Ganguli and colleagues: human red-teaming methods and scaling observations

What changed

Collected human adversarial conversations with language models and compared model size and training approaches, documenting how red-teaming outcomes, participant behavior and observed harms depend on the evaluation process.

Scientific scope

Collected human adversarial conversations with language models and compared model size and training approaches, documenting how red-teaming outcomes, participant behavior and observed harms depend on the evaluation process. A participant-generated dataset is not an unbiased estimate of real-world harm prevalence. Safety findings depend on recruitment, incentives, model access and rating procedures; human red-teaming complements rather than certifies deployment safety.

Evidence

The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. A participant-generated dataset is not an unbiased estimate of real-world harm prevalence. Safety findings depend on recruitment, incentives, model access and rating procedures; human red-teaming complements rather than certifies deployment safety.

Contributors

• The named authors in the research record · SCIENCE: Source-listed authors for the exact contribution; experimental subjects are not autonomous scientific authors. Source support and predecessor work retain separate roles. Individual effort shares and unresolved identities are not inferred.

• Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned

Provisional WRS assessment, edition 2026-10-23.4. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction