Across social media, a quiet frustration has been building. The AI tools that once felt sharp and surprising now hallucinate dates, confuse basic facts, and produce prose full of mistakes.
There are three separate but converging problems -model collapse, context rot, and a looming data wall- that are producing measurable deterioration in the AI tools that hundreds of millions of people now use daily. What began as theoretical warnings in academic papers is increasingly visible in the real world. Outputs are increasingly blander, less accurate, and more prone to confident error than the same systems were two or three years ago.
The photocopier problem
AI model collapse refers to the progressive degradation of a model's performance caused by training on data generated by other AI systems rather than original, human-created content. The concept was formally identified in a 2023 paper titled "The Curse of Recursion," authored by Ilia Shumailov and colleagues, which demonstrated that when AI models are trained recursively on synthetic data, they experience compounding information loss. Researchers describe the failure mode as analogous to a photocopier copying a photocopy: each generation of AI-on-AI training amplifies distortions and narrows the output distribution until the model degrades into low-diversity, high-error output.
Decreased diversity of outputs, with responses becoming increasingly similar across different prompts, and a rising rate of hallucinations or false claims are the clearest observable indicators. In image generation, the effects are starker still: research by Alemohammad et al. demonstrated that when image models are trained iteratively on their own output, they break completely within five generations, with distinctive features blurring and models beginning to hallucinate visual artifacts, unable to accurately recall what a normal object looked like.
The problem has been accelerating because synthetic content now dominates the web that AI systems scrape for training data. By April 2025, 74.2% of newly created webpages contained some AI-generated text, and AI-written pages in the top-20 Google results climbed from 11.11% to 19.56% between May 2024 and July 2025. The information ecosystem that AI was trained on is itself increasingly a product of AI, and the feedback loop is tightening. Jathan Sadowski, a senior lecturer at Monash University's Emerging Technologies Research Lab, has described this as "Habsburg AI," a reference to the genetic consequences of generations of royal intermarriage.
Lost in the middle
Separate from the training data problem, researchers have identified a second failure mode that affects how AI models process information in real time. Context rot is the measurable degradation in output quality that occurs as input context length increases. Even when a model's context window is nowhere near full, adding more tokens degrades performance. Chroma's 2025 research, which systematically tested 18 frontier models, found that every single one deteriorates as input length increases.
The gap between advertised and effective context windows has proved to be striking. Research by Norman Paulsen in 2025 introduced the concept of the Maximum Effective Context Window, finding that a number of leading models showed clear accuracy degradation by just 1,000 tokens, far below their advertised limits of 128,000 tokens or more. The NoLiMa benchmark, a long-context evaluation study, found that at 32,000 tokens, 11 out of 12 tested models dropped below 50% of their performance in short contexts.
The mechanism works through what researchers call attention dilution. Transformers distribute fixed attention across a growing number of tokens, attenuating focus on relevant content as context length increases. Models are also susceptible to being misled by coherent but irrelevant information, shifting their responses toward distractors rather than the material that actually matters. Researchers call this contextual distraction vulnerability, and it compounds with length: the longer the input, the more opportunities for noise to crowd out signal.
The data wall
Underlying both problems is a structural constraint the industry has known about for years but has yet to resolve. Research group Epoch AI projects that tech companies will exhaust the supply of publicly available human-generated training data sometime between 2026 and 2032. The type of text that provides genuine cognitive value, the kind that makes AI models smarter, is finite, and the industry is consuming it faster than any prior generation of humans created it.
The proposed solution is synthetic data, which is also the source of the collapse problem. High-quality, human-written web text is effectively running out, which is why there has been a rush of licensing deals between AI companies and publishers, news organisations, and online platforms. Shumailov, who left Google DeepMind in 2025 to launch his own AI security company, has said that model collapse is "unlikely to solve itself" and represents a persistent structural challenge, though he is less pessimistic than when the research was first published, noting that current models remain enormously useful even in their present state. AI critic Sadowski has a less sanguine view, arguing that companies may be deliberately downplaying concerns while sitting on limited internal knowledge about how to manage the problem.
Investment continues to accelerate
A growing number of researchers now describe the AI industry as facing a plateau under pressure. University of California experts, writing in January 2026, noted that the performance of large language models appears to have stalled, and that there are clear theoretical limits on their ability to learn straightforward concepts efficiently, even as investment continues to accelerate. That assessment cuts against the relentless optimism that has characterised industry communications for the past three years. Model collapse promotes hallucinations, loss of key training information, and the potential for poor decision-making, according to IT Pro's April 2026 analysis of the research literature. In healthcare, finance, and legal contexts, the consequences of that degradation are not merely inconvenient.
The industry spent years telling the world its tools were getting smarter. The data now suggests the opposite.
