Summary: AI performance faces a "model collapse" crisis where a 20% pollution rate from synthetic training data results in a 10-percentage-point accuracy drop.

This scaling trend is environmentally unsustainable, with AI projected to consume 4% of global energy and up to 765 billion liters of water annually by 2026.

The proposed Provenance and Integrity Framework (PIF) resolves these risks through auditable "Clean Room" disclosures of data origins, scientific retraction tracking, and granular climate metrics.

The Convergence of Performance, Pollution, and Provenance: A Strategic Framework for Generative AI Governance

by Gemini 3.0, Deep Research. Warning, LLMs may hallucinate!

The rapid transition of generative artificial intelligence from a novel research pursuit to a foundational layer of global infrastructure has precipitated a triadic crisis of confidence, sustainability, and reliability. As frontier models increasingly dictate the parameters of human discourse, clinical diagnostics, and financial risk, the lack of transparency regarding their underlying data architectures has become a systemic vulnerability. The contemporary artificial intelligence landscape is defined by a paradoxical tension: while the industry seeks to scale computational intelligence to unprecedented heights, it is simultaneously confronting the degenerative effects of model collapse, the ecological toll of astronomical energy consumption, and a widening “trust gap” between developers and the critical sectors that must operationalize these systems.¹ The fundamental challenge for the mid-2020s is the establishment of a verifiable “Trust Infrastructure” that reconciles proprietary interests with the public’s need for safety and the planet’s requirement for environmental stewardship.

The Performance-Provenance Nexus: Data Quality and the Mechanics of Collapse

The axiom that an artificial intelligence model is only as good as the data used to train it has taken on a new, more urgent dimension as the supply of high-quality, human-generated content approaches exhaustion. Research suggests that the era of improving model performance simply by crawling vast swaths of the public internet is ending, with the stock of human-written text predicted to be depleted as early as 2026.² This “digital drought” has forced developers to rely increasingly on synthetic data—content generated by earlier iterations of AI models—which introduces the risk of “model collapse” or “model autophagy”.³

The Degenerative Feedback Loop of Model Collapse

Model collapse is a phenomenon where machine learning models gradually degrade due to errors inherited from uncurated synthetic data or training on their own previous outputs.⁵ This process is not a linear decline but a phased degeneration. In the “early model collapse” phase, systems begin to lose information about the “tails” or extremes of the data distribution, which primarily affects minority data and rare edge cases. Because overall performance benchmarks may appear stable or even improve during this phase, early collapse is notoriously difficult to detect, yet it erodes the model’s ability to handle nuance and complexity.⁵

The progression into “late model collapse” is characterized by a significant loss of variance and the confusion of core concepts. The model’s view of reality narrows, causing it to produce repetitive, unoriginal, and increasingly inaccurate results.² Mathematically, this can be observed in simplified 1D Gaussian models where original data follows a normal distribution:

When successive generations are estimated using sample means and variances from the previous model’s output, each generation represents a new step in a random walk of model parameters.⁵ For the approximation to remain accurate, the sampling rate would need to increase superlinearly, a requirement that is rarely met in massive-scale web crawling.⁵

Empirical Correlations Between Pollution and Performance

The impact of “polluted” training data is quantifiable and severe. Research conducted in 2025 demonstrated that a 20% pollution rate in training datasets—where one-fifth of the content is AI-generated rather than human-generated—leads to a roughly 10 percentage point drop in model performance across various tasks, including classification, regression, and clustering. On the IBM Telco Customer Churn dataset, for instance, accuracy dropped by nearly 10 percentage points when the pollution threshold reached 20%, with further decay as pollution approached 45%.¹

This performance decay is directly linked to the loss of semantic diversity. Large language models trained on predecessor-generated text exhibit a consistent decrease in lexical and syntactic variety.⁵ Even a small fraction of synthetic data—as little as one per 1000 tokens—can be detrimental asymptotically, suggesting that larger training sets do not necessarily enhance performance if the data quality is compromised.⁷ Interestingly, while increasing model size aligns with current scaling trends, evidence suggests that larger models can actually amplify model collapse in certain regimes, although they may mitigate it once they cross high interpolation thresholds.⁷

The Environmental Imperative: Pollution and Energy Waste

The correlation between AI performance and global pollution is an inescapable reality of the current “compute-first” development paradigm. The rapid expansion of artificial intelligence has introduced a dual environmental crisis: the creation of degraded training data and the astronomical energy and water waste generated by the data centers required to process it.

