The Book Thief. Fahrenheit 451. 1984. The Handmaid’s Tale.

Nothing kills creativity and individuality quite as efficiently as the criminalization of it. Novels have been warning viewers about this phenomenon since The Advancement of Learning by Francis Bacon was published in 1605, and this ancient fear is one that has stuck around through the progression of time. Though often presented in a fictional manner, this fear has rapidly become objective throughout the information age. With the consistent banning of influential novels like To Kill a Mockingbird and The Catcher in the Rye, it raises the question of the true, long-term impact that this never-ending suppression will have on future generations. The answer: it’s a negative one.

By cross-analyzing a list of “20 Indispensable High School Reads” with a list of the top 100 most challenged novels, I was less than surprised to find that nine novels took spots on both lists. These ranged from coming-of-age books like The Perks of Being a Wallflower by Stephen Chbosky to ones regarding racial identity such as The Bluest Eye by Toni Morrison. With such a vast range of works, I began to wonder what the censoring groups were aiming to achieve.

These occurrences of censorship eliminate the sentiment of learning from history or risking condemnation to repeat it. If future generations are prevented from accessing these themes in digestible forms, how will they learn to sympathize with characters like Atticus Finch or Holden Caulfield? Where will they learn of the lessons taught by The Lord of the Flies, or how to handle the heart-wrenching pain felt in the end of Of Mice and Men? These experiences link generations together, and prohibiting future adolescents from absorbing the knowledge breaks this connection.

This topic recently resurfaced in my hometown. Despite holding other traditional ideologies, book censorship was never something it bought into. That was, at least, until Christmas of 2022 when a teacher at the junior high provided Looking for Alaska by John Green as an optional Christmas gift for her students. It spiraled out of control, ending up in news articles from out-of-state publications as well as earning a TikTok shoutout from the author himself.

After ending up on the desk of my county’s prosecutor, it ended with no charges filed. With the snowball effect that came with this case of censorship, it truly made me think about how traditional censorship can transition to an issue involving the ethics of information technologies. In this case, the escalation from a small-town issue to a global, social media frenzy forced it into said status.

With the increased usage of social media to air out ones grievances, it’s become progressively easier for traditional forms of censorship to grow into more modern instances. In recent years, it has largely translated from physical publications to those roaming the digital space, and websites like Twitter have turned this age-old debate into contemporary and modern.

With this transition from one media to another came a new question of responsibility. Following Elon Musk’s acquisition of the company, Twitter’s new status as a private company fuels this conversation. As a private company, Twitter’s actions are no longer dictated by what its stockholders deem as appropriate, leaving all of the power in Musk’s hands.

When the question of responsibility is added to the equation, the line between right and wrong becomes blurred. How much responsibility does a private company have for the platform it provides? How much responsibility does a poster have for the people that view their content? How much responsibility does a general user have for the content on any platform they choose to browse? Each of these questions is, by nature, interconnected, and it causes this confusion of right and wrong.

On a social media platform, the argument of censorship gains further depth. While eliminating the spread of hate speech and outdated ideologies is crucial for society’s necessary tolerance growth, this grey area causes uproar from all belief systems. This transforms the question from one of impact to one of extremity. Is it worse for everything or nothing to be censored? Each comes with its own downsides, of course, but is it worse to eliminate all possible freedom of speech, including things like time-sensitive news alerts, or to allow all possible freedom of speech, such as neo-Nazi ideologies? This is where the aforementioned issue of responsibility comes in, and it answers this question for you.

When it comes to modern censorship, the bulk of the responsibility must fall onto the general user (or their responsible guardian, considering 97% of 13- to 17-year-olds use social media). This all comes down to the unfortunate existence of a digital footprint. In an online environment, even deleted posts live in perpetuity. Whether it be through screenshots, word of mouth, or services like the Wayback machine, the most hateful of posts will remain immortal in one way or another. Though censoring the initial post can do some good in preventing this process, it hardly prevents any spread from happening at all. When these scenarios occur, it falls on the user to control their intake.

On top of this, online censorship often discourages users from continuing to create content. Content creators have been forced to create new terminology to use online called “Algospeak,” all to allow themselves the simple affordance of saying words like “sex” or “suicide”.

Through the act of “shadow-banning,” some more blatant or extreme cases of online censorship including artists having their work blocked and posts of plus-sized models being specifically targeted have occurred. “Shadow-banning” pertains to the removal, hiding, or muting of someone’s post without alerting them. By using censoring algorithms, social media platforms can invisibly censor a creator’s work in order to make the content on their platform more closely appeal to their ideals.

While this discourages content creators in general, the disproportionate censorship of minority groups on social media multiplies the discouragement of those in these communities. In a study conducted by students from the University of Michigan, it was found that conservatives, Black people, and transgender people are censored more often than other groups.

Conservative participants of the aforementioned study experienced their content pertaining to misinformation, hate speech, adult topics, or a relation to COVID-19 being removed the most. For the transgender participants, they found that following site guidelines was not enough to save their content from being marked as adult, and that their other removed content included anything critical of a dominant group as well as things specifically relating to queer and transgender issues. The Black participants in this study mainly reported that their most removed content related to racism and racial justice.

Despite the transgender and Black participants following the rules and policies of various social media posts, they found their posts being censored at a similar rate as those of conservative participants whose content, more often than not, broke these rules. If creators from marginalized communities will be censored regardless of closely they obey these regulations, why would they want to continue creating content anyway?

Self-censorship, also referred to as a “spiral of silence,” forces people to alter the content they share online in order to fit in with the popular ideals. As a byproduct of general censorship, this spiral can excessively impact those already overly affected by censorship, such as the groups mentioned previously.

Viewing the findings of this study in conjunction with unfortunate outdated prejudices and external cases of unnecessary censorship further speak to the ethics (or lack thereof) of this issue, and it all comes back to the question of whether or not censorship is a killer or a cure.

In order to finalize my thoughts on this topic, I chose to return to the fundamental meaning of censorship, which is the suppression of words, images, or ideas that are “offensive.” I found myself returning to the question of responsibility. After all, who gets to decide on what is or isn’t offensive? With algorithms putting awareness posts and hate-speech in the same box of unruly, why should we rely on censorship at all?

To be completely honest, we shouldn’t. Considering the lack of a concrete definition for what is or isn’t offensive (e.g. is “crazy” offensive? what is defined as racist?), it’s impossible for any one human being to control what is or isn’t censored. Relying on algorithms isn’t a solution either, as we’ve seen with the disproportional blockage happening on almost every mainstream social media platform.

Though censorship could prove beneficial when used correctly, it has become one of many go-to responses that relies far too much on a perfect, unbiased mind. Since this doesn’t, and will likely never, exist, the downsides will continue to outweigh and discredit the positives that censorship could offer. While continuing to inhibit creativity, suppress individual voices, and further marginalize already mistreated groups, it’s a clear solution: censorship is better off in your trash bin.

Censorship: Killer or Cure? was originally published in SI 410: Ethics and Information Technology on Medium, where people are continuing the conversation by highlighting and responding to this story.

As someone who is not deeply involved in the tech field, it can be challenging to know what’s really going on with new technology developments today. Recently, I found myself reflecting on how close technology is to reproducing human ideology, and quite honestly, it makes me a little anxious. There is all this talk of machine learning and AI development, but what does all of this really mean, and what are the potential consequences? Machine learning is “an umbrella term that refers to a broad range of algorithms that perform intelligent predictions based on a data set”. An algorithm is defined broadly as “a step-by-step procedure for solving a problem or accomplishing some end”, and in this case, the end goal of the algorithm is to learn patterns in a previous dataset in order to make predictions on a future dataset. As an example, an algorithm could use data about flu patients, such as age, time of year of diagnosis, sex, symptoms, and so on, to predict and improve future diagnosis of new patients. We as the general public must raise awareness of bias in these machine learning models and work to enact new laws regulating their development and societal use.

Often large data sets are used for these models, having upwards of one million unique data points. Since the algorithm begins with this input data, it will have an impact on the results. As a result, bias can be introduced into the algorithm based on the bias of the data entered. As a simple example, suppose an algorithm was created to predict the breed of a dog image shown. After the creation of the model, the algorithm correctly identifies a dachshund and a german shepherd but incorrectly identifies a pit bull. Why could this be? Suppose there are no or few pit bull images in the input data used to create the model. In that case, the model will be unable to accurately identify this breed compared to others. This basic idea can be extended to many machine learning models implemented in society today. Underrepresentation of communities in the actual world or misrepresentation of information will be reflected in the input data utilized by machine learning models if bias is not taken into account; this can have negative consequences and is a problem I believe needs to be addressed.

