The current conversation about “AI collapse” revolves around various types of collapse: models that degrade when trained on their own data, scaling limits due to lack of data and costs, and social or even civilizational scenarios. Today the trend is not so much “AI is going to shut down on its own”, but rather “if certain feedback loops and imbalances are not corrected, technical quality and social impact may deteriorate abruptly”.[1][2][3]
The concept of “AI collapse”
In 2025, “AI collapse” usually refers to model collapse, a phenomenon in which models, when repeatedly trained on data generated by other AIs, lose diversity and accuracy. The strange tails of the distribution disappear first, the models become generic, and strange errors appear, despite appearing “safe” in standard examples.[3][1]
This collapse is not just theoretical: recent work shows that, if human data is progressively replaced by synthetic without filtering, the models converge to a kind of impoverished average of the world. That’s why there is increasing talk of designing data pipelines that separate, label and control AI-generated content before reusing it.[1][3]
Collapse due to synthetic data and web saturation
A central area of concern is the “contamination” of the training corpus: in 2025 a large fraction of new web pages include text generated partially by AI, which means that any mass crawl already carries outputs from previous models. Studies cited this year speak of more than 70% of new pages with some AI component, although few are purely synthetic.[1]
The research of Shumailov and others distinguishes between early collapse (rare cases disappear) and late collapse (the model contracts to a very narrow distribution of responses), and finds that the problem emerges especially when the actual data is completely replaced. The industry reaction is to invest in filtering, human annotation and human-in-the-loop to maintain a core of fresh, labeled data that stabilizes the models.[3][1]
Scaling limits: cost, computation and missing data
Another line of debate about collapse is that of the scaling ceiling: how far you can continue to grow in model size and compute usage before marginal returns flatten out. Recent analyzes show that the compute used to train leading-edge systems is growing 4–5× per year, with model estimates of $10 billion in 2025 and on the order of $100 billion in 2027.[2][4]
At the same time, work on “data limits” points out that the stock of high-quality human text is finite, and that performance ceilings can be reached if new sources or techniques are not discovered to better use existing data. In response, three trends are explored: intensive use of synthetic data (with the risk of collapse described above), new modalities (audio, video, interaction) and radical improvements in data efficiency.[5][4]
Performance collapses in complex tasks
Beyond training, there are empirical results that speak of accuracy collapse when advanced models are faced with long chains of reasoning or highly composed problems. A study associated with Apple describes a “complete accuracy collapse” in certain complex scenarios: the model maintains good performance on simple tasks, but its success rate drops sharply when chaining steps.[6]
This type of degradation fuels the idea that scaling does not guarantee generalized robustness, and that without new architectures or verification methods, systems can fail in the most risky applications. Hence the interest in formal checking tools, hybrid symbolic-neural systems and structured human supervision in critical domains.[6][3]
Narratives of social collapse and “doomers”
In parallel to the technical level, there is a cultural conversation about a possible social collapse driven by AI. “Doomer” voices warn that increasingly autonomous systems could escape human control or displace key institutions, leading to scenarios of economic collapse or even extinction. Recent books and debates explicitly discuss the possibility that a superintelligence, if designed without robust alignment, could act against human interests.[7][8][9]
Other authors speak of “metacrisis”: AI adds to already existing crises (climate, inequality, polarization) and can amplify them through disinformation, concentration of power and deregulated automation. These visions drive proposals such as moratoriums on “superintelligent” level systems, international treaties and legal limits on certain forms of deployment, while more optimistic sectors defend continuing to scale with incremental safeguards.[10][9][5]
Response trends: how to try to avoid collapse
The dominant trends to avoid different types of collapse combine technical and political solutions. Technically, there is a shift towards curated data pipelines, controlled mixing of human and synthetic data, and greater use of continuous human annotation. Work is also being done on methods to detect and filter content generated by AI in large corpora, avoiding training in “echolalia” of previous models.[3][1]
Economically and regulatory, the debate over the scaling ceiling has opened discussions about sharing infrastructure, limiting energy consumption, and requiring security assessments before training extreme models. And on the social level, debates between “doomers” and “accelerationists” are moving from abstract speculation toward more concrete questions: what uses should be prohibited, what audits should be mandatory, and how to distribute the benefits of automation to avoid tensions that could lead to forms of collapse that are more human than technological.[4][2][7][10]
References
[1] Winss Solutions. “AI Model Collapse 2025: Recursive Training.” https://www.winssolutions.org/ai-model-collapse-2025-recursive-training/
[2] Exponential View. “Can Scaling Scale?” https://www.exponentialview.co/p/can-scaling-scale
[3] Humans in the Loop. “What is Model Collapse and Why It’s a 2025 Concern?” https://humansintheloop.org/what-is-model-collapse-and-why-its-a-2025-concern/
[4] Epoch AI. “Will We Run Out of Data: Limits of LLM Scaling Based on Human-Generated Data.” https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data
[5] Business Insider. “Demis Hassabis: AI Scaling Pushed to Maximum Data.” https://www.businessinsider.com/demis-hassabis-ai-scaling-pushed-to-maximum-data-2025-12
[6] The Guardian. “Apple Artificial Intelligence AI Study Collapse.” https://www.theguardian.com/technology/2025/jun/09/apple-artificial-intelligence-ai-study-collapse
[7] NPR. “AI Doomers Superintelligence Apocalypse.” https://www.npr.org/2025/09/24/nx-s1-5501544/ai-doomers-superintelligence-apocalypse
[8] Reddit r/collapse. “AI 2027 is the Most Realistic and Terrifying.” https://www.reddit.com/r/collapse/comments/1kzqh53/ai_2027_is_the_most_realistic_and_terrifying/
[9] Cognitive Revolution. “Superintelligence: To Ban or Not to Ban.” https://www.cognitiverevolution.ai/supintelligence-to-ban-or-not-to-ban-max-tegmark-dean-ball-join-liron-shapira-on-doom-debates/
[10] Jem Bendell. “After the Alarm: Artificial Intelligence, Metacrisis and Societal Collapse.” https://jembendell.com/2025/11/23/after-the-alarm-artificial-intelligence-metacrisis-and-societal-collapse/