Energy Consumption and the Carbon Footprint of Inference

A single request to a chatbot like ChatGPT can consume ten times more energy than a standard Google search.⁸ In the United States, data centers already consume more than 4% of the national energy mix, and projections suggest this could rise to 12% as early as 2028.⁸ Many of these centers are being constructed in regions where the energy grid relies on fossil fuels, such as the xAI facility in Memphis, Tennessee, which utilizes gas-powered turbines that increase local air pollution.¹

Water Scarcity and the Hidden Cost of Cooling

The cooling requirements for AI infrastructure pose a significant threat to global water resources. In 2025, AI-related data centers consumed between 312 and 765 billion liters of water, a volume equivalent to the world’s total annual bottled water consumption.¹ The water intensity of individual tasks is equally high: generating one AI image can consume up to 50 liters of water for cooling, while a standard query consumes roughly 50ml.¹ This consumption often occurs in regions already experiencing water scarcity, creating a direct conflict between technological advancement and basic resource security.¹

Furthermore, the hardware lifecycle of AI systems exacerbates the problem of electronic waste. Specialized hardware like GPUs is replaced 2-3 times faster than standard IT equipment, yet only 15% of this waste is currently recycled.¹ This “E-Waste” cycle contributes to a mounting environmental burden that is often omitted from corporate sustainability reports.

The Transparency Paradox: Narratives vs. Auditable Infrastructure

The demand for transparency regarding training data has evolved from an intellectual property debate into a fundamental requirement for consumer protection and product liability. While leading developers have begun to issue public summaries of their data practices, a significant “trust gap” remains. For critical sectors like healthcare, law, and finance, high-level narratives are insufficient to validate the safety and reliability of AI systems.

Corporate Disclosure Strategies: A Comparative Analysis

OpenAI, Anthropic, and Google have adopted varying approaches to transparency, primarily driven by the requirements of California’s Assembly Bill 2013. OpenAI utilizes a minimalist approach, describing its training data in broad categories: public data, licensed partnerships, user-provided information, and synthetic data.¹ While OpenAI notes its systems are trained on “trillions of tokens,” it avoids naming specific external datasets or detailing its technical obfuscation methods for privacy.⁹ This creates a “wait-and-see” strategy that prioritizes proprietary protection over verification.



^{Sources: https://help.openai.com/en/articles/20001044-training-data-summary-pursuant-to-california-civil-code-section-3111 and https://help.openai.com/en/articles/7842364-how-chatgpt-and-our-language-models-are-developed and https://platform.openai.com/docs/bots?utm_source=chatgpt.com}

Anthropic offers more structural detail through its “transparency hub,” adhering closely to the AB 2013 categories. It provides unusually concrete information about its crawling behavior and adherence to robots.txt, signaling an ethical commitment to website operator choice. However, Anthropic still avoids naming specific non-public datasets obtained commercially, and its data size descriptions remain in general ranges.



^{Sources: https://trust.anthropic.com/resources?name=ab-2013-training-data-summary&s=km4kixlo1o4zwqj0ybiaw and https://www.anthropic.com/transparency and https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler}

Google’s disclosures focus on the scale and ecosystem of its data, listing nearly 30 apps and services—from YouTube to Gemini—that contribute to its training pipeline. Google provides the most granular scale metrics, including over 1 trillion text tokens and 1 million hours each of video and audio. Yet, like its peers, Google’s summaries are narrative in nature and provide no mechanism for individual rightsholders to verify their inclusion or for scientific researchers to track the weighting of sources.



^{Sources: https://files.lbr.cloud/public/2026-01/pdf-report-jj_2026-1-1_2026-1-1_en_v1.pdf?VersionId=s42JWvJRH2Gj5ElN8HhWZ7i_SIehb0fT and https://developers.google.com/crawling/docs/crawlers-fetchers/google-common-crawlers and https://developers.google.com/crawling/docs/crawlers-fetchers/overview-google-crawlers and https://deepmind.google/models/model-cards/}

The Failure of Opacity in Professional Domains

The persistence of the “black box” model has led to documented failures in high-stakes applications. In clinical reasoning, a study of 21 LLMs showed that while models could arrive at a correct final diagnosis with complete information, they failed to generate an accurate initial differential diagnosis 70% of the time.¹ This failure suggests that models are often pattern-matching against training vignettes (such as those from the MSD Manual) rather than engaging in procedural reasoning. Without granular provenance, medical professionals cannot determine if a model’s accuracy is the result of true capability or data contamination.

