Artificial intelligence is increasingly consuming its own outputs, creating significant risks for the quality and reliability of future models. The rapid growth of AI has outpaced the availability of high-quality, human-generated data, forcing AI systems to increasingly rely on AI-generated content. This practice, while seemingly a practical solution to the data shortage, introduces serious problems, including the degradation of model accuracy and the potential for what experts are calling “model collapse.”

The Feedback Loop Problem

The core issue lies in the feedback loop created when AI systems ingest data that was originally produced by other AI systems. This process undermines the integrity of the AI’s learning capabilities. When AI models are trained on outputs that are not rooted in reality, the risk of their outputs becoming increasingly distorted and less reliable grows exponentially. This is not merely a technical glitch but a systemic flaw that could undermine the entire AI ecosystem if not addressed.

The New York Times has highlighted how this self-referential training can cause AI outputs to drift away from reality. This drift occurs because the AI is no longer learning from a diverse and accurate dataset but from a pool of content that may already be biased, incomplete, or outright erroneous. As AI continues to generate more of its own training data, the risk of amplifying these inaccuracies grows, leading to models that are less effective and more prone to producing flawed outputs.

Model Collapse: A Real and Present Danger

This problem is compounded by the sheer scale at which AI is being deployed. AI-generated content is now flooding the internet, filling websites, social media platforms, and even news outlets. Scientific American reports that large language models are generating vast amounts of text, which is increasingly indistinguishable from content created by humans. This saturation of synthetic content not only makes it harder to find reliable human-generated data but also increases the likelihood that future AI models will be trained on inferior data, further exacerbating the problem.

One of the most alarming consequences of this trend is the potential for “model collapse,” a scenario where AI models become so detached from reality that their outputs are no longer usable. This collapse is not just a theoretical risk; it is already being observed in some AI systems that produce outputs filled with biases, inaccuracies, and absurdities. These flawed outputs are then fed back into the training process, creating a vicious cycle that degrades the model’s performance over time.

Synthetic Data: Necessity or Risk?

AI companies are aware of these risks but are often left with few alternatives. As The Atlantic notes, the demand for more advanced AI models is pushing developers to use whatever data is available, including AI-generated content. The difficulty of distinguishing between human-generated and synthetic data means that even the most well-intentioned efforts to maintain data quality are likely to fall short. This situation is driving a reliance on potentially flawed training material, which could have far-reaching implications for businesses and consumers.

Despite these challenges, some experts argue that synthetic data is not inherently bad. There are specific scenarios where AI-generated content can be useful, such as in training smaller models or in situations where the accuracy of the output can be easily verified. However, these instances are the exception rather than the rule. The broader use of AI-generated content in training large models poses significant risks that cannot be overlooked.

Impact on Internet Integrity and Business Decisions

The proliferation of AI-generated content also raises broader concerns about the integrity of the internet as a whole. The dead internet theory, which suggests that much of the internet’s content is now generated by bots and AI rather than humans, is gaining traction. While this theory is still speculative, it reflects a growing unease about the direction in which AI is taking the digital landscape. If AI continues to dominate content creation, the internet could become less a reflection of human knowledge and creativity and more a repository of synthetic, machine-generated data.

This shift has significant implications for businesses that rely on the internet for information, customer engagement, and brand management. As AI-generated content becomes more prevalent, companies may find it increasingly difficult to trust the data they are using to make decisions. The risk of basing business strategies on flawed or biased information grows as the quality of online content declines. This could lead to poor decision-making, reduced competitiveness, and ultimately, a loss of consumer trust.

The Dead Internet Theory: More than a Conspiracy?

The dead internet theory also touches on deeper fears about the role of AI in shaping public discourse and influencing political outcomes. Some experts, like Jake Renzella and Vlada Rozova, have warned that AI-generated content could be used to support autocratic regimes, spread propaganda, and manipulate public opinion. While these concerns may sound alarmist, they are rooted in the growing influence of AI on the flow of information and the potential for AI to be used as a tool for social and political control.

Fortunately, there is evidence that the dead internet theory has not yet fully materialized. Forbes reports that the vast majority of viral content—such as provocative opinions, clever observations, and creative reinterpretations—are still generated by humans. However, the growing presence of AI-generated content on the internet cannot be ignored, and businesses must remain vigilant to the risks posed by AI’s self-consumption problem.

Addressing the Risks: Business Strategies for Safer AI

The issue of AI’s self-cannibalisation is not just a technical challenge but a strategic one that businesses must address head-on. Companies that are deploying AI systems need to be aware of the risks associated with using AI-generated content for training. They must invest in rigorous data governance practices to ensure that their models are trained on accurate, high-quality data. Failure to do so could result in AI systems that are not only ineffective but also potentially damaging to the business.

In conclusion, the practice of using AI-generated content to train new AI models is creating a significant risk of model collapse and degraded performance. As AI continues to grow in importance across industries, businesses must take proactive steps to address these risks. By prioritising data quality and investing in robust training practices, companies can mitigate the dangers posed by AI’s self-cannibalisation and ensure that their AI systems remain reliable, effective, and aligned with reality.