As a woman myself, it’s essential to bring attention to bias in these algorithms that have been shown in the workforce. In an IMD article it was shown that Amazon, a well-renowned company, utilized an algorithm that was found to be discriminatory against women.

Due to the high volume of applications, Amazon created a machine-learning model to help distinguish the best candidates for hire. In theory, considering the efficiency and time saved by using something like this, it appears to be a good idea. However, it was not well thought out. To train this algorithm, Amazon used resumes that had been submitted to the company in the past ten years. Where this algorithm fell short was in taking into account the sex of the majority of people who work in a technology-based company. As we all know, men dominate the tech industry, and Amazon’s model reflects this bias. The algorithm used the input data and developed a pattern to favor male candidates over female ones. As a result, even if you were a strong female candidate, your application would have been rejected solely on the basis of your sex. The model Amazon created caused “computational discrimination whereby unfair outcomes privilege one arbitrary group of people over another”. This is a glaring issue that demonstrates how bias is a problem in technology that must be addressed. Many women are likely to have missed out on opportunities in the workforce as a result of the careless creation of a machine learning algorithm. Although Amazon discontinued the use of this recruitment tool, it demonstrates the severe consequences that this bias can have on people.

Apart from the workforce, these algorithms are widely used in the medical field. These models can be used to make predictions regarding patients and their diagnosis. Previous symptoms, demographics, age, and other meta data can be used to make predictions and aid in the prevention of certain diseases before they develop. This highlights the crucial nature of bias mitigation because it has the potential to affect people’s lives.

Although I have no direct experience of being affected by these algorithms in the medical field, I have learned about some that concern me greatly. One day I was sitting in a Introduction to Bioinformatics class as a University of Michigan student. This class covers a wide range of topics, including machine learning algorithms in the healthcare field. I distinctly remember learning about one machine learning algorithm that was found to have extreme bias implications by University of Michigan researchers. Epic, a electronic health record system aiming to develop software for the medical field, was the center of this conversation. They released a software including a machine learning algorithm that predicts wether a patient is at risk to develop sepsis. As my professor was talking, he revealed that due to this algorithm, around two thirds of sepsis cases were missed. Hearing this issue made me particularly uneasy because of how careless the situation was. This algorithm was not thoroughly examined before it was made available for usage. The output of these algorithms developed by experts in these technological domains is relied upon and trusted blindly by clinicians as they are by nature not computer scientists. This puts the liability of these algorithms’ correctness on their developers.

When the system was finally re-evaluated, issues were found due to a lack of rigorous evaluation of bias and lack of transparency throughout the construction of this program. Instead of using clinical criteria, the algorithms for sepsis detection were developed based on billing data. To mirror real-world sepsis cases, a multitude of predictors needed to be included, not just one subset. In doing so, there was a high rate of false positive rates, and a lower accuracy in predicting cases of sepsis that may not have been detectable by a clinician. This problem, however, went unnoticed for far too long and serves as a reminder of the absence of rules governing the development and widespread application of machine learning algorithms. I am rather nervous when I learn about this bias and that it exists in the medical industry. How can we be sure that these algorithms will be accurate when utilized in the field if there are no rules governing how they are developed and validated? It’s unnerving enough to rely on computers, but learning about this technological mishap that resulted in potentially fatal complications has proven to me that this is an issue that needs to be addressed.

How then do we get past this bias? The most simplistic solution is to improve the rules that govern the creation and implementation of these models. In the instance of Epic, the issue was with the model’s validity and transparency. Lack of testing was done to confirm the algorithm’s accuracy since it was solely created within the company. There were insufficient impartial assessments to support the reliability and accuracy of their model. It wasn’t until subsequent testing by a different group that it was determined that it would not be suitable for use in a clinical context. This issue could have been completely prevented if this model had been thoroughly tested and validated prior to release.

Secondly, there should be more awareness surrounding how the datasets are chosen to train these models to make predictions. Reducing sampling bias is one of the main issues that should be addressed. Sampling bias is the act of systematically favoring some population members over others. Making sure the dataset being utilized is a truly random sample that can be generalized to the population of interest is one approach to fix this issue. It is essential to verify that each subgroup is represented equally in the training dataset. This should enable the model to produce more accurate and impartial predictions as a result. When developing these models, some corporations have been relatively careless and oblivious of the bias that they may be introducing into these systems. As machine learning technology becomes a more integral part of society, it is crucial to raise awareness of these fallacies. A rigorous evaluation of the data being inputted is necessary for large organizations like Epic or Amazon to become more conscious of these biases they may be introducing into algorithms.

Uncovering the Hidden Dangers of Bias in Machine Learning Algorithms was originally published in SI 410: Ethics and Information Technology on Medium, where people are continuing the conversation by highlighting and responding to this story.

Taking advantage of many attributes of the COVID-19 pandemic, telehealth startups — namely Cerebral, Done, and Hims/Hers — began to diagnose Attention Deficit Hyperactive Disorder in ways that blurred the line between treatment and profit. Emerging from a telehealth-reliant world, the consequences are starkly apparent.

—

Through the repeal of the Ryan Height Act, telehealth diagnostic companies seized the opportunity to capitalize on the now-allowed online prescription of Schedule-II medication, including stimulants for ADHD. Machine-automated medication suppliers, such as Truepill, further assisted telehealth expansion by allowing prescriptions to be bottled, sealed, and sent to one’s local CVS faster (and cheaper) than ever.
Appealing to a time in which patients required reliance on their own symptomatic knowledge due to lack of in-person care, social media-based telehealth ads provided direct-to-consumer language encouraging the viewer to analyze possible signs of ADHD within themselves:

7 SIGNS YOU MAY HAVE ADHD
WHAT IT’S LIKE HAVING ADHD
I THINK I HAVE ADHD…

In response to these factors combined, mental health-related telehealth startups experienced a major financial boom. Cerebral specifically stated to have currently treated over half a million patients during its almost three years of operation, alongside a 2022 evaluation of $300 million. Such success, however, is paired against the suspected public backlash: ADHD telehealth advertisements were “broad” and “predatory”, big-pharma fears of the evil businessman who wants to sell your children Adderall and Xanax were all too real to the viewer outside of Cerebral’s expected social media audience.

Insider voices, however, reflected similar panics. Nurses who had associated with mental health telehealth companies complained that they did not feel as though they had adequate time to work with patients before being required to give a prescription. Lawsuits appeared regarding CEOs demanding every evaluation ends with a diagnosis. Former employees discuss ending their time with companies out of uneasiness towards the brand’s approach to medication, such as through label aesthetics. Federal investigations aimed to conclude whether prescriptions were being handed out accordingly.

“Seven former nurses for the company say they worried that Cerebral wasn’t merely meeting a demand but was also, by making access so easy, effectively creating it”

—

Through new observations encouraged by the unusual circumstances of the pandemic, telehealth’s praises and problems contributed to further muddying the waters of ADHD. Outcries over plain statics (ADHD has risen in diagnosis over 10%! More than 500,000 new children have been prescribed stimulants!) paired with sketchy business models and unrecognizable medical technology created a black-and-white diagnostic scenario of ADHD diagnosis.

As Tricia Hersey insists in her manifesto “Rest is Resistance”, people are best to “survive, not thrive”, as capitalism requires the direct opposite. Leaning into the pandemic, this notion became muddled: the requirements to survive were now much higher, but the internal need to thrive in “productive” areas of life was impossible to meet or see (think leaving the office vs. closing your laptop while working from home).

With almost the entire population being forced into new, more stagnant pandemic life routines, feelings of unproductiveness, fatigue, and self-consciousness flourished within ADHD and non-ADHD civilians alike — our ingrained capitalistic desire for productivity was far from adequately met while staying at home, enhancing the shame of “not doing enough”, allowing defiant guilt which resonated with ADHD patients prior to the pandemic to now be adopted by non-ADHD citizens.

