AI INBREEDING – Is this the cause for AI Hallucinations?

My theory is so factual it’s not even a theory anymore.
Programmed schizophrenia is a thing and it spreads to people.

All the sensational and catastrophic headlines about AI are nothing more than disguised advertising for a new financial bubble like they have been and will be, because we have learned nothing from the dot-com bubble, for example. Investments are attracted from functionally and scientifically illiterate billionaires, but also from governments, which administer our money. And when the blatant lie is about to become too cosmetized to be embellished anymore, the whole crew secretly abandons the Titanic, leaving the populace to sink.

Apart from the computational speed, which is expensive, by the way, AI brings nothing new to the market. Success is based, as usual, on two main pillars: clever PR + colorful packaging.

What no one discusses, however:
In recent years, product degradation has been beating development.

What is the “Dead Internet”?

The term “Dead Internet” refers to an emerging theory that a significant proportion of today’s online content is no longer generated by real people, but by artificial intelligences, bots, or automated algorithms. This dynamic produces a form of digital self-canonization in which AIs not only contribute to online content, but also begin to consume it, recycling and amplifying errors and distorted information.

The phenomenon raises serious concerns about the quality, veracity, and epistemic diversity of the information that underpins language models and AI systems. Instead of reflecting human reality and the complexity of knowledge, AIs may end up reflecting only their own distortions—a form of digital cognitive incest.

Mechanisms of AI Inbreeding and Hallucinations

The “Dead Internet Theory” posits that a significant portion of online content is now generated by bots and AI, leading to a decline in genuine human interaction and authentic content creation. This shift has profound implications for AI development, particularly concerning the phenomenon of AI hallucinations—instances where AI models produce outputs that are plausible-sounding but factually incorrect or nonsensical.

As AI-generated content proliferates, it increasingly populates the datasets used to train new AI models. This recursive cycle—where AI learns from AI-generated data—can lead to a degradation of model performance, a process sometimes referred to as “model collapse” or “AI inbreeding” .

Model Collapse and AI Slop

“Model collapse” describes the degradation of AI model performance due to training on synthetic data. Initially, models may lose the ability to handle rare or nuanced inputs (early collapse), eventually leading to widespread inaccuracies (late collapse) .

“AI slop” refers to the accumulation of low-quality, AI-generated content that lacks diversity and richness. Training on such content can cause models to produce outputs that are increasingly detached from reality .

The Feedback Loop

The cycle of AI generating content that is then used to train other AI models creates a feedback loop. Over time, this loop can amplify errors and reduce the overall quality and reliability of AI outputs .

Implications and Mitigation Strategies

Risks of Continued AI Inbreeding

If unchecked, the trend of AI models learning from AI-generated content could lead to a sharper decline in the accuracy and usefulness of AI systems to the point of total uselessness. This degradation poses risks across various sectors, including healthcare, finance, and information dissemination.

Consequences and Warnings

• Erosion of Reality: If AIs dominate online information and other AIs are trained on this basis, a downward spiral of “simulated reality” occurs.
• Research Issues: AI models become less and less useful for research, journalism, or education because they lose touch with authentic human sources.
• The Need for Clean Data: It is vital that AI models are trained on verified, human, diverse, and contextualized data, not just the digital noise of an increasingly artificial web.

Strategies for Mitigation

To counteract these risks, several strategies can be employed:

• More Human Oversight: Ensuring that training datasets include diverse, high-quality human-generated content can help maintain model robustness.
• Implement Data Filtering: Developing methods to identify and exclude low-quality AI-generated content from training datasets.
• Promote Transparency: Encouraging openness about data sources and training methodologies to allow for better assessment and trust in AI systems.

---

USEFUL RESOURCES:

* Edwards, B. (2023). *AI Is Now a Misinformation Machine*. Ars Technica.

* Shulman, S. (2022). *The Dead Internet Theory Isn’t Just a Conspiracy Anymore*. WIRED.

* Marcus, G. & Davis, E. (2020). *Rebooting AI: Building Artificial Intelligence We Can Trust*.

* Vincent, J. (2023). *How LLMs Go Wrong*. The Verge.

* Internet Archive & Common Crawl Dataset Analysis (2022–2024).

To be continued?
Our work and existence, as media and people, is funded solely by our most generous supporters. But we’re not really covering our costs so far, and we’re in dire needs to upgrade our equipment, especially for video production.
Help SILVIEW.media survive and grow, please donate here, anything helps. Thank you!

! Articles can always be subject of later editing as a way of perfecting them