gpt2-foreshadowed-capability-risk-dual-scaling

IN derived (depth 1)

Created 2026-06-21T10:16:19+00:00 · Reviewed 2026-06-21T14:41:08+00:00

GPT-2's staged release over misuse concerns, combined with measurements showing 1-7% exact duplicate training data in its outputs, illustrated early tensions between scaling language models and managing associated risks such as memorization and potential misuse.

Summary

GPT-2 was a concrete early case where a team had to withhold a working model from the public because of legitimate safety concerns, while at the same time the model was demonstrably regurgitating chunks of its training data verbatim. It established that as language models grow in size, the risks of memorization and misuse don't stay theoretical — they become practical problems that force a tradeoff between shipping capability and protecting against harm.

Justifications

SL — Memorization evidence plus staged release together establish that capability and risk scale in tandem

Antecedents (all must be IN):

  • IN gpt2-memorization-1-to-7-pct-exact-duplicates — GPT-2 models were measured to contain 1% to approximately 7% exact duplicate outputs from training data, and repetition prompting can trigger memorized excerpts.
  • IN gpt2-1point5b-params-staged-release — GPT-2 had 1.5 billion parameters, was released February 2019 with a staged release citing misuse risks, and the full model was released November 2019

Dependents

These beliefs depend on this one: