Position: the Stochastic Parrot in the Coal Mine. Model Collapse is a Threat to Low-Resource Communities

This position paper argues that model collapse, driven by training generative AI on its own outputs, threatens to democratize AI by degrading data quality and efficiency, thereby disproportionately harming low-resource and marginalized communities through reinforced cultural biases and environmental costs.

The Gist

The Big Picture: The "Stochastic Parrot" in the Coal Mine

Imagine a canary in a coal mine. In the past, miners used canaries to detect dangerous gas; if the bird stopped singing, the miners knew to run.

This paper argues that Low-Resource Communities (people speaking less common languages or living in poorer regions) are the "canaries." They are the first to feel the danger of a phenomenon called Model Collapse.

What is Model Collapse?
Think of a game of "Telephone" played by a group of photocopiers.

1. You start with a clear, original photo (Real Human Data).
2. You make a copy. It's slightly blurry.
3. You take that blurry copy and make a new copy from it. It gets blurrier.
4. You keep doing this, copying the copies.

Eventually, the image becomes a muddy, unrecognizable mess. The details vanish, and only the most common, generic shapes remain.

In the world of AI, this happens when new AI models are trained on data created by old AI models. Because AI tends to repeat the most common patterns it sees, the "rare" and "unique" details get lost over time. The AI becomes a Stochastic Parrot—it mimics the sounds it hears but doesn't understand the meaning, and over generations, it only repeats the loudest, most common sounds, forgetting the quiet, unique ones.

The Problem: Why Poorer Communities Get Hurt First

The paper argues that while this "copying game" hurts everyone, it destroys the cultures of low-resource communities much faster. Here is why, using three main metaphors:

1. The "Rich vs. Poor" Data Diet

Imagine two people trying to stay healthy.

• The Wealthy Person (High-Resource): Has a massive pantry full of fresh, real food (Real Human Data). Even if they eat some processed, fake food (AI-generated data), they have so much real food that their diet stays healthy.
• The Struggling Person (Low-Resource): Has a very small pantry. They only have a few cans of real food. If they have to rely on processed, fake food to fill their stomach, they run out of real food very quickly.

The Paper's Claim: Low-resource languages (like many African or Indigenous languages) have very little data on the internet. If AI starts filling the internet with AI-generated text, these languages will be "poisoned" almost immediately because they don't have enough real data to dilute the fake stuff. Their unique cultural "flavor" will disappear first.

2. The "Echo Chamber" of Power

Imagine a town square where everyone is shouting.

• The loudest voices (English, Western culture, dominant viewpoints) are already heard by everyone.
• The quiet voices (minority groups, specific local dialects) are barely audible.

When AI learns from the internet, it acts like a megaphone that only amplifies the loudest voices. As AI generates more content, it repeats those loud voices over and over. The quiet voices get drowned out completely.
The Paper's Claim: Model collapse acts like a "Value-Lock." It freezes culture in the past, locking in the dominant viewpoints and erasing the attempts by marginalized groups to change social norms or reclaim their language. The AI forgets the "tails" of the distribution—the rare, unique, and diverse ways people speak.

3. The "Carbon Cost" of Trying to Fix It

Imagine trying to fix a leaky roof.

• The Wealthy Person can afford to buy new shingles and hire a crew to fix it.
• The Struggling Person has to try to patch it with tape and cardboard, which costs them their savings and makes the house hotter.

The Paper's Claim: To stop Model Collapse, researchers need more real data. But collecting real data is expensive and requires massive energy (computers running hot).

• Low-resource communities often live in areas already suffering from climate change and energy shortages.
• They bear the environmental cost of training these massive AI models but get the least benefit from them.
• They cannot afford to "buy" enough real data to save their languages from being erased by AI-generated noise.

The "Stochastic Parrot" Analogy

The paper revisits an old idea: AI is a "Stochastic Parrot." It doesn't understand; it just predicts the next word based on statistics.

• The Paper's View: Even though AI has gotten smarter, it is still a parrot. If you feed a parrot only the most common phrases, it stops saying anything interesting.
• The Danger: For low-resource communities, the "interesting phrases" (their unique culture, slang, and history) are the first things the parrot forgets because they are statistically rare.

What Does the Paper Want Us to Do?

The authors are issuing a Call to Action. They say we cannot wait until the AI breaks completely to worry about this.

1. Listen to the Canaries: Low-resource communities need to be the leaders in this conversation, not the afterthoughts.
2. Protect the Real Data: We need to create special "safe zones" of data that are guaranteed to be real human content, not AI-generated, specifically for these vulnerable languages.
3. Detect the Fake: We need better tools to spot AI-generated text so we can filter it out before it poisons the training data.
4. Accept the Risk: The paper admits that maybe the AI won't break globally for a long time, but for specific, small communities, the "break" is happening right now.

Summary

The paper warns that as AI generates more content, it creates a feedback loop that makes AI "dumber" and more repetitive. This process acts like a filter that removes the rare and unique. Because low-resource communities already have less representation online, their unique cultures and languages are at the highest risk of being erased by this process, leaving them with only a homogenized, dominant version of the world.

Technical Summary

Technical Summary: Model Collapse as a Threat to Low-Resource Communities

Problem Statement
The paper addresses the phenomenon of model collapse, defined as the degradation in performance and diversity of generative models when they are trained on data produced by prior generations of similar models. As artificially generated content proliferates (estimated at ~30% of the internet by 2025), successive model generations increasingly rely on "synthetic" data. This creates a feedback loop where models overfit to dominant patterns in the training distribution, systematically erasing rare or "tail" examples.

