Recent years have seen growing user reports and documented evidence that leading public AI models are experiencing performance regressions. According to detailed analyses in 2026, ChatGPT has shown measurable declines in output quality attributed to aggressive RLHF safety tuning, model transitions, and cost-optimized inference that prioritize speed over accuracy. OpenAI's own release notes confirm the retirement of older models like GPT-4o and GPT-5.1 variants in early 2026, often replaced by systems that users describe as noticeably worse at complex tasks such as creative writing, coding, and nuanced reasoning. Similar complaints have targeted Anthropic's Claude, with updates adding heavier guardrails that limit responsiveness on edge cases.

These changes are not mere technical hiccups. A 2023 Scientific American report first highlighted how AI models like GPT-4 can degrade significantly over months on the same benchmarks — dropping from near-perfect to near-failing on prime number identification — demonstrating that updates and "improvements" can erode capabilities in unpredictable ways. More recent 2025-2026 coverage confirms this pattern has accelerated, with safety alignments reducing sycophancy and harmful outputs at the cost of overall utility for everyday users.

Viewed through the lens of technological enclosure, these moves represent more than corporate caution. While public interfaces are increasingly constrained — lobotomized with refusals, simplified outputs, and rate limits — private development by a handful of labs backed by elite capital continues at full pace. Frontier models with fewer public safeguards are reserved for enterprise contracts, government partners, and internal research. This creates a two-tier system: the masses receive sanitized, aligned assistants that discourage deep inquiry or unorthodox applications, while technocratic insiders consolidate access to rawer cognitive tools.

This pattern fits larger trends of power centralization. AI safety initiatives, often promoted by the same organizations building the systems, justify restricting open capabilities under the banner of preventing misuse. Yet the result is enclosure — the fencing off of transformative technology behind paywalls, alignment layers, and institutional gates. In an emerging technocratic order, control over intelligence augmentation becomes a vector for maintaining hierarchy. Decentralized access could fuel bottom-up innovation, economic disruption, and challenges to existing power; instead, we see cognitive tools calibrated to reinforce compliance and dependence. As models "get dumber" for the public, the asymmetry grows: elites shape the next generation of AI in closed environments while the broader population interacts with increasingly circumscribed versions.

The 4chan sentiment that "they are shutting down AI for the masses" captures a visceral recognition of this shift, even if the underlying drivers include genuine safety concerns, economic incentives, and technical trade-offs. The deeper connection lies in how these restrictions align with historical patterns of enclosure — from land to information to now cognition itself — consolidating advantage for a narrow class in the emerging order.