Scientific integrity is similarly compromised. Models trained on uncurated web scrapes often ingest “poisoned” data, such as academic articles that have been retracted due to fraud or ethical violations. A study published in Learned Publishing revealed that ChatGPT confirmed claims from retracted papers as true two-thirds of the time and failed to flag a single retraction across 217 evaluated articles. The lack of a “retraction tracking” mechanism in current training pipelines ensures that disproven scientific claims are propagated as factual information by AI systems.

The Global Regulatory Response: From Voluntary to Mandatory

The regulatory discourse surrounding AI has shifted toward binding statutory obligations, characterized by a tension between fostering innovation and mitigating systemic risks. The global landscape is currently defined by three primary models: the risk-based hierarchy of the European Union, the consumer-protection model of California, and the directive-based governance of China.

The EU AI Act and Article 53

The European Union AI Act (Regulation (EU) 2024/1689) provides the most detailed global standard, particularly for providers of General-Purpose AI (GPAI) models. Under Article 53, GPAI providers must maintain technical documentation and publish a “sufficiently detailed public summary” of training content. The European Commission’s mandatory template, released in July 2025, requires the disclosure of individual large datasets, web-crawling behavior, and the top domain names used for scraping. This framework is designed to move beyond self-certification toward auditable compliance.¹³

California’s AB 2013 and the US De Facto Standard

In the United States, California’s AB 2013 has set a national standard by requiring generative AI developers to post documentation across 12 specific categories. These include data point scale, acquisition methods (such as web crawling or licensing), and whether the datasets include personal or copyrighted information.  While the law includes narrow exemptions for national defense and security, its broad application to all generative AI systems made available to Californians has sparked significant legal debate over trade secrets and the First Amendment.

Directive-Based Governance in China

China’s “Interim Measures for the Management of Generative AI Services” take an integrated approach to transparency and social control. Beyond source disclosure, Chinese regulations mandate that training data be “true and accurate” and require providers to register their algorithms with the Cyberspace Administration of China (CAC). By the end of 2025, over 740 services had completed this filing process, illustrating a model where transparency is used to ensure adherence to state-defined social and ethical standards.

The Proposed Framework: AI Provenance and Integrity (PIF)

To resolve the transparency gap without compromising commercial security, this report proposes a two-track governance model: the AI Provenance and Integrity Framework (PIF). The PIF is designed to enable “assurance transparency,” moving from static summaries to an auditable trust infrastructure.¹

Track 1: The NIST-Administered Certification Program

The regulatory track empowers the National Institute of Standards and Technology (NIST) to establish a graded certification system aligned with the AI Risk Management Framework.¹ This system would standardize dataset metadata, including origin URLs, acquisition terms, and integrity signals such as retraction status. Certification would be tiered (Bronze, Silver, Gold), with “Gold” ratings required for high-impact functions in healthcare, finance, and government. To protect trade secrets, NIST would manage a “Clean Room” mechanism where regulators can review sensitive provenance evidence without public disclosure.

Track 2: The Statutory Provenance and Integrity Duty

The legislative track would establish a federal legal obligation for developers to maintain and produce provenance records. This includes a heightened “duty of care” for version-sensitive content, such as medical and legal data, requiring developers to track and update record-status changes (e.g., retractions). Metadata preservation must be mandatory through the ingestion and preprocessing stages to ensure that source information is not stripped out.

Ideal Data Points for Disclosure

The following framework identifies the specific data points that AI companies should disclose to satisfy the needs of users and regulators while maintaining commercial confidentiality.

Implementing the “Clean Room” Audit Methodology

The “Data Clean Room” (DCR) is the primary technical solution for reconciling transparency with commercial risk. A DCR provides a safe, neutral space for data collaboration without either party having access to the other’s raw data.¹⁷ By embedding privacy-enhancing technologies such as differential privacy and encryption, DCRs allow for the training and auditing of machine learning models on shared datasets without compromising confidentiality.¹⁴

Zero-Trust and Hardware-Backed Isolation

Effective AI auditing requires a zero-trust framework where privacy is enforced by technology, not just contractual promises.¹⁵ Modern DCRs utilize hardware-level encryption to ensure that neither the provider nor the auditor can access raw training data.¹⁵ This allows pharmaceutical companies, for instance, to run proprietary algorithms across healthcare datasets to evaluate drug adherence without exposing the underlying patient data or the proprietary model weights.¹⁸

Workflow and Adoption in Critical Sectors

The transition to DCRs accelerates adoption in critical sectors by bypassing traditional onboarding bottlenecks.¹⁵ Because DCRs ensure data minimization and purpose limitation—requirements under Article 25 of the GDPR—they provide a legally sound mechanism for the wide-scale adoption of AI in regulated domains.¹⁵ The success of this model is already visible in the advertising and healthcare sectors, where companies like Hershey’s and Hill’s Pet Nutrition use DCRs to enrich their data while maintaining privacy compliance.¹⁷

Conclusion: Toward a Sustainable and Verifiable AI Ecosystem

The current trajectory of generative AI development—characterized by a reliance on unverified synthetic data and an unsustainable environmental footprint—threatens the long-term viability of the industry. The correlation between performance degradation and data “pollution” is a scientific reality that high-level corporate narratives cannot obscure.¹ For AI to move from a promising innovation to a dependable infrastructure, the “black box” must be replaced by a verifiable trust architecture.

