from Udacity, a program that offers an in-depth collegiate-level education in all facets of artificial intelligence. During the nanodegree, I built many deep learning models as part of the coursework, including models that utilized cutting-edge transformer technology. Since then, I have developed numerous AI models academically and professionally, for example, while working as an intern at Optum, a part of the UnitedHealth Group, last Summer. At Optum, I developed a machine learning model capable of classifying stress levels within front-line workers, ie. medical staff. Using information from IoT devices, my team and I created a burgeoning machine learning model capable of classifying stress from numerous biosignals. 

During the summer, in addition to working at Optum, I began to develop the precursor to the Aegis model, which, unlike the newer prototype, could only process images. The model proved incredibly successful, and though it utilized only one modality, it was highly accurate. I then wrote a research paper about the efficacy of Deep Learning models when categorizing Pro-Eating Disorder images, which is publicly available on Cornell University's Scholarly Repository arXiv (https://arxiv.org/abs/2212.13949). Based on the success of my previous model, I have worked to create Aegis, which analyzes both images and text, thereby increasing accuracy.

Also on my team is Tal Feldman, a Truman and Schwarzman Scholar with expertise in Deep Learning and Artificial Intelligence. Tal is currently working at the Department of Defense and previously worked as an Artificial Intelligence researcher at the United States Federal Reserve and as an Advisor on Artificial Intelligence for the United States State Department. Tal is currently a senior completing his undergraduate education at Wake Forest University and the London School of Economics. During his undergraduate education, he published several groundbreaking AI research papers and has given numerous presentations on the future of AI research. Thanks to his industry experience, Tal is an asset to the Aegis team. 

My own experiences with Pro-Eating Disorder social media content is what first led me to create this solution and has since then fueled me throughout its development. But, if this model is to serve a vast array, all with different experiences, I knew that I would have to speak with others who have first-hand suffered the harmful effects of Pro-Eating Disorder content. Sadly, I know many people who, just like me, fell prey to the false promises of such posts. However, we're glad to share their experiences and provide input on how to best prevent users from viewing Pro-Eating Disorder content, because, like me, they wanted to prevent others from injury. 

Each person I spoke to told a story similar to mine. At first, social media was a safe place that helped them weather the tumultuous quarantine, but, before long, their for-you-pages were plastered with Pro-Eating disorder content, which they fell prey to. An important realization that I came to after speaking with my peers, is that Pro-Eating Disorder content did initially presents itself in its most potent form, but in the vessel of a seemingly harmless post, a wolf in sheep's clothing, if you will. Over time, more of these posts began to appear, and as they engaged and resonated with them, more extreme versions of Pro-Eating disorder posts came and soon began to engulf their entire social media sphere. 

The fact that Pro-Eating Disorder posts are disguised as harmless content, can trick both the algorithm and the user, leading them down a slippery slope. Most importantly, Pro-Eating Disorder content is not uniform and often lacks telltale signs which shield them from moderation. That is why artificial intelligence has such a crucial role to play in the uncovering of Pro-Eating Disorder content, for it can recognize the dastardly content for what it is and report it immediately. 

To this day, no substantive solutions have been offered to the ongoing encroachment of Pro-Eating Disorder content. That is not to say that others have not tried to raise awareness about the issue or lobby social media corporations to take steps to mitigate the problem, but none of these efforts have managed to garner large-scale results. A lack of manpower is the primary obstacle to sufficient removal of Pro-Eating Disorder content--the movement is far too large to remove or even track by a team of humans. That is why artificial intelligence will prove to be an invaluable tool in the fight against Pro-Eating Disorder content. 

The Aegis model is the first of its kind. No others, to my knowledge, have created a model capable of analyzing textual and visual data from a social media post and classifying it as either Pro-Eating Disorder content or regular content. The model uses state-of-the-art transformer models--a highly advanced type of deep learning model-- trained on a proprietary dataset. The Aegis model is then able to "read" a given social media post and nearly instantaneously produce a highly accurate classification, something no other technology can do. In essence, Aegis is the first technological solution to the ongoing Pro-Eating Disorder content crisis on social media and has the potential to be the most effective tool to combat this harmful content. 

However, the scope of Aegis's benefits is not confined to social media moderation but may spark a considerable change in the way the matter in question is viewed not only on social media but in the real world. An effective tool, such as Aegis, has the potential to not only help mitigate the problem but also raise awareness about it. Too long has the matter of Pro-Eating Disorder content been written off as unavoidable, and Aegis will show that Pro-Eating Disorder content can be stopped. Moreover, will create a system of accountability that ensures social media companies maintain their policy of deterrence against Pro-Eating Disorder content, rather than sitting idle in the face of it. And the widespread awareness this model may bring will have the general public keeping social media accounts too. Aegis is not a perfect solution to the problem, but it is the first attempt to combat such content online, and it may catalyze the creation of future technologies that, together, may stop the Pro-Eating Disorder Movement once and for all. 

Lastly, it's important to highlight Aegis's possible impact on AI implementation on social media. Some researchers have created Deep Learning models similar to Aegis, but they uncover posts that spread prejudice rather than Pro-Eating Disorder ideals. However, these efforts suffer from a lack of funding and awareness since most believe that the goal of an AI capable of uncovering such sentiment is unfeasible. But if Aegis is successful, it will show the world that AI can revolutionize the online sphere and safeguard against those who seek to spread malice and hate. In turn, this may inspire others to join the movement, even other teenagers, so that, together, we may work to make social media safer for all. 