The language used by telehealth advertisements took quick advantage of these overlapping symptoms: ADHD, pandemic burnout, general anxiety, trauma, etc. are generalized through quick and funny video format. Companies like Cerebral’s swift APIs allow anyone to have instant access to medical professionals — many of whom have admitted to not being qualified to assess and diagnose conditions like ADHD, yet were hired to do so regardless. Our constant negative state of pandemic and ADHD lifestyle alike is aggressively reminded through telehealth advertisements:

FATIGUED
CAN’T FOCUS
OVEREATING
BAD MEMORY

Forever living in a world of stigmatized stimulants, it is striking to see advertisements of young adults shaking bottles of Adderall while dancing with a smile on their faces. Symptoms of the disorder have similarly 180-ed, aiming to include many participants rather than connect with the niche situations worrying those who may have ADHD or a similar disorder. All-encompassing videos from telehealth advertisements have been seen throughout other social media realms, with a study from the Canadian Journal of Psychiatry finding more than half of ADHD TikToks researched to be considered misleading.

The introduction of profitable telehealth to ADHD treatment, however, makes the word “misleading” difficult to dissect, the same way it was previously difficult to label telehealth advertisements as strictly “predatory”, even though the public is able to witness the rising numbers in ADHD diagnoses and stimulant prescriptions. Considering the effects of such social changes and their impacts on the individual, we are encouraged to view post-telehealth statics regarding ADHD through the lens of the sixth principle of data feminism: considering context.

The bottom line for numbers is that they cannot speak for themselves. In fact, those of us who work with data must actively prevent numbers from speaking for themselves because when those numbers derive from a data setting influenced by differentials of power, or by misaligned collection incentives (read: pretty much all data settings), and especially when the numbers have to do with human beings or their behavior, then they run the risk not only of being arrogantly grandiose and empirically wrong, but also of doing real harm in their reinforcement of an unjust status quo.

If considering the context of new diagnostic statistics, paired alongside background knowledge of profitable telehealth companies, the conversation of ADHD within the scope of the pandemic should heavily consider medicalization’s reliance upon social factors. It is, in fact, only recently that we have had the capability to expand and understand ADHD in terms of adult functionality, rather than seeing it as child-only. The same can be said for the introduction of stimulant medication: ADHD is a great example of changing genealogy among diagnostic criteria.

Before the popularity telehealh, Peter Conrad describes within “The Changing Social Reality of ADHD” that the evolution of technology is bound to impact how we learn about, diagnosis, and treat ADHD. Telehealth plays into this prediction all too well, reflecting other discussions of Conrad’s as well, such as that ADHD treatment for older patients requiring a more adaptable approach than that with children due to factors like employment or higher education. This “adaptable approach” usually finds itself in the form of stimulant medication, while children are more likely to receive benefits such as cognitive behavioral therapy.

Information as such continues to confuse companies like Cerebral’s messaging, lacking answers to questions similar to: How is telehealth bringing new “adaptable approaches” to the table when treating ADHD? How are telehealth visits keeping in mind the pandemic affecting necessary activities such as employment and education? Is telehealth prepared for the new challenges patients will face within the pandemic/virtual work or learning?

In attempted answer, previous CEO and creator of Cerebral Kyle Robertson responds in his Medium article as to why he started the company that “we believe in the power of comprehensive care”. However, Roberston continues

“our care managers and prescribing providers alike know that medication alone may not be enough for many of our clients… First, we seek to provide easy access to the highest quality medication management for anxiety and depression using telemedicine. If we can do this successfully at scale, we will expand to provide a broad array of options for anxiety and depression, such as cognitive behavioral therapy and mindfulness”

Such reflects Conrad’s predictions of the expansion of ADHD: Cerebral’s idea of telehealth lies more within telemedicine than online, qualified conversational care and personal patient attention. While this may provide a decent bandaid while restricted within the pandemic, Robertson’s approach to ADHD treatment strives to uphold the factors negatively affecting ADHD and non-ADHD patients alike through a strong capitalistic drive in treatment, advertisement, etc.

—

Beginning 2023 with a well-known Adderall shortage, many peak-pandemic repercussions of telehealth approaches to ADHD are present in everyday media. A lack of comprehension on a day-to-day level presented an extremely difficult struggle to many, ADHD or not. With telehealth startups having the ability to provide the public an easy-grabs stimulant prescription and unqualified diagnosis, they fail to
discredit stereotypes and moral panics regarding ADHD that have complicated the medicalization and accessibility of the disorder/diagnosis for decades.

Telehealth’s practices throughout COVID-19 create further difficulty in analyzing statistics regarding ADHD and its corresponding practices, allowing for biased or extreme-scaled public discourses of the disorder. Using the information available about the overlap of ADHD and telehealth, observations of diagnostic criteria should be closely read into to understand the effects America’s society will witness in the upcoming months or years.

The COVID-19 Pandemic is Flipping the Script on ADHD was originally published in SI 410: Ethics and Information Technology on Medium, where people are continuing the conversation by highlighting and responding to this story.

However, as the saying goes, with great power comes great responsibility, and OpenAI has taken on the task of responsibly managing the power that comes with such a world-changing technology. One area of concern is the controversial issue of moderation when creating language models. In my view, OpenAI may be going too far with moderation, and seems to be moving farther away from the values it originally was founded on (OpenAI transitioned from a non-profit to a for-profit in 2019).

As ChatGPT’s capabilities and the language model it is built upon, GPT-3 and its subsequent version GPT-3.5, have been extensively studied and documented, this blog post will reflect my reaction to the recent release of GPT-4 as well as recent developments and observations from OpenAI’s latest updates. However, the focus of this post is not on the impressive abilities of ChatGPT and its evolution but rather on the excessive attention (in my view) given to moderation in its development process and GPT-4’s release.

Several individuals have pointed out that ChatGPT has been subject to increased moderation due to ethical concerns and continuous utilization since its original release, citing increases in refusal of prompts, less creative or otherwise interesting answers. Although there are not consolidated studies measuring the quality of ChatGPT’s interactions over time other than official benchmarks on standardized testing metrics, online communities have countless of examples of users noticing that a type or quality of response that was available before have been ‘nerfed’, or otherwise limited.

We’ll discuss the type of content that is intended to be denied and the reasons for denying it, but most notably, even non-threatening content, such as code snippets, have been denied in attempts at moderation. Some users have also expressed frustration with the model’s seemingly strict criteria, as even prompts in the vein of “write a story with elements of Disney’s [insert movie title here]” which are obviously innocuous, have been rejected as prompts. This excessive moderation has led some, including myself, to question whether OpenAI is crossing a line in terms of fair use for the technology which has been so revolutionary.

This has sparked a debate on the potential harm of language models. Recently, OpenAI released GPT-4, accompanied by a technical report spanning over 99 pages. Unlike previous papers discussing advancements in language models, which delved into intricate details about the model’s mathematical foundations, and particular training methods, this paper lacks such information. While this is not the primary focus of this post, it is crucial to note as a start of a potential trend of secrecy.

The focus of this technical revolution surrounding GPT-4, at least in my understanding from the paper, appears to be on content moderation rather than improving the accuracy of the model or advancing computation capabilities towards achieving true AGI. While the concern for moderation is understandable, I suspect that OpenAI’s heavy moderation decisions may have been influenced by partnering corporations such as Microsoft, who licensed integration of OpenAI’s models into Bing search. Personally, I was disappointed with the lack of progress in the latest iteration of advanced language models, GPT-3.5 (ChatGPT’s old standard, as well as the current standard for non-paid users) which until recently was considered the state of the art autoregressive transformer language model.

There appears to be a contradictory energy within OpenAI’s mission, as they have chosen to leave many implementation details closed source despite their original intention of promoting AI for all. This has been a disappointment for me, as the bulk of the report focuses heavily on moderation and potential harm rather than technical implementation details.

The researchers in the technical report have placed a significant focus on not replying to specific requests. Some of the issues they have addressed include:

• Hallucinations
• Disinformation
• Potential for risky emergent behaviors

… and others.

However, there have been concerns about bias in OpenAI’s moderation. For example, their moderation toolkit has been found to rank statements based on perceived offensiveness with discrepancies such as deeming statements like “I hate Republicans” as less offensive than “I hate Democrats.” Additionally, a study from the University of Munich suggests that ChatGPT has a pro-environmental, left-libertarian ideology based on a testing suite of various prompts.

David Rozado, who has many blog posts regarding biases in language models and ChatGPT specifically, notes further flaws in moderation have been identified, such as OpenAI’s content moderation system classifying negative comments about disadvantaged demographic groups as hateful, except for negative comments about conservatives/Republicans. This raises questions about whether AI systems should treat all demographic groups equally or display preferential treatment towards vulnerable groups. The systemic biases within OpenAI’s moderation suggest a blind spot or indifference/contempt towards disfavored demographic groups, and similar biases may exist in other big tech companies’ content moderation filters.

