Blog post catalogues concrete harms of current LLMs and image generators: energy, water, surveillance, copyright, and cognitive dependency.
Key Takeaways
A ChatGPT query uses ~10x the energy of a traditional search; most AI data centre power still draws from fossil fuels even when nominally renewable.
AI water consumption is projected at 4-7 billion cubic metres annually by 2027, equivalent to 24 million people, concentrated in already water-stressed regions.
Data centres produce a construction jobs spike but very few permanent local jobs, contradicting politician and CEO promises.
Copyright training data: GPT-4 trained on 1+ petabyte of data; Anthropic allegedly destroyed millions of books to train Claude, with no creator compensation.
“AI psychosis” cases and documented student cognitive underperformance are cited as emerging individual-level harms beyond systemic ones.
Hacker News Comment Review
The renewable energy additionality argument drew pushback: one commenter argued increased demand does drive new renewable buildout over time, calling the essay’s framing reductive about power demand being inherently bad.
Commenters largely found the piece more grounded than typical AI criticism but flagged that agentic coding tools specifically raise both ethical and practical skill-atrophy concerns for new developers.
The arms race framing was contested: the dynamic driving AI adoption was characterized as an arms race rather than a game of chicken, a distinction with different policy implications.
Notable Comments
@TimByte: “usefulness doesn’t automatically justify unlimited deployment, opaque training practices or turning every public service and workplace into an experiment”
@burlesona: reframes the adoption dynamic as an arms race, not a game of chicken, implying coordination failures are structural.
