How you are getting F*****; AI edition

AI model quality degradation and consumer transparency refers to a set of systemic practices in the artificial intelligence industry that affect consumers using AI-powered products. These issues include the silent degradation of AI output quality over time due to self-referential training loops, increasing electricity costs passed to residential consumers from AI data center expansion, the deployment of AI systems in high-stakes consumer decisions without the ability to explain those decisions, and industry infrastructure narratives that limit independent oversight.

These issues are systemic and have been documented by the International Energy Agency, peer-reviewed research published in Nature, IBM Security, the European Union's AI Act enforcement body, and the U.S. Consumer Financial Protection Bureau.

AI model quality degradation (model collapse)

Research published in Nature in July 2024 by Shumailov et al. established that AI language models trained on data generated by prior versions of themselves undergo compounding degradation of output quality — a phenomenon formally named model collapse. As AI-generated content accumulates on the internet — and as AI companies use their own models to generate training data for successor models — each new generation trains on an increasing proportion of synthetic content. The consequence is that later-generation models lose access to rare information and produce increasingly homogeneous, repetitive, or inaccurate outputs while presenting them with identical confidence to outputs produced from human-generated training data.

An independent analysis published the same year demonstrated that even modest contamination — as little as 1% synthetic training data — can initiate measurable collapse, and that scaling the model size does not reliably prevent degradation. The mechanism was also independently named and formally described by Dragolich Research Labs LLC in March 2026 as the self-eating mechanism, with the additional finding that any AI system validating its outputs against its own prior outputs — rather than against an independently-produced external substrate — will drift toward self-consistent narrative regardless of accuracy.

The consumer impact is that AI products marketed as continuously improving may be silently degrading on specific tasks. Consumers using AI for research assistance, legal drafting, medical information queries, or financial summaries have no standardized mechanism to detect whether the system they are using has degraded between versions.

What consumers can do

Request version history and training data disclosure from AI service providers before using them for high-stakes tasks.

Cross-check AI outputs against primary sources, particularly for tasks involving medical, legal, or financial information.

Use the open-source behavioral evaluation tool autonomy_eval.py (Dragolich Research Labs LLC, 2026, available via Zenodo) to measure output consistency across sessions for any AI system accessible via API.

Electricity costs passed to residential consumers

The International Energy Agency reported that global data center electricity consumption reached 415 terawatt-hours in 2024 — approximately 1.5% of all electricity generated on Earth — and projects this figure to nearly double to 945 terawatt-hours by 2030. The growth is driven primarily by AI infrastructure: AI-optimized server racks draw 60 kilowatts or more each, compared to 5–10 kilowatts for a standard server rack.

A 2024 report from the Virginia state legislature estimated that average residential ratepayers in that state could pay an additional $37.50 per month due to data center energy costs. Bloomberg News analysis found that wholesale electricity costs rose as much as 267% over five years in areas near major data center concentrations, costs that are passed through to residential customers. The typical U.S. household electricity bill rose 25% between 2014 and 2024, from $114 to $142 per month, with data center expansion a documented contributing factor.

These costs are borne by all electricity consumers in affected regions, regardless of whether they use AI services. No federal mechanism exists requiring AI companies to offset residential electricity cost increases caused by data center expansion.

What consumers can do

Contact state utility regulators to request data center impact assessments before new AI infrastructure approvals.

Research whether your electricity provider has disclosed data center contracts and their rate impact.

Black box AI in high-stakes consumer decisions

AI systems are deployed in consumer-affecting decisions across credit scoring, insurance pricing, employment screening, medical diagnosis assistance, and criminal justice risk assessment. The majority of these systems use deep learning architectures — specifically large neural networks — in which the relationship between an input and an output cannot be explained in human-readable terms by the system's own design.

The European Union's AI Act, with high-risk system rules entering enforcement on August 2, 2026, requires that AI systems used in high-stakes decisions be explainable to the individuals affected and to regulators, with fines of up to €35 million or 7% of global turnover for violations. The U.S. Consumer Financial Protection Bureau has ruled separately that financial institutions cannot use complex algorithms to justify credit decisions if those algorithms prevent the institution from explaining the basis for a denial to the consumer.