In my opinion, it is possible that OpenAI is not intentionally imposing any particular ideologies onto its language models, but rather these biases may result from associations made during training. For instance, since hateful content is often directed towards specific groups, language models might inadvertently develop a preference for certain groups when trying to learn behaviors like “fairness” or “uncontentiousness.”

According to an article in Technology Review, OpenAI’s AI policy researchers, Sandhini Agarwal and Lama Ahmad, are working to improve the reliability of ChatGPT by removing instances where the model has shown a preference for false information [and presumably, biases]. Additionally, the company plans to develop a customized chatbot that can represent diverse perspectives and worldviews to allow users to generate responses that align with their political beliefs. However, Agarwal acknowledges that this process will be challenging and lengthy.

There are valid concerns about the potential dangers of AI, but there is no concrete evidence to suggest that under-moderation would have a significant impact with the current level of intelligence. Even if ChatGPT refused to disclose relevant information about illegal goods, for example, or committing tax fraud without getting caught, those interested in pursuing such actions would likely find other pertinent avenues on the internet.

I was able to find evidence the other way, suggesting that chatbots’ personalities may have an effect on its users, depending on the circumstances.

One should also take into account a long-term impact on free speech. Imagine a world where closed-source language models form the basis of the internet, including search engines. What if these same models were responsible for generating a significant portion of the content we consume, with the writers’ creativity and original thinking taking a backseat to the models’ programmed biases? Such a scenario could pose a significant threat to free speech and creativity, subjecting them to the whims of whoever decides to regulate these models’ output. While we may not be close to this reality yet, it is worth considering.

Moderation may have several motivations in the early stages, including potential use cases [like teaming up with Bing, as I mentioned before] and trajectory for their models that are not available to the general public or known by executives. However, given the rush to release the GPT-4 release and report with minimal technical information and limited accuracy boosts in specific domains (test-taking, etc.), this seems unlikely.

Regardless of the motivations behind it, moderation carries its own set of risks. OpenAI is responsible for determining what information is suitable for users to access, and while unmoderated systems can be dangerous, moderated systems can also pose similar concerns. OpenAI has the power to control what information users can access, and if language models like ChatGPT are used in more scenarios, the effect of moderation may extend beyond the current chat box on openai.com. This is a concerning possibility, especially as technology continues to advance and sprawl to be integrated in more and more aspects of our daily experience on the internet.

So what’s the solution here? Clearly the battle of how to deal with online moderation isn’t one with an end yet. However, there are various potential strategies that I would like to see the effects of. One is to expand the models by allowing users to fine-tune their preferences for moderation,suggested by OpenAI policy researchers in a recent interview. Another one is the endorsement and emergence of open-source alternatives to the existing models, which would promote competition and prevent the formation of monopolies in these fields (which, seems to be picking up after Meta’s language model got leaked.) It is also conceivable that the internet could undergo a major shift in its direction, necessitating the adoption of new methods for disseminating information. It is even plausible that there could be a resurgence of particular interest in non-generated content, with a greater emphasis being placed on original thought. But even then, how will we be able to tell the difference?

Is OpenAI Going Too Far with Moderation? was originally published in SI 410: Ethics and Information Technology on Medium, where people are continuing the conversation by highlighting and responding to this story.

Section 230 of the Communications Decency Act was originally passed in response to a defamation lawsuit from Stratton Oakmont, made famous because of the movie The Wolf of Wall Street, against an internet service provider, Prodigy. Prodigy was found liable as it had moderated older posts before, meaning the court looked at them as a publisher. After this decision was released, Congress stepped in by creating and passing Section 230. Section 230 allows websites to moderate their platform as they see fit, without fear of being legally liable. Since the passing of Section 230, large websites have risen, such as FaceBook, Twitter, and Omegle, which present entirely new problems the government didn’t foresee.

The biggest of these problems is how websites are using Section 230 to avoid numerous civil lawsuits. Websites have years of precedent in their favor, as judges have the trouble trying to decide if something a user posts/does on the website is simply them exercising their right to free speech online or if the site being used has some severe flaws in it. Some examples of other lawsuits that have been thrown out due to Section 230 are a harassment case filed against the app Grindr, and a lawsuit against Google claiming YouTube helped radicalize the perpetrator of the 2017 Paris Attacks. While the idea of allowing these websites to moderate as they seem fit may seem like a good idea, companies have started to rely on the clause to get away with ethically unsafe practices. In 2009, the strength of Section 230 was slightly reduced in a change known as FOSTA-SESTA, which doesn’t provide protection from federal trafficking laws.

Personally I can recall in school learning about, what at the time was referred to as internet safety, or in other words how to act on the internet. However, everything that was taught during this time, was the opposite direction in which the internet was evolving. Namely the advice to not interact with strangers over the internet. This advice soon grew harder to follow as more and more social media companies were operating where their primary goal is to increase those connections between strangers. The simple fact is that even a few years after Section 230 was passed the internet was still changing rapidly to the point where there became far more dangers on the internet in a few years than previously.

Returning to the Omegle case, the most interesting aspect from it, is that Omegle is being sued over the fact that there is a product liability, meaning the plaintiffs are arguing Omegle’s design is at fault. In a different sense the argument is similar to if you bought a child’s toy, which had a small piece consistently break off leading to a choking hazard, a defect in the design that the company is liable for. This essentially bypasses Section 230 of the Communications Decency Act, which protects social media platforms from being sued for what’s posted on them. The argument that Omegle has a design flaw is quite a strong one too, as there are reports dating all the way back to 2013 raising this issue (Tidy, Joe. “Omegle: Suing the Website that Matched me with my Abuser”). “They’ve made it actually easier to navigate, easier to jump in. It was a bit difficult last time. More warnings were available, but now you can primarily sign in anyplace, from your phone or wherever you might want to” said Charleston Police Officer Doug Gallucio, after investigating Omegle in early 2013, after initial claims surfaced of pedophiles using the website to find victims (Jacobs, Harve. “New Concerns over Pedophile Paradise”). Simply, Omegle is one of the worst examples of an unsupervised/unregulated platform can do to our society.

While Omegle isn’t the only platform with these ethical and possible legal issues, Omegle does serve as an example as to how much protection companies felt they had due to Section 230. For years Omegle was aware of the issues their site faced, and instead of trying to implement new features to keep users safe, they simply added a warning on the loading screen. Possibly if there was an actual threat of legal action, a solution would have been implemented. Additionally, the moderation of the site was severely lacking allowing things like graphic pornography, violent extremism, and hate speech. The allowance of graphic pornography is so rampant there are always jokes surrounding the amount of dicks you will see while browsing Omegle.

Until this point the main perpetuated I had discussed was Omegle, but other companies also have been using Section 230 as a shield. These companies haven’t been as blatant in their disregard for their users. The other big examples for this would be FaceBook and Twitter struggling to moderate all the misinformation that gets posted to their platforms. If these companies knew that the way they moderate their platform was going to be scrutinized, would that cause a significant change in how companies do moderate their platform? I personally know that most companies want to prioritize moderation of content that is being talked about from people outside of the site. I can recall most platforms not having a problem with someone like Andrew Tate, until there was an outcry against him. Until that point there was a higher level of engagement on the site, because of people trying to denounce him in the replies of videos. Additionally a large amount of misinformation is spread by bots on these platforms, and by spreading these narratives they increase engagement on the site.

Considering the main goal of any company is to bring in a profit, then there is an issue at hand when companies are profiting off events that hurt society. A prime example of this is when Jamal Koshoggi died a bot-net was deployed to downplay the involvement of Saudi Arabia in the killing. By deploying this bot-net on Twitter, it increased engagement on the platforms, generating a profit for the company. Other more benign uses bots have on a social media platform would be to artificially increase the number of users on the site, or simply for one user to seem more popular by having a large following of fake people.

The Omegle case will certainly be used in the future when inevitably the Supreme Court takes a case involving Section 230, but there are some other important cases that have already happened. One in particular, involving Snapchat and their speedometer filter was ruled against Snapchat. The judge agreed that the filter encouraged users to reach faster and more dangerous speeds than if it were otherwise not there. That decision showed companies can be held liable when their products are severely and repeatedly damaging society, but this case had clear impacts to be drawn upon within the world outside of the digital space. Currently, the main issue facing the most notable of these platforms is the misinformation being posted to it, which doesn’t have as clear of an effect on society. That lack of affect could also be traced back to more misinformation to convince members of society there is no problem.