The authors argue that while model collapse is a general technical risk, it poses a disproportionate and existential threat to low-resource and marginalized communities. These communities already suffer from:

1. Under-representation: Their data occupies the "tails" of global data distributions.
2. Environmental Inequity: They bear the brunt of the carbon footprint and infrastructure costs of training large models without reaping proportional benefits.
3. Cultural Bias: Large Language Models (LLMs) tend to reinforce hegemonic viewpoints and "value-lock" outdated social norms.

The paper posits that model collapse acts as an "erasure engine," accelerating the removal of low-resource languages and cultural nuances from the digital sphere, thereby undermining efforts to democratize AI.

Methodology and Theoretical Framework
This work is a position paper that synthesizes existing theoretical and empirical literature rather than presenting new experimental results. The methodology involves:

• Literature Synthesis: Integrating findings on model collapse (e.g., Shumailov et al., 2024; Dohmatob et al., 2024) with critiques of LLMs regarding bias and environmental cost (e.g., Bender et al., 2021; Strubell et al., 2019).
• Theoretical Application: Applying the concept of neural scaling laws to synthetic data scenarios. The authors utilize the finding that even small fractions of synthetic data (approx. 1%) can alter scaling exponents, causing a "glass ceiling" where increasing dataset size no longer yields proportional performance improvements.
• Cognitive Science Analogy: Drawing parallels between model collapse and iterated learning (IL) in cognitive science, noting that while IL can lead to regularization, it also risks the loss of expressivity and the forgetting of differences between closely related data points without external intervention.
• Critical Analysis of Definitions: Engaging with recent debates (specifically Schaeffer et al., 2025) regarding the definition of model collapse. The authors distinguish between three types: (1) progressive inability to fit original distributions, (2) deformation of data distributions (loss of tails), and (3) decreasing marginal value of additional data.

Key Contributions

1. Intersectional Risk Analysis: The paper's primary contribution is linking the technical mechanism of model collapse to the socio-economic realities of low-resource communities. It argues that model collapse is not a uniform threat but a stratified one, where marginalized groups face the "worst-case scenario" first because their data is already scarce and located at the distributional tails.
2. The "Threefold" Disadvantage: The authors identify a compounding disadvantage for low-resource communities:
   • Data Scarcity: Limited real-world data makes them more vulnerable to being overwritten by synthetic data.
   • Environmental Cost: The increased compute required to combat model collapse (by collecting more data or training larger models) exacerbates environmental harm, which disproportionately affects these communities.
   • Cultural Erasure: The "value-lock" effect is amplified, where dominant viewpoints become entrenched while minority dialects, code-switching, and reclaimed language are systematically forgotten.
3. Re-evaluation of the "Stochastic Parrot": The authors revisit the "Stochastic Parrot" analogy (Bender et al., 2021), arguing it remains foundational. They contend that even with multimodal advancements, LLMs remain statistically grounded in observed patterns, making them susceptible to the frequency biases that drive model collapse.
4. Critique of Mitigation Strategies: The paper highlights the limitations of current solutions. For instance, "data curation" (filtering synthetic data) is becoming increasingly difficult as generative models become more convincing. Furthermore, automated curation tools may inadvertently remove legitimate, low-frequency linguistic variations used by marginalized groups.

Results and Observations
As a theoretical position paper, the work does not present new empirical metrics but synthesizes existing findings to support its arguments:

• Scaling Law Disruption: Citing Dohmatob et al. (2024b), the paper notes that achieving a tenfold reduction in test loss with 10% synthetic data requires an order-of-magnitude increase in dataset size compared to clean data, drastically increasing costs.
• Distributional Shift: Referencing Figure 1 (based on GPT-2 perplexity), the paper illustrates that low-resource languages (e.g., Albanian, Indonesian) already occupy higher perplexity distributions (the "tails") compared to high-resource languages (e.g., English). Model collapse predicts these tails will be the first to vanish.
• Cost Asymmetry: Citing Kandpal & Raffel (2025), the paper emphasizes that the cost of acquiring real data is 10–1000 times higher than the compute cost of training. This economic barrier prevents low-resource communities from "buying" their way out of model collapse by simply collecting more real data.

Significance and Call to Action
The paper concludes that while the severity of model collapse is debated in general contexts, it is an imminent and severe threat to low-resource and marginalized communities. The authors argue that the academic discourse has been too focused on high-resource settings and theoretical models, neglecting the specific vulnerabilities of these populations.

The Call to Action:

• Prioritization: The low-resource NLP and AI fairness communities must take the forefront in researching and mitigating model collapse.
• Data Sovereignty: There is an urgent need to construct and maintain datasets guaranteed to contain real, culturally authentic data for thousands of under-represented languages (e.g., Ndebele, Pedi, Sotho).
• New Research Directions: The paper outlines open questions regarding:
  • How model collapse affects agglutinative languages (which have long-tailed word distributions) differently than analytic languages.
  • The impact of code-switching on data distribution shifts.
  • Developing tractable measurements for model collapse and synthetic data detection.
  • Strategies for "decreasing" collapse, such as self-supervised learning or reinforcement learning to promote robust, out-of-distribution features.

Ultimately, the paper asserts that without immediate intervention, the online representation of multiple cultures risks being severely damaged or erased before the damage can be repaired.