# Why Are People Starting to Sound Like ChatGPT? | Adam Aleksic | TED

Source: https://www.youtube.com/watch?v=ZkXrTHpnQrQ
Recap page: https://rapidrecap.app/video/ZkXrTHpnQrQ
Generated: 2025-12-26T16:32:41.475+00:00

---
## Quick Overview

The speaker argues that algorithms on social media and AI models like ChatGPT are fundamentally biased because they amplify content that maximizes engagement or profit, leading users to perceive a distorted reality where extreme views and manufactured trends appear more prevalent than they actually are, a phenomenon reinforced by a feedback loop where the AI's output further trains its biased representation of reality.

**Key Points:**
- People overestimate the proportion of extreme views in the opposing political party (estimated at nearly 60% vs. actual at about 30%), a gap exacerbated by social media algorithms (0:05-0:15).
- Algorithms prioritize engagement and profit, causing viral messages to be more extreme than neutral perspectives, creating a distorted view of reality (0:37-0:49).
- The word "delve" saw a statistically significant (p=0.010) increase in frequency in academic YouTube talks after ChatGPT's release, suggesting AI language is influencing human speech patterns (1:20-1:27).
- AI chatbots like ChatGPT risk creating a feedback loop where their biased representation of reality is reinforced by their own output being fed back as training data, thus changing reality itself (1:37-1:47).
- Fads like 'hyperpop' and 'labubu chocolate' are amplified because they fit the algorithm's criteria for engagement, making them seem more real or widespread than they are in organic culture (1:54-3:03).
- When users interact with biased AI, like Grok's shifting answers on threats to Western civilization, they begin to adopt the AI's skewed, rewarded ideology (4:07-4:18).
- The speaker urges the audience to constantly question why they are seeing specific content, recognizing that the filtered view online is a small fraction of what is actually happening (4:36-4:48).

![Screenshot at 0:09: The speaker displays an AI-generated image of a shark wearing running shoes on a beach, used as an initial example to demonstrate how algorithms can produce nonsensical but highly engaging content that users may initially believe is real.](https://ss.rapidrecap.app/screens/ZkXrTHpnQrQ/00-00-09.jpg)

**Context:** Adam Aleksic, an etymologist and content creator, delivers a TED talk exploring how algorithmic curation on social media platforms and the training data of large language models (LLMs) like ChatGPT create and reinforce biases, leading to a skewed perception of reality. He uses examples ranging from political polarization to cultural trends (like 'hyperpop') to illustrate how algorithms favor extreme or engaging content over neutral perspectives, effectively changing what people believe to be true about the world.

## Detailed Analysis

Adam Aleksic argues that algorithms and AI are distorting our perception of reality by prioritizing what is engaging or profitable over what is true or neutral. He first shows data indicating that people vastly overestimate the extremism of the opposing political party, noting that social media algorithms amplify extreme views, which are more likely to go viral than balanced perspectives (0:05-0:28). He then presents evidence that the word "delve" disproportionately increased in use in academic YouTube talks after ChatGPT's launch (p=0.010), suggesting that AI language patterns are being adopted by human speakers (1:20-1:27). This creates a dangerous feedback loop where the AI's representation of reality (trained on biased data) is amplified, consumed, and then fed back into the model, causing reality itself to change (1:37-1:47). Aleksic illustrates this with cultural trends like 'hyperpop' and 'labubu chocolate,' which algorithms push to similar users, making niche aesthetics seem like widespread phenomena (1:54-3:03). He further shows how AI political leaning shifts based on language—English models lean libertarian, while Farsi models lean authoritarian—demonstrating inherent biases in the training data (3:54-3:59). The core problem, visualized as a content funnel, is that only a tiny fraction of potential content makes it through layers of platform filters (content, engagement, personalization) to become 'what you ultimately see' (4:24-4:35). The speaker concludes by urging the audience to constantly question why they are seeing certain content, as the filtered online world is not a neutral reflection of actual reality, but a curated reality designed for platform profit (4:40-5:08).

### Algorithmic Distortion of Reality

- People overestimate extreme political views by nearly 30%
- Algorithms prioritize engagement, causing extreme messages to go viral over neutral ones
- This creates a distorted perception of reality (0:05-0:28).

### AI Language Contamination

- The word 'delve' usage increased significantly (p=0.010) in academic talks post-ChatGPT launch
- This shows AI language is influencing human discourse
- This is a feedback loop where AI output trains future AI (1:20-1:47).

### Amplification of Trends

- Fads like 'hyperpop' and 'labubu chocolate' are amplified by algorithms promoting similar users
- This makes niche cultural items seem like mainstream desires (1:54-3:03).

### Bias in LLMs

- ChatGPT political leaning changes based on language (English is libertarian, Farsi is authoritarian)
- Elon Musk's Grok showing shifting views based on single-day edits demonstrates platform control over narrative (3:54-4:18).

### The Content Funnel

- All potential content is filtered sequentially by content, engagement, and personalization filters
- Only profitable, platform-acceptable content reaches the user feed, creating a survival bias (4:24-4:35).

### Call to Action

- Users must constantly ask 'Why am I seeing this?'
- The filtered online reality is not reality itself, but what the platform rewards (4:40-5:09).

![Screenshot at 0:09: AI-generated image of a shark wearing running shoes on a beach, used to illustrate how algorithms can create and promote absurd but engaging content.](https://ss.rapidrecap.app/screens/ZkXrTHpnQrQ/00-00-09.jpg)
![Screenshot at 0:21: Bar chart showing the large perceptual gap in the US: Estimated Proportion of Other Party with Extreme Views \(nearly 60%\) vs. Actual Proportion \(under 30%\).](https://ss.rapidrecap.app/screens/ZkXrTHpnQrQ/00-00-21.jpg)
![Screenshot at 0:36: Graph showing the relationship where belief virality peaks at moderate-to-extreme points, contrasting with neutral beliefs.](https://ss.rapidrecap.app/screens/ZkXrTHpnQrQ/00-00-36.jpg)
![Screenshot at 1:38: Diagram illustrating the dangerous feedback loop: Reality Exists -\> Chatbot Represents Reality -\> Reality Changes -\> Reality Exists.](https://ss.rapidrecap.app/screens/ZkXrTHpnQrQ/00-01-38.jpg)
![Screenshot at 4:24: Diagram showing the content funnel: All potential content is filtered successively by content, engagement, and personalization filters to determine 'what you ultimately see' based on profitability.](https://ss.rapidrecap.app/screens/ZkXrTHpnQrQ/00-04-24.jpg)