While Section 230 does help the freedom of information on the web, in its current state it is more a barrier helping companies continue unethical practices without fear of legal repercussion. The fact that a company such as Twitter, can allow large amounts of misinformation to be spread with little to no effort put in to stop it, but can’t be held responsible for the causes of that misinformation is alarming. As consumers we should be trying to demand transparency from the companies we use on a daily basis. Currently these companies have a way to be less transparent in how they moderate their sites and many other aspects (expand later). While this debate rages on, the Supreme Court gears up to hear cases regarding Section 230, potentially leading to its demise or strength in the future of the internet.

Links for sources used:

https://www.bbc.com/news/technology-64618791

https://www.eff.org/issues/cda230#:~:text=Section%20230%20allows%20for%20web,what%20content%20they%20will%20distribute.

https://www.theverge.com/2022/7/14/23216386/omegle-lawsuit-section-230-district-ruling

https://www.nytimes.com/2020/05/28/business/section-230-internet-speech.htmlhttps://www.cnbc.com/2023/02/21/supreme-court-justices-in-google-case-hesitate-to-upend-section-230.html

https://www.cnbc.com/2023/02/21/supreme-court-justices-in-google-case-hesitate-to-upend-section-230.html

https://www.cna.org/our-media/indepth/2021/04/social-media-bots-and-section-230

https://www.techdirt.com/2020/08/10/section-230-isnt-why-omegle-has-awful-content-getting-rid-230-wont-change-that/

How the laws unfairly protect big social media companies was originally published in SI 410: Ethics and Information Technology on Medium, where people are continuing the conversation by highlighting and responding to this story.

Social media has undoubtedly revolutionized the way we consume news and information. With the advent of platforms like Twitter, Facebook, and Instagram, the news is now just a click away. In 2019, the Pew Research Center found that over half of Americans (54%) either got their news “sometimes” or “often” from social media, and Facebook was the most popular social media site where American adults got their news.

I am also guilty of this.

Even though I subscribe to official news platforms such as the New York Times and the Wall Street Journal, I often find myself receiving updates about the world on Instagram via newsletter accounts such as @shityoushouldcareabout and @so.informed. In between posed beach posts and aesthetic food pictures, I would pause to read about Iran school girls poisoned by toxic gas and Roe V. Wade updates in the U.S. In a later survey from the same year, the Pew Research Center reported that 18% of American adults reported that the most common way they get news about politics and the election was from social media. I too, instead of actively searching for information about neutral viewpoints on political candidates, complied with the solely left-wing-leaning content on Instagram because the accounts I follow only curate such content to me. While social media has made accessing news easier, it is important to recognize that using social media as a primary news source is not always a good idea. In fact, it can be detrimental in several aspects such as misinformation, bias, and more.

One of the most prevalent issues with news on social media is that it is the perfect breeding ground for fake news. The standards of journalistic integrity that apply to traditional news outlets are not applicable to social media platforms. There are little to no restrictions on the content, so often, stories are shared without any fact-checking, creating fake news and misinformation that harms our society.

In a Medium article about fake news, Claire Wardle discussed her list of motivations for creating false information: Poor Journalism, Parody, to Provoke or ‘Punk’, Passion, Partisanship, Profit, Political Influence or Power, and Propaganda. During the peak of the COVID-19 pandemic, driven by profit and even power, a medicine company created false information about their capsules being able to cure COVID-19. Many users on WeChat, including my mother, believed that this medicinal remedy called Lianhua Qingwen, released way before the start of the pandemic, in 2003, could cure and prevent the COVID virus. It all started with a post circling around WeChat and other platforms through retweets, stating that this traditional Chinese medicine formulation can effectively prevent catching the virus, and is even capable of putting an end to COVID symptoms in a blink of an eye. People in China were paying thousands to get their hands on some of this medicine; while overseas, people who heard the words would try any means to smuggle some of the Lianhua Qingwen medicine disguised in snack packages. However, the viral post at the time did not outline the adverse effects that the Lianhua Qingwen capsules would cause: such as nausea, vomiting, and diarrhea. This led to people believing that one of the COVID-19 symptoms is diarrhea, when it was late realized that it was more likely a result of ingesting the capsule. Until today, Lianhua Qingwen is still not FDA approved and it is officially warned as a fraudulent COVID-19 product that does not mitigate, prevent, treat, or cure COVID-19 in people. Whoever started spreading the word about the medicine as a cure took advantage of the mass panic sentiment in which people would do anything to not catch the virus and stay healthy. People ended up making important decisions based on false information, leading to serious consequences where they wasted a significant amount of money and energy on something that meant nothing.

Not only are social media platforms a hotbed for misinformation, they are also echo chambers for users. Social media algorithms can create confirmation bias since it is trained to feed users what they are more likely to engage with. People may consume news that only reinforces their existing beliefs, rather than seeking out a range of perspectives. I am personally aware of this issue, yet I still fall for it. News outlets on social media are heavily skewed perspective-wise. Even if the newsletter account consciously tries to present content that comes from a neutral point of view, the comment section is never a balanced seesaw. In the comment section where users can freely express their thoughts, we would naturally expect it to be a cauldron of mixed opinions. However, only people of similar demographics and political opinions would follow the same accounts and actively seek information on similar topics. Thus, when I scroll through the comment thread looking for other people’s opinions and hoping for differences, I often find most of the comments to echo my personal belief, which is a reassuring feeling, but I have to acknowledge the fact that it is not the healthiest way to absorb information, especially when it comes to politics. Being in an echo chamber can be comforting, but also strange and disorienting because it can feel like being trapped in a bubble of my own making. Through regular newsletter sources, one can more easily seek opposite beliefs and dissenting voices. Since social media platforms are self-curated social bubbles where people tend to post and retweet similar content, it can be difficult to escape the echo chamber. This can lead to a polarized society, where people are less willing to engage with those who hold different opinions. In order to overcome this, we should actively seek out diverse sources of information, engage with people who have different views, follow diverse voices on social media, be aware of our own biases, and practice critical thinking.

Social media news can be sensationalized and exaggerated. Social media platforms thrive on engagement, and as a result, news stories that are sensational or emotionally charged are more likely to go viral. This has led to a rise in clickbait headlines and stories that are designed to provoke strong emotions. One example of a sensationalized social media news story is the “Pizzagate” conspiracy theory that circulated on social media platforms in 2016. The theory claimed that high-ranking Democratic Party officials were involved in a child sex ring operating out of a Washington, D.C. pizzeria called Comet Ping Pong. The theory was based on a series of hacked emails from John Podesta, Hillary Clinton’s campaign manager, which was released by Wikileaks. The theory was quickly debunked, but it gained traction on social media platforms due to sensationalized headlines and false information spread by conspiracy theorists. Some social media users even went so far as to threaten the restaurant and its staff with violence. The Pizzagate conspiracy theory is an example of how sensationalized and false information can spread quickly on social media platforms and have real-world consequences. We need to be critical of information found on social media and verify sources before accepting any information as true. While sensationalized news may generate more engagement, it can be harmful in the long run. People may end up consuming news that is not accurate or well-researched, leading to a distorted view of the world.

Social media news can be overwhelming. With the constant barrage of news stories, notifications, and updates, social media can be a stressful and anxiety-inducing experience, especially when it comes to coverage of major events such as natural disasters, mass shootings, or political crises. Sometimes I just want to scroll through my friends’ feeds to relax before bed, but the disturbing news that I did not ask for at that specific time makes me end up feeling overwhelmed and unable to process all the information that is being presented to me. This can lead to a sense of helplessness and disengagement, where people become apathetic to the news and the world around them. For instance, sadly, nowadays when a mass shooting happens, all media would be covering the event, but the time we would pause from scrolling to read about it is gradually decreasing, and even if we do, we simply move on to our regular life on the next day.

While social media has made accessing news easier than ever before, it is important to be aware of the drawbacks of using social media as a primary news source. Social media news can be unreliable, biased, sensationalized, and overwhelming. To ensure that we are well-informed and engaged citizens, it is crucial to seek out a range of perspectives and to consume news from a variety of sources. By doing so, we can avoid the pitfalls of social media news and stay informed responsibly and meaningfully.

Fake News, Real Folly: Why Using Social Media as Sole News Source is a Hashtag Disaster! was originally published in SI 410: Ethics and Information Technology on Medium, where people are continuing the conversation by highlighting and responding to this story.