Within the next year, my team and I hope to launch the Aegis model as part of a social media bot program on the major social networking sites Facebook, Instagram, Twitter, and Tumblr.  Altogether, these social media sites have over three billion users. Moreover, these sites have documented Pro-Eating Disorder communities and are most popular with teenagers and young adults. Once launched, these bots will use the Aegis model to classify posts they view as either Pro-Eating Disorder or not, and then report the post accordingly. Tumblr, Instagram, Twitter, and Facebook all have policies prohibiting Pro-Eating Disorder content. Thus, any Pro-Eating Disorder content reported to the site should be removed. However, as I said previously, social media sites do make substantial profits from Pro-Eating Disorder content and may not be inclined to remove it. Therefore, to increase accountability for social networking sites, Aegis will keep track of every post that is reported and removed. Reporting data will be displayed on the Aegis website and will be coupled with a social media presence proliferating the same information. Ultimately, the goal of Aegis is to report and, in turn, remove Pro-Eating Disorder content while also growing awareness about this issue and encouraging engagement. For there is no force as formidable as public opinion, and my team and I hope to use that power, along with the force of AI, to put an end to Pro-Eating Disorder content. Thereby, Aegis may protect online users from the influence of Pro-Eating Disorder posts and save lives.  

The core technology that powers the solution is artificial intelligence and its derivative Deep Learning. The Aegis model, the core of the solution, is a multimodal Deep Learning model built from an interconnected network of transformers. Transformers are deep learning units capable of wholistically analyzing input and focusing on the most informative section, similar to how a human can distinguish a face from the background in a portrait. The different sections of the Aegis model can read simultaneous textual and visual inputs, and then classify them. The Aegis model can also accurately analyze text or images on their own, allowing the model to classify a majority of posts on social media. Transformers and multimodal Deep Learning models are recent innovations and have yet to be appropriated to Eating Disorder research and offer a cutting-edge tool in the fight against Pro-Eating Disorder content. 

As of yet, the Aegis model has not yet been launched and therefore has no current. However, the solution, when launched, is capable of helping any social media user, especially those who are predisposed to or are suffering from an eating disorder. Focusing only on the United States, the country in which my team and I are located, there are over 28.8 million people with eating disorders, seventy percent of which are statistically likely to be on social media. As a result, the solution will help decrease the potential eating disorder triggers for over 20.1 million Americans, with the potential to help countless others. 

Social media by definition is open to everyone, and, subsequently, so is the Aegis model. Anyone, no matter race, age, sex, or creed can develop an eating disorder. For that reason, the Aegis model is available to everyone on social media and will help protect users from the harmful influence of Pro-Eating Disorder content. Aegis works in the background to ensure that all users are safe. 

The most substantial barrier to the solution is a lack of funding. To launch the Aegis model, it is necessary to have a computer, be that physical or a rented virtual one, with continued processing power and internet access to run it.  Regardless of what hardware the model is launched, it will require funding to maintain the operation. 

The other possible obstacle that may inhibit the scope of the solution's success is the cooperation of social media platforms. The potency of the solution relies on the belief that social media platforms will remove the flagged content once it is uncovered and reported. However, social media companies are not subject to many regulations and can choose whether or not a post violates their regulations. Thus, if a social media platform is not keen to remove Pro-Eating Disorder content for any reason, it may reject the flagged report, stifling the ability of the model to remove content. Though this is a serious issue, it will not dull the ability of the Aegis model to make serious positive changes. For you see, though social media platforms may reject reports, they are unable to silence the data. Once Aegis reports a post it will keep track of whether or not it is removed, thus generating an important statistic for the general public. This would create an important system of accountability for social media companies as the general public would be privy to the outcomes of the reports, encouraging social media sites to take a stand against Pro-Eating Disorder content. 

The model prototype does not require much funding to maintain functionality and is easily sustainable. The business model upon which our expanded solution rests is the value of its social impact. The source of revenue for Aegis comes from outside grants and benefactors who want to support the effort. Aegis, regardless of funding, will be launched on several social media sites and, based on further funding, expand to other social media sites. 

Aegis is, by definition, free for everyone and works to help all people regardless of status. Therefore, to continually fund our solution, outside funding will be necessary. There exist many grants which finance mental health and social safety initiatives, of which we hope to secure more funding to ensure the continuation of Aegis. In addition, charitable donations by groups or individuals who aim to assist in the fight against Pro-Eating Disorder content will also help to fund Aegis. 

Even without funding, Aegis will be able to function as a strong detterent for Pro-Eating Disorder content. Since Aegis only requires an internet connection and the requisite software, it is possible to launch Aegis on a social media site from a laptop. Thus, Aegis can still make an impact even before funding is secured. However, if Aegis runs on a local device, the model will be less potent because of hardware constraints, especially the lack of storage and processing power. This limitation will relegate the Aegis model to only a couple of social media sites, but the impact of the model will still be significant. Moreover, any advantages and protections that Aegis's work provides will still exist with or without funding.

As Aegis continues to grow and solidifies itself as an important tool in the online sphere, more people will become aware of the danger of Pro-Eating Disorder posts and will be inclined to give money to the initiative. Similarly, philanthropic organizations will be inclined to give money as well. Of course, part of our team's job is to seek out potential donors or grants and apply to them. By remaining proactive and stressing the importance of Aegis, it is possible to create financial sustainability. 

Lastly, the end goal of Aegis is the incorporation of our model into standard moderation bots on social media websites. With enough awareness and the backing of public opinion, social media websites will be inclined to incorporate AI measures to combat the proliferation of Pro-Eating Disorder content on their platform. In that case, social media platforms could purchase the Aegis model to incorporate into their existing algorithms and further improve upon. Such a purchase would continue to fund further initiatives in the future and would convey an active effort by social media companies to combat harmful posts, making the online sphere safer for everyone.