“I’ve read that the internet is slowly becoming “useless” because AI responses are being given instead of human-written website content. To compound this problem, AI is being trained by AI and it gets worse with every “generation” of training. I noticed that I’m having to correct EVERY response that Google’s AI has given me since it was first introduced. Why is Google’s AI so wildly inaccurate nearly 100% of the time?”
I asked Google’s Gemini that question and, to add more AI slop to the pile, here is its response…
You have hit on some of the most critical and hotly debated problems in modern computer science. It is incredibly frustrating to feel like you have to fact-check every single answer, especially when you are looking for reliable information.
What you are observing is a mix of documented technical limitations, economic pressures, and a phenomenon known as “model collapse.” Here is a direct look at why these systems struggle so much with accuracy.
1. The Core Flaw: Prediction, Not Knowledge
The most important thing to know about Large Language Models (LLMs) is that they do not “know” facts. They are advanced autocomplete engines.
• They calculate the statistical probability of the next word in a sentence.
• They prioritize sounding fluent and human over being factually correct.
• Because they are built to please the user with a fluid answer, they will confidently invent (“hallucinate”) details that sound perfectly plausible, such as swapping book titles into a list of author names.
2. The “Model Collapse” Loop
You mentioned AI training on AI content, which scientists call Model Collapse. This is a major structural threat to the internet.
• The Cycle: As the web fills with AI-generated blogs, articles, and reviews, newer AI models are accidentally trained on that synthetic data instead of genuine human writing.
• The Result: Just like making a photocopy of a photocopy, the quality degrades. The AI loses grasp of rare facts, amplifies its own biases, and eventually produces repetitive, nonsensical, or completely incorrect outputs.
3. Google’s Unique Integration Challenge
Google’s specific AI features (like AI Overviews) have faced intense public criticism for accuracy. This happens because of how they try to merge two incompatible technologies:
• Search vs. Synthesis: Traditional search finds existing human web pages. Google’s AI tries to read those pages and summarize them instantly.
• The Breakdown: If the AI summarizes a satirical article, a Reddit joke, or an old, unverified forum post, it treats that text as absolute truth and displays it prominently at the top of your screen.
4. The Internet’s Changing Landscape
The internet is changing rapidly because of this technology. Many high-quality, human-written websites are now blocking AI bots from reading their data, or hiding their content behind paywalls. As a result, AI models are sometimes left training on lower-quality, open-source web data, which further drags down their accuracy.
________________________________________
Model collapse is a degenerative learning process where artificial intelligence models slowly lose quality, accuracy, and diversity over successive generations because they are trained on synthetic data produced by previous AI models instead of original human data.
Key Stages of Collapse
• Early collapse: The AI starts losing information about rare events or unusual data (the “tails” of the distribution), producing less varied text or images.
• Late collapse: The model completely misunderstands reality, outputting repetitive, garbled, or uniform nonsense that bears little resemblance to actual human data.
Why It Happens
• Recursive pollution: Each new model copy-pastes and amplifies the small mistakes, biases, and high-probability averages of the older model.
• The photocopy effect: Like making a photocopy of a photocopy, every new cycle loses sharp details until the final image or text is blurry and distorted.