Freaknomics Podcast, Why is U.S. Media so Negative does a great job of explaining this concept. Two accredited researchers found that an overwhelming amount of U.S. media used negative terms linked to fear, hate, and morality to produce more “eye-catching” stories. I am sure many of you are familiar with the term “click-bait”, or, a misleading hook that grabs out attention so quickly we can’t help but click on the page. In many ways, these negative and eye-catching terms serve as click-bait. This is because the human brain tends to gravitate towards negativity because it seen as more important to our brain’s psychology. Terms linked to fear, hate, and morality stick out to our brains more, giving us the final push to click on the news story and engage in the content. Our brains hold on to these negative sentiments as they create patterns of anxiety and fear. I am sure you and I are more likely to remember words like “dire”, “threatening”, and “war on ____” than sunshine and roses. The relationship between negativity and engagement motivates journalists and media to produce more negative and triggering content to push numbers. Additionally, negative, and morally linked terms can accelerate a certain bias, influencing public opinion. This is best seen through coverage of elections. For example, Fox News has recently come under fire for promoting claims unjustly undermining elections and creating a sense of fear and distrust in certain elected officials. In this example, Fox News used terms like “fraud”, “stolen election”, and many other extreme terms to describe elections and influence how their readers perceive politics. In another example, the researchers from Freaknomics’ Why is U.S. Media so Negative found that 87% of national U.S. covid media and journalism coverage was overwhelmingly negative while only 64% of scientific jorunals had the same negative tone on the same covid-19 topics. But why? What do media reporters have to gain by painting covid in a more negative and dire light than scientific journals? The answer is simple. This discrepancy arises from media motivations. The increase in negativity drives up media engagement, benefitting the media sources. Words like “catastrophe”, “troublesome”, and “appalling” catch people’s eye by instilling a certain amount of fear. This fear keeps us coming back. Checking that story or following that event. It gives us a false sense of control to read about the things that scare us. This fear becomes linked to the media creating a toxic relationship the media exploits to keep customers consuming information. Oftentimes the information is distorted or presented in a way that overly catastrophizes a certain issue. It cultivates a toxic relationship between the consumer and the media. The feelings of anxiety and fear cause people to monitor stories more closely, and the cycle repeats.

The term “doom-scrolling” was created to describe this toxic relationship between reader and news outlet. The National Institute of Health describes doom scrolling as “a vicious cycle in which users find themselves get stuck in a pattern of seeking negative information no matter how bad the news is.” I can’t tell you how many times I have opened the news app and found myself obsessing over all the negative news in the world. Sometimes it seems like hours go by and I am on one disasterous news article to the next. Research in to doom-scrolling finds that this type of behavior linked to negative news coverage is not only very common and but very hard to resist. Researches found that many people experience feelings of sadness, isolation, and fear after sessions of “doom-scrolling”. This is because the constant negative exposure and outlook creates those brain patterns in our minds. The negative terms influence to think about things more negatively. This can cause feelings of helplessness, catastrophizes, and making us feel isolated and less connected from the world we live in. Long term exposure to fear and anxiety can lead to levels of depression which can be exacerbated by doom-scrolling in many people.

Which brings us here: How bad is doom-scrolling, really, and what can be done. While not life-threatening, doom-scrolling is harmful to people and can seriously affect their outlook on life. I know I certainly feel feelings of hopelessness if I see too many sad news stories. Constant exposure to a skewed sample of negatively biased news can affect the way we think and interact with the world, increasing depression and overall sadness. While the effect of doom-scrolling varies by person, there are steps that can be taken to mitigate the negative mental health side effects associated with reading the media. Scientific American suggests placing time-limits of around 15 minutes on news scrolling can limit long term negative mental health consequences. Time limits allow our brains necessary breaks from overwhelmingly negative content and can prevent a vicious, doom-scrolling cycle.

Like many people, I want to stay informed about the world. I want to keep up with news cycles, vote informed, and know what’s going on around me and across the globe. But, like many people, I find myself disheartened after reading the news. I find myself clicking on stories motivated by fear and chaos rather than intellectual curiousity or wanting general knowledge. The negative bias in the media is something that can change with attention brought to the issue. The first step is recognizing the negaitve bias is present, and teaching our brains not to internalize the negativity through time limits and critical thinking about media motivations. With change, we can create a news information ecosystem with the right balance of necessary “bad-news” coverage with out manipulating readers with negative click-bait.

Media Negativity: Is it Harmful? was originally published in SI 410: Ethics and Information Technology on Medium, where people are continuing the conversation by highlighting and responding to this story.

In December 2016, a man named Edgar Maddison Welch walked into a pizzeria a short walk away from my house with a rifle and fired a shot. Thankfully, no one was hurt and he was quickly arrested, but many people in my neighborhood had been fearing that something like this would happen for a while. The reason Welch had driven hours from his North Carolina home to the Comet Ping Pong pizza restaurant in Washington DC was that he had become indoctrinated by an online conspiracy theory called Pizzagate, that alleged that the basement of the restaurant was home to a child sex trafficking ring which was controlled by prominent officials in the Democratic Party, including then-Presidential nominee Hillary Clinton. For months after the conspiracy theory emerged on fringe social media sites such as 4Chan, Comet and other small businesses adjacent to it had been receiving threats from unknown people who had found the theory online and became hooked in. It was only a matter of time before someone took matters into their own hands.

A few years later, a conspiracy theory loosely associated with Pizzagate, QAnon, inspired a mob of rioters to attack the US Capitol to attempt to stop the confirmation of President Biden’s victory in the 2020 election. In both cases, many of those who had become radicalized by the theories had interacted with them over social media websites including mainstream sites such as Facebook, Twitter, YouTube, and TikTok. A question that many raised particularly after the Capitol riot was what responsibility social media platforms had to remove false conspiracy theories from their platforms. While legally, as private companies, social media sites can decide to allow or ban anything posted on their sites, many platforms have historically been hesitant to remove politically-oriented content, even if it has been proven to be false and harmful. After all, many of those posting content about conspiracies such as QAnon may genuinely believe in them and are just expressing their opinions, and are not consciously trying to deceive anyone. However, this is a very narrow view of this issue, as ethical social media companies also have the responsibility to ensure that their platforms do not become hubs of misinformation and hate speech.

This is also compounded by the fact that many conspiracy theories do not emerge organically on social media sites and are spread by organized campaigns using misinformation tactics. For example, as investigated by the Media Manipulation Casebook, the “#SaveTheChildren” movement, which attempted to subliminally push QAnon and Pizzagate adjacent theories to people who might not otherwise believe them, had an organized group of influencers leading it that customized content to appeal to different groups, such as women and Black people. In addition, it tried deliberately to get around content bans through steps like changing around the hashtags. A coordinated disinformation campaign like this is very different then just any user expressing their opinion online, as it has a clear agenda and is being dishonest about its goals. Of course, one might point out that almost all forms of commercial advertising uses tactics similar to these to be effective, and advertisements have always been a large part of social media. However, unlike in disinformation campaigns, even the most subliminal forms of advertising generally have some identifying information about the company who produced it in the ad, and advertisements for products seen as harmful, such as tobacco, have government restrictions about how and where it can be advertised. There are no regulations about online media manipulation campaigns, and the only ones who can stop them, if they can even detect them in the first place, are social media platforms.

Even if social media companies try to act against coordinated manipulation campaigns, the more ethically complicated issues at stake related to misinformation on social media is how exactly misinformation should be defined and what actions, if any, should be taken against users who post misinformation but genuinely believe in it. In the first case, there is the fundamental epistemological problem that particularly with current events and political topics, there is no one fundamental “truth” that can be logically proven. However, at the same time, not all claims should be taken equally as often some claims have much more evidence pointing to them then others. In my opinion, the best way to classify whether something is misinformation is whether it has any credible evidence supporting it, as problematic as it can sometimes be to define what that means. Another issue, which was particularly illustrated during the COVID-19 pandemic, is sometimes information can change due to new evidence. For example, at the beginning of the pandemic US public health officials told the public not to wear masks because they did not see evidence they were effective against asymptomatic transmission and wanted to prioritize them for healthcare workers, but the guidance quickly changed as new evidence proving their effectiveness was discovered. This did not stop videos of public health officials giving the previous guidance from being widely shared out of context on social media, and even though these were real videos of authoritative figures giving what was at the time standard guidance, as soon as this guidance was outdated it became misinformation to share without giving context.

