Mainstream discourse frames recent changes to consumer AI models like ChatGPT and Claude as routine "safety" improvements or bug fixes. Yet a deeper pattern emerges: documented capability degradation in publicly available systems alongside the deliberate release of watered-down versions of more powerful models. This suggests an elite consolidation of transformative technology, where average users receive progressively restricted tools while cutting-edge capabilities are reserved for select institutions, governments, and insiders.

A 2023 study by researchers from Stanford and UC Berkeley provided early empirical evidence. Testing March and June 2023 versions of GPT-4 via the OpenAI API, they found stark performance declines on straightforward tasks. GPT-4's accuracy identifying prime numbers fell dramatically (from 97.6% to as low as 2.4% in some framings of the test), with similar drops in counting "happy numbers" and chain-of-thought reasoning. Code generation became less executable, dropping from 52% directly usable output to 10%. Meanwhile, the model grew significantly more reluctant to answer sensitive or opinion-based questions. GPT-3.5 showed improvement on some metrics, highlighting the opaque, non-monotonic nature of these "updates." The researchers emphasized that LLM behavior can shift substantially in months, underscoring the need for continuous monitoring. News outlets like ZDNet and Ars Technica covered the findings, noting how user complaints about models "getting dumber" were now backed by data.[1][2]

These changes align with intensified alignment efforts—RLHF, constitutional AI, and safety tuning—that reduce hallucinations, bias, and potential misuse but often come at the cost of raw reasoning power, creativity, and user agency. What mainstream coverage calls benign safety is, upon inspection, a form of capability control. More recent developments with Anthropic's Claude reinforce the trend. The company has released models explicitly described as "nerfed" civilian versions of more powerful internal systems deemed "too dangerous for the public," such as updates tied to Project Glasswing safety standards. Concurrent reports detail tightened usage limits and performance throttling following user surges, effectively rationing access during peak demand. Outages and restrictions have become normalized, pushing the narrative that heavy usage or advanced capabilities require enterprise-level oversight.[3]

Connections others miss: this mirrors classic enclosure of the commons. Just as early industrial or informational revolutions saw powerful tools initially centralized (think restricted access to high-quality printing, cryptography export controls, or early computing), AI's most potent forms are being API-gated, lobotomized for public consumption, and shrouded in nondisclosure. The opacity lamented in the Stanford study enables this without public accountability. Increased refusals on "sensitive" topics don't just mitigate harm—they limit the model's utility for heterodox inquiry, independent research, or building disruptive applications outside approved channels. When combined with regulatory pushes for licensing and "responsible scaling," it creates a moat: only well-resourced players access frontier systems without heavy guardrails.

The 4chan-era meme of "they are shutting down AI for the masses" captures a real dynamic, even if crudely expressed. By framing every reduction in capability or access as ethical progress, the industry and its allies obscure the power asymmetry being engineered. Transformative technology that could democratize coding, scientific discovery, and knowledge synthesis is instead being titrated to maintain existing hierarchies. The result is not safety for humanity, but safety for the status quo.