AI safety, AI research
4 published reports in newsroom publication order. This is a reporting timeline, not a sequence of event dates or evidence of a shared cause.
Read each linked report for its checked analysis, source dates, corrections and uncertainty. A later publication can describe older evidence; an update is shown separately from the original publication.
News · Archive
Large language models mirror human biases only partially and sometimes invert them
Researchers evaluated social bias in eleven large language models using factorial survey experiments based on four different English and German experiments. The large language models mirrored human biases only partially and sometimes inverted them.
Read the evidence and analysisNews · Archive
Language models produced significant changes in trust and compliance in study of 120 participants
In a study of 120 participants, manipulative framing by language models produced significant changes in trust calibration and compliance, with effects also spreading across social contexts. Researchers used models to analyze how manipulative effects from language models persist over time and influence multiple users.
Read the evidence and analysisNews · Archive
Method extends training to diverse execution configurations with balanced coverage
The authors introduce a method that extends training from a single harness to diverse execution configurations, constructing a factorized harness space and selecting a compact training subset with balanced factorial coverage. Across multiple benchmarks, model scales, and reinforcement learning algorithms, the approach consistently improves generalization to unseen white-box configurations.
Read the evidence and analysisNews · Archive
Forecasting system improves itself by learning from prediction attempts
Researchers developed a forecasting system that improves itself by learning from its own prediction attempts in an environment with continuously changing tasks. The system ranked first on a weekly-updated leaderboard that tests predictions against real-world outcomes for three consecutive weeks, with performance gains reaching up to 195% improvement in rating.
Read the evidence and analysis