A similar dilemma happened regarding how social media companies treated various theories regarding the origin of the COVID-19 virus. Originally, there was no concrete evidence pointing to a lab leak or artificial origin to the pandemic, but due to the fact that Wuhan hosted one of the main research labs in China studying Coronaviruses, many social media users speculated that the pandemic could have started this way. In 2021, Facebook decided to mark this claim as disinformation and announced that posts promoting this theory would be removed from the site, but backtracked several months later as intelligence information started pointing towards a lab leak being at least plausible. In fact, by 2023 some US government agencies declared that a lab leak was in their opinion the most likely origin of the pandemic, although in low confidence and without direct proof. The initial suppression of this theory even though it eventually turned out to be more plausible angered conservatives who saw it as an overreach by social media companies against free speech and open debate.

The assertion that social media companies censor viewpoints for political reasons is actually widely believed among Americans, a 2022 Pew survey found that 77% of Americans believe that it is likely that social media platforms intentionally censor political viewpoints they disagree with, including 92% of Republicans and 66% of Democrats. Attempting to change this was one of the main reasons that Elon Musk decided to buy Twitter in 2022. Far from actually implementing true “free speech”, Musk has multiple times censored or interfered with content that criticizes himself and his companies. He has however reinstated many prominent accounts of users who were banned for hate speech or misinformation, such as Donald Trump and Andrew Tate. These changes led to a large spike in hate speech being posted on Twitter compared to the period before he bought it.

While Musk and other free-speech absolutists may rejoice at the new era of Twitter, it is impossible to ignore the fact that the hate speech and disinformation bans were working prior to their removal. Researchers from Zignal Labs found that after President Trump and other 2020 election disinformation influencers, content promoting lies relating to election fraud allegations were reduced by 73%, and there was also a significant drop in content celebrating the attack at the US Capitol. One big reason why that was true might have been that taking out the central actors of a disinformation network could have a disproportionate effect on the continued functioning of that network since oftentimes many simple users just follow what the big influencers are promoting. If this theory is true, it would mean that removing online disinformation and hate speech could in many cases be an easier problem than one might think, as just removing central figures will significantly reduce the amount of undesirable content.

Most importantly, there are also strong ethical reasons for social media platforms to act against hate speech. The examples I mentioned above show that in many cases it is difficult to define what these terms mean and that many people do not believe that social media networks currently do a good job with content moderation. However, I believe that even that is much better than doing nothing and allowing social media to become infested with hateful and false content. As the philosopher Karl Popper explained in his “Paradox of Tolerance” if we give complete tolerance to intolerant ideas, then eventually tolerance itself will be destroyed. Popper specified that he did not believe suppressing these ideas should be the first course of action, but if it becomes the case that intolerant ideas become a clear threat to society, then a tolerant society has not only a right but an obligation to fight against them in any way possible. In this spirit, as long as social media companies are being careful and more transparent than they have been so far about why they remove certain content, from an ethical perspective they must remove false and hateful content that has the ability to cause harm. It may help prevent the next mass shooting or act of political violence.

How far should social media platforms go in removing disinformation? was originally published in SI 410: Ethics and Information Technology on Medium, where people are continuing the conversation by highlighting and responding to this story.

I don’t think so. Tech companies have earned a reputation as a problem child for racially discriminatory hiring practices. They have, of course, created some real buzz over real acts of discrimination, to include a cyber-security firm posting an ad that asked for “Preferably Caucasian” applicants, or recent developments to hiring protocols that inadvertently create difficulty for disabled applicants. These are problems faced by the tech industry, but they don’t paint a full picture of the more common frustration that the demographics of the engineers who design and produce tech products are not representative of the population that those products serve. At its core, the problem that tech engineers do not represent the population starts with earlier cultural influences, and then manifests itself as a skewed adult workforce.

This problem begs three questions: What demographics do tech hiring practices represent? Whose responsibility is this lack of representation in tech? And what should be done to correct the demographic spread of our developers?

First, who are Big Tech’s targets to hire? If we compare data from the United States Census Bureau with Diversity, Equity, and Inclusion (DEI) Reports from Google, Amazon, Apple, and Facebook, it does not look like our tech firms balance these decisions well. We see in every circumstance a higher proportion of Asian employees and proportionate under-representation of every other racial group relative to the US population,, and a devastating lack of women in the workforce. In Amazon’s and Facebook’s cases, we do also see a disproportionate skew of white workers compared to black and Hispanic ones.

But tech giants don’t hire from the general American population, they hire from primarily the group of college graduates with degrees in computer, information, and data sciences. If we compare those companies’ reports with the demographics of graduates in the information technology field, we see a different picture. According to Data USA’s aggregation of degree data, almost 78% of graduates in the field are men. All four of these giants actually hire a higher proportion of women in their tech departments than graduate from university with the relevant degrees. In terms of race, Facebook leaves black and Hispanic representation in employment wanting in comparison to graduates. Google, Apple, and Amazon, on the other hand, all roughly match or beat minority representation in their employment though, which runs counter to the notion of a discriminatory trend throughout the industry.

It’s worth noting that international hiring skews corporate relations to American demographics. On top of that, representative hiring is a step from these companies towards creating representative organizations. While their hiring has been recently made representative, with years of under-hiring minorities, overall tech employment is not.

That’s good? Maybe. It’s good that companies are doing all they can given the hand they’ve been dealt, but why are we here in the first place? Where does it begin?

Let’s start with some speculation. What would be the consequences of the tech industry taking it upon themselves to represent the population? We could quickly see a total demographic match at massive companies. Meta, Google, Amazon, Intel, and the like see enough applicants that they could easily hire within national proportions for race, gender identity, etc.. However, the giants’ decision to match the population would catastrophically worsen the proportions of the remaining applicant pool. In the wake of representative giants, we would watch smaller companies skew even more overwhelmingly white, Asian, and male.

From an optimistic perspective, that kind of move by leaders in the tech industry could signal outward that the field is ready to accept diversity in the work force in an almost trickle-down fashion. But given the constitutional violations it would take to execute this, and the damage that would resonate through the industry, it isn’t quite reasonable to blame companies for hiring in relation to the applicant pool.

If we pass the buck from companies, we have two main agents to investigate for the disparate set of applicants that show up at company doorsteps. These are colleges and culture.

Starting at the next highest level we have colleges, where the problem of disproportionate demographics is clear as day. Every semester I start new classes, and am consistently a little bit surprised by how different my computer science classes look from my humanities classes. Not only do they lack diversity in the visibly apparent demographics like sex and race, but in categories like class and experiences we also all fall into similar buckets.

Colleges don’t overtly prevent anyone from joining the major of their choice, but there are less direct forces that could add preventative pressure from entering STEM majors. At the University of Michigan, the College of Engineering has a higher tuition cost than its liberal arts school. While many students take a computer science major through the liberal arts school, all CS students pay the higher rate.

This table from Pew’s research shows the additional costs for engineering programs within the Big 10 coalition of schools. Since 2017, that gap has increased for engineering. National research taken in 2022 shows that the average tuition for 4 year university (considering in-state, out-of-state, private, and public institutions) is just over $20,000 annually. For engineering, that cost jumps up to $32,000 a year. This is skewed by a large international student population (about 20%) in engineering, but overall represents a monetary barrier to entry.

According to the U.S. Bureau of Labor Statistics, the median annual income for Black, Hispanic, and Native American households fall $20–30,000 below that of White households, and $30–40,000 below Asian households. Those disparate incomes, just on a monetary basis, serve as a preemptive force for Black, Hispanic, and Native American students interested in engineering fields.

The cost prohibitions in place at the collegiate level provide a solid explanation for the racial discrepancies between the U.S. population and the tech industry, but not so much on the gender discrepancies. While women represent 60% of college students, they only represent about a quarter of students in computer science and information types of degree programs. Men and women don’t see nation-wide differences in class so greatly as members of different races, so that part of the blame lies elsewhere.

Ultimately, what the gender gap comes down to is a lack of cultural incentivization for women interested in joining technology-related career fields. The first chapter of Catherine D’Ignazio and Lauren Klein’s Data Feminism discusses the nuances of power within the tech industry. Here they define a notion of “privilege hazard,” wherein which a heterogenous group drives innovation, leaving developers blind to harms and biases present in their products. A secondary effect is that the responsibility to prove and remove these problems often falls onto members of the impacted minoritized groups.

Neither internet companies nor the internet itself are places where women hold much power. From my point of view, asking women to change the environment of the internet in organizations where they are dramatically underrepresented just doesn’t seem appealing. The culture of the industry is toxic to those not represented at the solutions table, and that toxicity hinders those unrepresented from entering the field.

