← Latest · Archive

SEV-3OpenAI
2 sources standard

OpenAI published a six-month follow-up on GPT-2 in August 2019, detailing its staged release strategy and observed misuse patterns since the model's initial partial release in February 2019 [source].

The company originally withheld GPT-2's full 1.5-billion-parameter version, citing concerns about malicious applications including misinformation generation, spam, and phishing. The follow-up reported that OpenAI released progressively larger model versions—117M, 345M, and 762M parameters—while monitoring for abuse.

OpenAI stated it found "no strong evidence of misuse" during the observation period. The company noted that detection of synthetic text remained feasible, though it acknowledged that automated detection tools showed imperfect reliability. Human evaluators could identify GPT-2 outputs approximately 52% of the time for the 1.5B model, only marginally better than chance.

The report documented that third parties replicated GPT-2's architecture and trained comparable models independently during the withholding period. OpenAI cited this as evidence that staged release provided limited security benefit, since determined actors could reconstruct similar capabilities.

The company announced it would release the full 1.5B parameter model in November 2019, contingent on continued absence of strong misuse evidence. OpenAI framed the staged release as an experiment in responsible disclosure, stating the approach allowed time for the research community to develop detection methods and study potential harms.

The follow-up provided no quantitative metrics on detection accuracy across model sizes, nor detailed taxonomies of attempted misuse cases. OpenAI indicated it would continue collaborating with external researchers on synthetic text detection and policy frameworks for future releases.

Why this is an AI incident

Launch-archive bulk classification (10 May 2026). Source signal originates from a real AI provider, regulator, or model-comparison probe; the harm or behavioural change described would not have occurred without the AI system being deployed in the role described. Editor reviewing the archive may amend the rationale per-wire.

Counterfactual "but-for" test per the Editor's Guide.

Codes M1, F10
Providers OpenAI