risk · R10
Model Collapse and Data Scarcity Crisis
As AI-generated content floods the internet, models trained on this synthetic data are showing progressive quality degradation—'model collapse'—while high-quality human-generated training data is projected to be exhausted by 2028.
Updated
Observatory assessment: 68 / 100. High
Deep dive
A fundamental paradox threatens AI's continued improvement: the more successful AI content generation becomes, the more it poisons its own future training data. Research on 'model collapse' is now visibly manifesting as a flood of low-quality 'AI slop' across the internet, from social media to dedicated streaming channels, prompting public backlash.
Simultaneously, the supply of high-quality human-generated training data is finite and diminishing. Epoch AI's estimate of exhaustion by 2026-2028 appears increasingly prescient, as the 'data wall' is now described as a 'data bottleneck' slowing progress in visual and physical AI, not just text models. The crisis is driving extreme acquisition tactics, with reports of AI firms buying and destroying rare books to secure unique data. While responses like synthetic data and licensing deals (Reddit-Google, News Corp-OpenAI) are underway, they face their own limitations and cannot fully resolve the core scarcity problem.
Trajectory and projections
Near term
By 2026, major AI labs acknowledge diminishing returns from scaling alone. Data licensing costs exceed compute costs for some training runs.
Mid term
By 2030, 'Data Commons' initiatives create curated, verified human datasets. Synthetic data techniques mature but require careful validation.
Long term
By 2035, AI development shifts from data-hungry approaches to more efficient architectures (neurosymbolic, few-shot). The 'data crisis' is seen as a temporary growing pain.
Mitigation strategies
- Investment in curated, high-quality public training datasets
- Development of synthetic data validation and quality standards
- Fair compensation frameworks for human data contributors
- Research into data-efficient model architectures
- Public pressure and platform policies for labeling, flagging, or banning low-quality AI-generated content ('slop')
Top evidence
Hollywood is hiring ‘slop janitors.’ The price is visible.
AI Incident Database
The article discusses the emergence of 'slop janitors' in Hollywood, workers tasked with cleaning up low-quality, AI-generated content. This highlights the current limitations of generative AI in producing high-quality content without significant human intervention and the economic implications for creative professionals.
AI in scientific publishing: Slower, worse, and more expensive
Science (AAAS)
The article highlights that the integration of AI into scientific publishing is currently making processes slower, resulting in lower quality outputs, and increasing costs. This suggests that AI, rather than improving efficiency and quality, is having a detrimental effect on a critical sector for knowledge dissemination and academic integrity.
The epistemic debt of generative AI
Nature Machine Intelligence
The concept of "epistemic debt of generative AI" suggests a growing risk related to the reliability and truthfulness of information produced by AI systems. This could lead to an amplification of disinformation, degradation of knowledge quality through model collapse, and an erosion of trust in digital information.
Will the Open Internet Survive the War on Bots? Mapping the Debate Over “AI Preferences”
Center for Democracy & Technology
A conference will bring together experts to debate the future of the open internet in the context of 'AI Preferences' and the 'War on Bots.' This discussion addresses risks like AI-amplified disinformation, data scraping, and model collapse, while seeking governance solutions for the internet's integrity.
Can Generative Intermediaries Deliver When It Comes to Information Quality?
Center for Democracy & Technology
The article questions whether generative AI intermediaries can maintain information quality, highlighting concerns about how these new tools are reshaping information access. This raises risks regarding the reliability and accuracy of content consumed by the public.