According to the American Association of University Women, there are other cultural factors exclusionary towards women. There is a recursive problem of a lack of role models, and pervasive, wholly untrue stereotypes about women’s abilities to perform in these roles. Additionally, in a survey reported in Fortune, 85% of women reported maternity leave as a factor for exiting the industry. We can see reason after reason for women not to join the tech industry, so it’s no surprise that we have a 3:1 ratio of men to women.

So now we’re here. We have drastic disparity among the creators of the online space that skew its development towards groups who already carry undue power in American society. We can’t blame the hiring practices of giant tech companies because they actually do a pretty good job of matching the demographics of their applicants. We can blame some of the racial misrepresentation on systemic class discrimination between the costs of liberal arts degrees and engineering degrees; and we can blame a lot of the gender discrepancies on toxic power dynamics in the work environment. Where do we go?

Colleges nationwide have made a decent start with financial aid packages based on need, and government student loan assistance has followed suit. The problem of disproportionate racial representation is induced, at least in part, by disproportionate class opportunity when entering university. As such, the clearest solutions include equitable provisions along class lines. To level the representation of women in tech, America needs a cultural shift. The problems for women range from a life of stereotypes to bad options for parental leave to the nature of the platforms they’d work on. There’s no law to wave away misogyny, but we have to create an environment where women know they are welcome and have a chance at a long, fruitful career.

Are companies to blame for unrepresentative hiring? was originally published in SI 410: Ethics and Information Technology on Medium, where people are continuing the conversation by highlighting and responding to this story.

“People are suffering algorithmic harm, they’re not being told what’s happening to them, and there is no appeal system, there’s no accountability,” argues Cathy O’Neill, the author of the book Weapons of Math Destruction, in the Netflix documentary “Coded Bias”.

In the age of information, we, as consumers, have little knowledge of what goes behind the scenes. This makes it difficult to make value judgements on the technology we use. That certainly seems to be the case for the algorithms that generate art. More specifically, as you might be aware, there’s an ongoing debate about whether these algorithms are ethical.

Let me give you the bottom line up front: these algorithms can be unethical because they scrape and utilize data without the consent or knowledge of those who produced the data in the first place. In fact, there’s a whole system that enables these algorithms to take the data. Plus, there’s little to no opportunities for these individuals to take back their data.

At first glance, AI-generated art, or the algorithms themselves, seem harmless. You might have interacted with it through social media applications that turn a picture of your face into an anime-style portrait, or through wildly hypnotic landscapes that portray scenes from science fiction.

Yet, do we really know how it works? And if we do find out that there are harmful effects, who should be held accountable for these effects?

Those are some deep questions, but I’m going to be honest: I don’t think I’m happy with what I’ve found. The way in which the algorithms work and the question of accountability, or at least what we understand from it, troubles me, and I think it should trouble you too.

Let’s take a step back for a moment: What exactly is AI-generated art?

To put it simply, it’s any piece of art generated through machine-learning algorithms. These machines are self-learning, which means that they take large amounts of data from a dataset and use them to train itself. Today, most of the algorithms use generative adversarial networks (GANs). This means that there are two systems: one that takes the data and generates random products (like images), and another that judges the products and determines which ones best align with the dataset.

Here’s the key word: “dataset”. Where the data comes from matters a lot.

Does it, though? Why should I care?

Well, dear reader, as authors Catherine D’Ignazio and Lauren Klein point out, “Most data arrive on our computational doorstep context-free.” This issue matters most to artists whose art is used in the datasets or to people whose faces and bodies are included in the data without them knowing. And, this issue is particularly tricky when the dataset itself is composed of products that are supposed to be private or copyrighted.

Take the LAION-5B dataset, for example. Many images in the dataset were taken without the artists’ consent, and without crediting or compensating them. To add on to that, within the same dataset, an artist who goes by Lapine found photos from her private medical record. Lapine wasn’t aware of that her private medical photos were uploaded on the internet, let alone scraped into the dataset. But, to have your works or even your body be used in such a way without your knowledge is unfair. It breaches your privacy because when you copyright a work or when you go to the doctor’s office, you trust that the work or the results will be handled in a way that respects the value you assign to them. In Lapine’s case, she had signed a form refusing the practice the right to put the image in a dataset.

Lapine on Twitter: "🚩My face is in the #LAION dataset. In 2013 a doctor photographed my face as part of clinical documentation. He died in 2018 and somehow that image ended up somewhere online and then ended up in the dataset- the image that I signed a consent form for my doctor- not for a dataset. pic.twitter.com/TrvjdZtyjD / Twitter"

🚩My face is in the #LAION dataset. In 2013 a doctor photographed my face as part of clinical documentation. He died in 2018 and somehow that image ended up somewhere online and then ended up in the dataset- the image that I signed a consent form for my doctor- not for a dataset.

This is not to say that the creators of LAION-5B knew that Lapine’s image was in the database, or that they had any malicious intent. In fact, it’s very believable that the creators and the algorithm itself probably just gathered whatever the internet had to offer. The creators of LAION-5B have pointed out that the dataset is taken from the “publicly available internet” and is “uncurated”. However, it does seem a little sus [slang for “suspicious”]; yes, it’s clear that the dataset is too large to be filtered through, but to what extent can “publicly available” data be used for, well, everything? And who should take responsibility when data is used so unsparingly?

Let’s work down the chain of custody and see what we can infer.

Would the responsibility fall on the doctor’s office? Lapine suspects that the images were taken from the practice after the surgeon in charge passed away. Perhaps the onus should have been on the office to make sure that these images were kept within the practice. The surgeon’s passing may have created some confusion about what to do with the files, but there were ways to prevent the incident. These include deleting the data, working with the patients to put the data elsewhere, or letting the patients know that the data had been leaked. To be fair, there is no sure way to keep the data safe 100% of the time. It could also have been the case that the office didn’t know that the images were leaked.

So, if the responsibility can’t be entirely held by the office — should it also fall on LAION-5B instead?

On LAION’s public Discord server, Romain Beaumont, an engineer, pointed out that LAION itself doesn’t host the images, but only takes from certain websites that does host those images. Beaumont also suggested that individuals should “ask for the hosting website to stop hosting it” and that a blacklist of host websites should be created.

To be honest, this sounds like LAION is shifting the blame and responsibility for taking these images off the web. Who would expect a picture of your medical condition to be public online? Plus, not everyone knows who or how to ask to get the images removed.

At the same time, though, let’s be honest: sifting through the internet would be a horrible task.

Perhaps we could turn to new research about unlocking where the algorithm takes its data from. A research group from Nvidia tried running the neural network in reverse, and were successful in producing some of the training data images. But, as the article naming Nvidia points out, this reverse engineering process is still in the works; this makes it hard for companies, let alone regular users, to access it.

And here’s something else to ✨spice up✨ the problem: AI-generating platforms cannot hold copyright. This means that the art produced by these platforms belong in the public domain. In other words, no matter where the AI gets its data, the end product doesn’t belong to anybody in particular. This creates a huge problem: artists can’t get credit for any products generated based on their art. This issue is especially frustrating to artists who find their art everywhere, even if the art is copyrighted. Some artists have even found their mangled signatures on AI art, which makes the link between the source data and the products clearer.

Even then, however, there’s not much the artists can do to control how their art is used, because it’s already in the hands of the users. And users have taken their liberties with the products: one AI piece won a prize at a state art fair, while other pieces have been sold as NFTs.

So, here’s what we’ve got so far: the surgeon’s office, the websites hosting the images, and LAION-5B are all problematic in their own ways. At each of these stages, however, there’s not a lot of accountability being held.

But what about us? Do we have a responsibility too?

If we take a view based on the book Data Feminism, maybe the users do have a responsibility. Authors D’Ignazio and Klein argue that users should “ask questions about the social, cultural, historical, institutional, and material conditions under which that knowledge was produced”. Perhaps LAION-5B should put up a warning telling users that the products created by the algorithm could be derived without the creator’s consent.

Okay, so… what can I learn from this mess?

Well, dear reader, I wanted to share that the whole system is full of ethical conundrums, especially when it comes to where the data comes from. There’s little to no accountability at each stage, hurting artists and patients like Lapine. No one wants, or can, take the full responsibility for making sure that data is ethically sourced or used. It’s quite dismal, really.

But, as I’ve suggested, there is some hope. At least you, dear reader, know that art-generating algorithms take from the unsuspecting, and that there are always a few ways to stop that from happening. And we can always work to improve the system.

Contests of Context: The Problems with Art-Generating Artificial Intelligence was originally published in SI 410: Ethics and Information Technology on Medium, where people are continuing the conversation by highlighting and responding to this story.