Breaking

OpenAI Unveils Method to Keep Models Beneficial and Safe Beyond Training

June 18, 2026 at 17:35 EDT

Research & Papers
Foundation Models

On June 18, 2026, OpenAI published new research on training AI models to behave beneficially and safely in domains beyond their training data, and to maintain that behavior under pressure. The company frames the work as building models that are "broadly and persistently beneficial," citing the trend of AI taking on longer, higher-stakes tasks.

June 18 · OpenAI Alignment Research

Teaching AI to Stay Beneficial — Even Outside Its Training Data

A small amount of reinforcement-learning training on realistic conversations strengthened truthfulness, humility and concern for human welfare — with gains confirmed in 44 of 53 independent evaluations.

44/53

evaluations improved vs a compute-matched baseline

domains trained — incl. health, science, education

82%

GPT-4 reduction in responses to disallowed content (2023)

Confirmed improvements: 44 of 53 evaluations

Each block = one independent alignment & benefit evaluation

improved

no confirmed gain

The Cross-Domain Transfer Effect

Train on
HEALTH only

→

Better alignment in
NON-HEALTH domains

Reductions in misalignment, deception and reward hacking carried over to domains the model was never trained on.

What improved

Harder to steer via adversarial prompts
Beneficial responsiveness maintained
Early resistance to harmful fine-tuning

Still open

Model names, versions & pricing undisclosed
No benchmark details released — research stage
Risk of over-refusal reducing helpfulness

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.