GPT-2: Too Dangerous To Release (2019)

Scope of “Too Dangerous to Release” (GPT‑2 in 2019)

• Early concern was mainly about spam, phishing, and misinformation, not coding or AGI.
• Some argue the caution was reasonable given unknowns in 2019.
• Others see it as an early instance of “AI is dangerous” marketing to attract attention, funding, and favorable regulation.

Did the Predicted Harms Materialize?

• Many say yes:
  • Explosion of low‑cost, hard‑to‑detect content: spam, SEO slop, propaganda, scams, AI‑written documents and emails.
  • Significant cheating in education; homework and essays trivial to fake.
  • Deepfakes and synthetic media eroding trust in images, video, and news.
  • New scam vectors (voice/video fraud, political hoaxes) and CSAM concerns.
• Others push back, asking for concrete quantification and noting:
  • Internet “enshittification” and content farms predated LLMs.
  • Some harms (e.g., memes of politicians as religious figures) say more about politics than AI itself.

Impact on Work and Software Engineering

• Several developers say their workflow is transformed: they write little or no code directly; non‑programmers ship apps in a day.
• Others counter that:
  • Simple scaffolding was always possible via cheap freelancers/offshoring.
  • Large/complex systems still break down without expertise and create security risks.
  • Maintenance and long‑term quality remain the hard part.
• Debate over whether new models will “solve” bad engineering vs. just enabling more low‑quality systems.

Societal, Economic, and Psychological Effects

• Reported negatives: loss of trust, job displacement (e.g., translators, juniors), higher resource use (RAM/GPU/energy), degraded online culture, and some users’ loss of hope about the future.
• Some hope for a Star‑Trek‑like post‑scarcity society; others warn that advanced AI under current capitalism could massively concentrate wealth and power.
• A minority note possible upsides:
  • Reduced reliance on toxic platforms (people stop scrolling and post less).
  • Productivity and mental‑health benefits from replacing tedious web search.

Governance, Safety, and Trust in Labs

• Disagreement over safety messaging:
  • Some see phased releases, model cards, and explicit danger framing as responsible.
  • Others view “too dangerous” rhetoric as a bid for regulatory moats and monopoly power, given perceived dishonesty and past behavior from major labs.
• A few argue it’s immoral to restrict access when models are trained on public human data; they advocate open weights or at least broad availability.