Welcome back, AI prodigies!
In today’s Sunday Special:
📜The Prelude
🗣️Fighting Harm vs. Free Expression?
🤖How AI Moderates Content
🔑Key Takeaway
Read Time: 7 minutes
🎓Key Terms
Acoustic Model (AM): Listens to speech and identifies which sounds correspond to which words.
Machine Learning (ML): Leverages data to recognize patterns and make predictions without being explicitly programmed to do so.
Convolutional Neural Networks (CNNs): A network of specialized filters that detect edges, corners, and textures across images and videos.
🩺 PULSE CHECK
Should social media platforms be responsible for moderating content?
📜THE PRELUDE
Imagine being paid to watch viral clips and trending videos all day. But, there’s a catch: you also see the internet’s unfiltered underside. The cruelty packaged for clicks and the suffering reduced to pixels.
The first video shows a slave auction in Libya. The second video depicts a terrified family trapped in a collapsing building. The third video portrays cannibals cooking a dismembered calf. You’re not angry. You’re not shocked. You’re not disturbed. You’re emotionally numb. In a single workday, you just watched hundreds of videos capturing the absolute worst of humanity.
It’s hard to imagine, but more than 100,000 content moderators face this digital nightmare every day. No wonder about 54% of them exhibit PTSD symptoms, with roughly 20% experiencing severe physiological stress indistinguishable from combat veterans.
Over the past decade, AI has helped lighten the load for content moderators by automatically flagging, freezing, and filtering harmful media. So, how do social media platforms decide what stays and what goes? How does AI separate the harmful signals from the social noise?
🗣️FIGHTING HARM VS. FREE EXPRESSION?
⦿ 1️⃣ Too Much OR Too Little?
All social media platforms walk a tightrope: too much censorship or too much free speech. Striking the right balance between content moderation and freedom of speech remains vital for a thriving democratic society. When social media platforms get this balance wrong, the consequences can be catastrophic, ranging from enabling genocides to interfering with democracy.
⦿ 2️⃣ The Rohingya Genocide?
Facebook’s content moderation team failed to halt the hate that fueled the Rohingya genocide. The Rohingya people are a stateless, Indo-Aryan ethnolinguistic group who predominantly follow Islam. They’ve faced ongoing persecution since the 1960s in Myanmar for their cultural identity and religious beliefs.
On August 25th, 2017, the ARSA, an insurgent militia that resorts to armed resistance, launched coordinated attacks on police posts and reportedly killed 12 security personnel. Ruthless military dictator Min Aung Hlaing immediately posted the following on Facebook: “We openly declare that absolutely, our country has no Rohingya race.” This powerful statement triggered a military-linked propaganda network that spread across the social media platform, reaching over half the country.
In the following months, approximately 288 Rohingya villages were burned, 6,700 Rohingya people were slaughtered, and 500,000 Rohingya people fled the country. Meta CEO Mark Zuckerberg eventually admitted Facebook was weaponized to incite violence against the Rohingya people. Facebook hasn’t faced any legal justice or financial penalties for the algorithmic amplification that supported an ethnic cleansing.
⦿ 3️⃣ Erasing Criminal Evidence of War Crimes?
Between 2012 and 2018, Syrian Archive, a non-profit initiative that aims to document human rights violations across Syria, collected nearly 1.2 million videos capturing everything from civilians suffering chemical attacks to targeted airstrikes on hospitals. Most were uploaded to YouTube, serving as crucial evidence for ongoing war crimes.
Despite this critical evidentiary value, YouTube removed over 120,000 videos documenting these atrocities, citing violations of community guidelines against “sensitive content.” This decision significantly restricted public access to evidence of crimes against humanity. Several human rights groups warned that the removal of these recordings threatened the integrity of international justice processes.
Syrian Archive insisted these videos weren’t “gratuitously violent” but documentary evidence of human rights violations. A YouTube spokesperson later admitted: “With the massive volume of videos, sometimes we make the wrong call.” Syrian Archive was able to recover thousands of aggressively flagged videos documenting violent extremists. Still, the reinstatement of videos depicting war crimes across Syria has been inconsistent. For example, the VDC YouTube channel, which housed over 32,000 video interviews of local Syrian residents describing the bombings, was suddenly terminated without notice.
🤖HOW AI MODERATES CONTENT
⦿ 4️⃣ Reducing Errors and Biases?
Most content moderators work under intense pressure, where speed and accuracy are imperative. They must make split-second decisions on complex and nuanced content, all while striving to maintain consistency. This relentless combination of time, pressure, and the need for careful judgment triggers a perfect storm for stress and burnout, often resulting in massive mistakes.
AI excels at leveraging historical content moderation patterns to reduce subjective biases and consistently apply community guidelines and policy regulations across the roughly 1.1 billion social media posts uploaded daily.
⦿ 5️⃣ Example: The Group Harassment Case.
Imagine someone uploads a short clip showing a person on the ground being verbally attacked with racial slurs and physically assaulted with kicks. The caption states: “Here’s what happens when these people don’t learn! #LessonLearned.”
Here’s how today’s modern AI moderation systems tackle this short clip:
🔵 Classify the Caption:
First, they convert captions into tokens: a word or part of a word. For example, “unbelievable” might be broken into four tokens: “un,” “bel,” “iev,” “able.”
Second, these tokens are transformed into embeddings: a list of numbers that machines can understand. For example, the concept of “these people” is represented by numbers like: “{0.23, -1.51, 0.83, 0.07,....}.”
🟠 Interpret the Video:
First, the CNN is given video frames. These video frames appear as grids of numbers (i.e., “pixels”), where each number represents a color or brightness level.
Second, the CNN uses small filters (i.e., think of them as “little windows”) that slide over the video frames. Each small filter looks for specific shapes, like edges, corners, or textures, where colors in the video change sharply.
Third, the CNN highlights specific shapes that seem important for identifying specific objects. It achieves this by repeatedly detecting statistical patterns in pixels associated with a prone body and bloody spots. Then, it outputs a high violence probability (e.g., P = 0.79).
🟡 Interpret the Audio:
Second, the AM analyzes the layered auditory signals themselves, such as high-pitched yelling and low-frequency thumping.
🔑KEY TAKEAWAY
All social media platforms tread a razor-thin line between censorship and chaos. AI is increasingly being deployed to act as a safety net that catches hate and harassment before humans have to. The scariest part?! Deciding who trains the AI on what counts as hateful.
📒FINAL NOTE
FEEDBACK
How would you rate today’s email?
❤️TAIP Review of The Week
“These super-sized editions on Sunday continue to be excellent reads, Thanks!”
REFER & EARN
🎉Your Friends Learn, You Earn!
{{rp_personalized_text}}
Share your unique referral link: {{rp_refer_url}}