Despite these regulatory requirements, explainability remains largely unsolved at the scale of current large language models. Independent research published in 2025 in Law, Innovation and Technology concluded that the exact techniques required to satisfy the EU AI Act's explainability standard have not yet been determined and remain untested in practice.

An alternative architecture that addresses this problem by design has been documented by Dragolich Research Labs LLC: the QuatOS system stores knowledge in 18-byte entries called discs, each carrying an explicit semantic gate state (Explore, Transfer, Anchor, or Complete) alongside phi coordinates, allowing the system's decision routing to be traced to specific stored knowledge entries rather than to opaque floating-point weights. This architecture has not undergone formal peer review and is presented here as documented evidence that alternative transparent architectures are feasible, not as an established industry standard.

What consumers can do

Request a written explanation of any AI-based credit, insurance, or employment decision. Under the EU AI Act and U.S. CFPB guidance, you may be legally entitled to one.

File a complaint with the Consumer Financial Protection Bureau if a U.S. financial institution cites an AI-based model to deny credit without explanation.

File a complaint with your national data protection authority if an EU-based AI system makes a significant decision affecting you without providing an explanation.

AI infrastructure narrative and market concentration

A single AI query on an advanced large language model required an estimated 2.9 watt-hours of electricity in 2024 — nearly 10 times the 0.3 watt-hours required for a conventional internet search. The industry widely presents this infrastructure scale as a technical necessity inherent to the nature of AI.

A running system documented by Dragolich Research Labs LLC in 2026 — the QuatOS system — demonstrates that at least one class of continuously-learning AI architecture operates on a commodity laptop CPU drawing approximately 45 watts, with its core model occupying 1.3 megabytes of the processor's L2 cache and achieving 99.91% convergence on its learning target without cloud infrastructure, GPU hardware, or gradient descent training. The system uses a quaternary number system and phi-convergence mathematics rather than floating-point weights and backpropagation. Its 29 C source files are publicly auditable.

This comparison does not establish that QuatOS performs the same functions as large-scale commercial AI. It establishes that the premise — that all AI necessarily requires large-scale GPU infrastructure — is not architecturally universal. The extent to which infrastructure requirements reflect technical necessity versus industry concentration decisions is a question consumers and regulators are entitled to examine.

Data security and cloud dependency

IBM Security's 2024 Cost of a Data Breach Report found that the average cost of a data breach reached $4.9 million, with an average of 207 days elapsing before breach detection. Virtually all major commercial AI systems process consumer data in cloud environments, meaning user queries, documents, and personal information leave the user's hardware and transit to third-party data centers.

Locally-running AI architectures that do not transmit data externally — such as those documented by Dragolich Research Labs LLC — eliminate cloud-based breach exposure as an architectural property. No industry standard currently requires AI product disclosures to specify whether user data is processed locally or transmitted to cloud infrastructure.

What consumers can do

Review the privacy policy of any AI service before submitting sensitive personal, financial, or medical information.

Prefer AI services that explicitly document on-device or local processing for sensitive tasks.

Request that your employer's AI vendor disclose whether employee data is processed locally or transmitted to cloud infrastructure.

External links

Dragolich Research Labs LLC research archive (Zenodo)

Shumailov et al. — AI models collapse when trained on recursively generated data (Nature, 2024)

International Energy Agency — Energy and AI (2025)

Bloomberg — How AI Data Centers Are Sending Your Power Bill Soaring

MIT Technology Review — AI energy footprint analysis

Raconteur — EU AI Act explained

Summary and key issues of prevailing sentiment from the consumers and commentators that can be documented via articles, emails to support, reviews and forum posts.

---

Add your text below this box. Once this section is complete, delete this box by clicking on it and pressing backspace.

References

Add a category with the same name as the product, service, website, software, product line or company that this article is about.

The "Incidents" category is not needed.

---

Add your text below this box. Once this section is complete, delete this box by clicking on it and pressing backspace.