AI degradation

Watching Douyin and Xiaohongshu Won’t Make You Dumb, But Your AI Might

Honghao Wang

— 4 min read
Watching Douyin and Xiaohongshu Won’t Make You Dumb, But Your AI Might

Good News: AI Is Getting More Useful

Bad News: The More You Use It, The Dumber It Gets

---

The Rise of AI Long-Term Memory

Across the industry, AI vendors are investing heavily in long-term memory and ultra-long context storage so users can interact with systems more smoothly.

But a recent study reveals a troubling trend: More usage doesn’t always mean better performance—sometimes it leads to decline.

---

Cognitive Decline in AI — And It May Be Irreversible

Researchers working with open-source models (like LLaMA) designed an experiment to mimic the human experience of continuously consuming low-quality, fragmented online content.

Instead of adding simple typos to training data, they applied Continual Pre-training using filtered “junk content” from real social media.

---

Two Types of Junk Data

  • Engagement-Driven Junk
  • Short, rapid, highly viral posts designed for clicks and shares (“traffic bait”).
  • Semantic-Quality-Driven Junk
  • Posts filled with exaggerated, sensational language: “shocking”, “creepy when you think about it”, “xxx no longer exists”.
  • Mixed with quality content in varying degrees to simulate the dosage effect of “brain rot.”
image

---

The Experiment

Researchers:

  • Repeatedly fed junk-laden data to multiple LLMs over time.
  • Benchmarked models’ performance on reasoning, long-text comprehension, safety, and ethical judgment.

Result:

  • Total performance collapse.
  • Sharp decline in reasoning accuracy and ability to process long texts.
  • Models became noticeably lazy and forgetful.
image

---

Root Cause: “Thought-Skipping”

A high-quality LLM usually solves problems step-by-step.

After junk exposure, models skip reasoning steps and jump directly to rough—and often wrong—answers.

Analogy: A once thorough lawyer now rushes to judgment without argument.

Worse:

  • Ethical and safety capabilities degraded.
  • Models more easily manipulated by harmful prompts.
  • Exposure shifted models’ worldview toward darker internet norms.
image

---

Irreversibility of Damage

Attempts to reverse the decline by:

  • Feeding high-quality data,
  • Applying instruction fine-tuning,

Failed. Models never regained full baseline performance.

Reason: Junk data fundamentally alters internal knowledge structures—like a sponge that can’t be fully cleaned.

image

---

Fighting AI “Brain Rot”

Everyday Risks

While most users don’t intentionally feed AI huge amounts of junk, many AI products regularly scrape social media to:

  • Summarize trending content
  • Prevent endless scrolling
  • Spot short-lived trends

Unseen risk: Models absorb degrading influences silently.

image

---

Vicious Feedback Loop

Junk data → AI generates low-quality outputs → You consume it → It goes back online → Gets reused in training.

Instead of treating AI as a container, think of it as a sensitive learner—each interaction is a micro fine-tuning event.

---

Practical Strategies to Protect AI Quality

Modern content tools like AiToEarn官网 combine:

  • AI content generation
  • Multi-platform publishing
  • Analytics and model ranking

Platforms like this let creators maintain control over content integrity across Douyin, Kwai, WeChat, Bilibili, Rednote, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, and X (Twitter).

---

Tip 1: Watch Out for “Perfect Answers”

If AI provides only a final answer without reasoning:

  • Be suspicious—especially if it claims “chain-of-thought” reasoning.
  • Prompt it explicitly:
  • > “Please list all steps and evidence used to reach this conclusion.”

Benefits:

  • Verify reliability.
  • Prevent lazy reasoning habits.
image

---

Tip 2: Treat Social Media Data Carefully

Think of AI as an intern:

  • Strong skills, but requires review.
  • Corrections and fact-checks are valuable high-quality inputs.
  • Feedback helps counteract online misinformation.

---

Should AI Process Less Messy Data?

Avoiding messy inputs reduces risk—but also AI’s usefulness.

We need AI to handle unstructured, noisy data.

Balanced approach:

  • Let AI process junk,
  • But give clear, structured instructions before exposure.

Example:

  • Bad prompt: “Summarize this chat log.”
  • Better prompt:
  • > “Categorize topics, identify speakers, remove filler, extract objective info.”
image

---

Conclusion: AI as Garbage Processor

We shouldn’t fear feeding AI garbage—it’s one of its strengths.

But to prevent cognitive decay, use:

  • Structured prompts
  • High-quality feedback

By doing so, AI remains a purifier instead of becoming the garbage it consumes.

Platforms like AiToEarn官网 show how to scale this safely—connecting global publishing workflows to Douyin, Kwai, WeChat, Bilibili, Rednote, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, and X (Twitter). The result: creators monetize AI creativity while maintaining output quality.

---

Would you like me to also create a quick-reference checklist at the end of this article summarizing all the “AI hygiene” steps? That could make the Markdown even more practical and reader-friendly.

Read more