The Pinnacle of Enshittification, or Large Language Models

Michał Górny, a long-time Gentoo developer, writes a comprehensive indictment of LLMs from the perspective of someone maintaining open-source infrastructure that's directly harmed by them. The piece covers data ethics, environmental costs, copyright, community fracturing, labor, and psychological dependency. His conclusion: LLMs are "fundamentally unethical regardless of utility" — asbestos, not a tool with trade-offs.

The parrot argument

Górny's technical framing is simple: LLMs do next-token prediction. They remix training data into statistically plausible output. "You can train a parrot to respond to some questions, but said parrot won't comprehend the questions." He argues society's embrace of "bullshit" — plausible-sounding discourse that means nothing — has primed people to mistake LLM output for intelligence, because the output matches the form of what they're used to accepting.

Infrastructure under siege

This is where Górny speaks from direct experience. LLM scrapers hit Gentoo's Bugzilla like a "constant DDoS attack," ignoring robots.txt, consuming bandwidth and CPU that legitimate users need. Administrators deploy protections (rate limiting, CAPTCHAs, blocking) that create accessibility barriers for the people the service is meant to serve. The scrapers don't care — they retry, rotate IPs, and keep coming.

The irony: the scrapers are consuming open-source project infrastructure to train models that will generate code that creates more work for those same projects when it arrives as low-quality contributions.

FLOSS community fracturing

The open-source world has split into three camps on LLM-generated code:

Reject all LLM contributions — some projects refuse them entirely
Require human review — accept LLM-assisted work if a human vouches for it
Lean into LLM generation — use LLMs heavily for code, docs, and communication

The third camp worries Górny most. Dormant projects suddenly releasing massive code changes that downstream maintainers can't reasonably review. LLM-generated bug reports that "sound plausible at first but are entirely wrong," followed by questionable fixes and broken releases. The curl project abandoned its bug bounty program because LLM-generated reports flooded it with plausible-sounding garbage.

Trust erosion goes both ways: people question whether suspicious code came from a real developer or an LLM, and formal writing styles are treated with suspicion. Some users have started mimicking chatbot language because they think it sounds more professional.

Copywashing

A specific pattern Górny flags: using LLMs to recreate existing projects while stripping the original attribution. The training process launders the provenance — the LLM absorbed the original code, regenerated something functionally equivalent, and the new version has no attribution trail.

The irony came full circle when Claude Code's leaked source was itself copywashed to prevent DMCA takedowns. The same "training isn't copying" argument companies used against copyright holders was used against them. Górny doesn't celebrate this — he sees it as evidence that the entire framework is broken.

Environmental and hardware costs

LLM training and inference consume hardware at a scale that distorts global markets. Górny notes video card and memory prices skyrocketing as data centers absorb production capacity (see hold-on-to-your-hardware for the detailed numbers). Rapid model obsolescence generates e-waste. Energy and water consumption for cooling is massive. Companies that made climate pledges have "quietly removed" them to accommodate AI infrastructure expansion.

Economic unsustainability

Górny echoes the ai-subsidy-economics argument: LLM companies lose billions, surviving on continuous subsidies and circular financial arrangements. Without fundamental breakthroughs, they can't reach profitability. But they may persist through psychological dependency and billionaire underwriting, even without a viable business model.

Labor displacement

LLMs are explicitly marketed as worker replacements. Górny lists concrete cases: artist commissions lost to generated assets, translation studios closing as companies switch to LLM translation despite quality issues, experienced workers replaced by junior staff using AI, remaining employees expected to produce more. The output is "bleak repetitive slop" — quantity over craft.

The local model counter-argument

Some argue smaller, self-hosted models solve the ethical problems. Górny is skeptical: the performance gap between industrial and local models is severe enough that most users won't downgrade, and the argument doesn't address the ethical problems of how the models were trained in the first place. This pushes back on titit-local-ai's position, though Neward was focused more on sovereignty and economics than ethics.

Where this sits in the wiki

This is the strongest anti-LLM piece in the wiki, complementing the more finance-focused critique in subprime-ai-crisis. Where Zitron argues the economics will collapse, Górny argues the technology is harmful regardless of whether it's profitable. The cult-of-vibe-coding piece by Bram Cohen takes a middle position — LLMs are useful tools if you apply rigor, harmful if you don't.