Low-background Steel: content without AI contamination

Original Article ↗ Hacker News Discussion ↗

Value of “low-background” human content

  • Many welcome the low-background steel analogy: pre-AI, human-origin text is finite and increasingly precious for linguistics, historical research, and as clean training data.
  • Concern: mixing AI text into corpora (e.g., word-frequency datasets) permanently distorts language statistics and future research baselines.
  • Some see growing demand for “100% organic / human” content, akin to organic food, even if boundaries are fuzzy and enforcement imperfect.

AI training on AI output: risk vs practice

  • One camp fears “model collapse”: recursively training on synthetic data leads to degenerate, self-referential language and concepts.
  • Others note that:
    • Smaller test models trained on newer web scrapes (post-2022, likely AI-polluted) perform as well or slightly better than on older-only scrapes.
    • Synthetic data, when curated and filtered (including by other models), already improves frontier systems, especially in images and some text tasks.
  • Counterargument: current evaluations may be too crude to detect subtle degradation; first principles still favor avoiding opaque AI-generated data when possible.

Hallucinations, misinformation, and citogenesis

  • Repeated examples of confident but wrong answers (e.g., the MS‑DOS “Connect Four” easter egg) illustrate how LLM hallucinations can:
    • Be quoted online as fact.
    • Then be re-ingested as “evidence,” collapsing the distinction between “not in the training set” and “never existed.”
  • Some models now say “I’m not aware of…” rather than fabricating, but users still report overconfident falsehoods, especially in technical and historical niches.

Copyright, incentives, and reluctance to publish

  • Creators worry that AI systems will absorb their work and rephrase it without attribution, undermining career or reputational benefits of publishing.
  • Others stress that “knowledge isn’t copyrightable”: reading and re-expressing ideas (human or machine) has always been allowed, as long as you don’t copy protected text verbatim.
  • Disagreement centers on whether AI’s scale makes this morally or economically different from human learning.

Detection, labeling, and provenance

  • Many doubt humans (or current detectors) can reliably distinguish polished AI from human text.
  • Proposed technical schemes include:
    • Special Unicode ranges or invisible tags marking provenance.
    • HTML / metadata flags for “AI-generated” or “AI-edited.”
  • Critics argue such signals are trivial to strip or forge; any marker system risks false security and rapid circumvention.
  • Social solutions suggested: reputation systems, “organic” labels, and web‑of‑trust style relationships with known human creators, rather than hoping for perfect technical guarantees.

Archives, books, and alternative data sources

  • Archives (Wayback Machine, Project Gutenberg, shadow libraries) and pre‑AI snapshots are seen as long-term reservoirs of uncontaminated text.
  • Some advocate building personal physical libraries and relying more on paper references, both for robustness and as a check against AI-driven factual drift.

Attitudes toward AI vs “organic” content

  • Several commenters say they care less about origin and more about quality, and would rather see search engines penalize low-quality content, human or machine.
  • Others describe a growing aesthetic preference for rough, concise, obviously human writing over “ultra‑polished,” generic AI-style prose, and have started explicitly writing “organic” content again.