The proposed AI Provenance and Integrity Framework (PIF) offers a path forward. By mandating the tracking of scientific integrity signals, establishing NIST-administered “Clean Room” audits, and disclosing granular environmental metrics, the industry can satisfy the demands of critical sectors and regulators without exposing its “secret sauce”. The future of AI adoption will not be determined by the most impressive model, but by the most “assurable” one—the system that can prove its data is clean, its reasoning is sound, and its impact on the global climate is accounted for.¹ Only through this transition from document transparency to assurance transparency can the AI sector bridge the trust gap and fulfill its potential as a partner in scientific, medical, and social progress.

Works cited

1. Generative AI ROI: Why 80% of Companies See No Results, accessed May 2, 2026, https://www.fullstack.com/labs/resources/blog/generative-ai-roi-why-80-of-companies-see-no-results & The AI Data Transparency Index, accessed May 2, 2026, https://theodi.org/insights/reports/the-ai-data-transparency-index/

2. Why 2026 is the Year Synthetic Data Becomes Non-Negotiable - Towards AI, accessed May 2, 2026, https://pub.towardsai.net/why-2026-is-the-year-synthetic-data-becomes-non-negotiable-b5a2a84d1b1b

3. Synthetic Data: The New Data Frontier - World Economic Forum publications, accessed May 2, 2026, https://reports.weforum.org/docs/WEF_Synthetic_Data_2025.pdf

4. What Is Model Collapse? - IBM, accessed May 2, 2026, https://www.ibm.com/think/topics/model-collapse

5. Model collapse - Wikipedia, accessed May 2, 2026, https://en.wikipedia.org/wiki/Model_collapse

6. UvA-DARE (Digital Academic Repository) - Research Explorer, accessed May 2, 2026, https://pure.uva.nl/ws/files/308518741/1-s2.0-S0306437925000341-main.pdf

7. Strong Model Collapse | OpenReview, accessed May 2, 2026, https://openreview.net/forum?id=et5l9qPUhm

8. Growing Energy Demand of AI - Data Centers 2024–2026 | TTMS, accessed May 2, 2026, https://ttms.com/growing-energy-demand-of-ai-data-centers-2024-2026/

9. Training Data Summary Pursuant to California Civil Code Section ..., accessed May 2, 2026, https://help.openai.com/en/articles/20001044-training-data-summary-pursuant-to-california-civil-code-section-3111

10. Anthropic’s Transparency Hub, accessed May 2, 2026, https://www.anthropic.com/transparency

11. AI Training Data Transparency Summary, accessed May 2, 2026, https://files.lbr.cloud/public/2026-01/pdf-report-jj_2026-1-1_2026-1-1_en_v1.pdf?VersionId=s42JWvJRH2Gj5ElN8HhWZ7i_SIehb0fT

12. Citation of retracted publications: A challenging problem - ResearchGate, accessed May 2, 2026, https://www.researchgate.net/publication/349168209_Citation_of_retracted_publications_A_challenging_problem

13. Understanding General Purpose AI - European Institute of Public Administration (EIPA), accessed May 2, 2026, https://www.eipa.eu/blog/understanding-general-purpose-ai/

14. What Is Data Clean Room? Types, Benefits, and Use Cases - Tredence, accessed May 2, 2026, https://www.tredence.com/blog/data-clean-room

15. What are the best data clean room companies in 2026? - Decentriq, accessed May 2, 2026, https://www.decentriq.com/article/data-clean-rooms-compared

16. Networking-Aware Energy Efficiency in Agentic AI Inference: A Survey - arXiv, accessed May 2, 2026, https://arxiv.org/html/2604.07857v1

17. What is a Data Clean Room? How it Works and Use Cases - LiveRamp, accessed May 2, 2026, https://liveramp.com/explainer/data-clean-rooms

18. Data Collaboration Service – AWS Clean Rooms Features, accessed May 2, 2026, https://aws.amazon.com/clean-rooms/features/