OpenAI slows advanced AI work for safety checks

OpenAI President Greg Brockman recently described how safety concerns have slowed the company’s most advanced AI work. The firm is pausing or extending development timelines to run extra tests and alignment checks before wider deployment.

Company leaders decided to slow progress on the most capable models so engineers and researchers can address potential harms. OpenAI has redirected staff and compute time toward safety evaluations, larger red‑team exercises and alignment research to identify failure modes and risky behaviours.

The change affects top‑tier projects that push model capability. Those efforts now face longer internal review cycles and more experiments designed to surface adversarial vulnerabilities, misinformation risks and other threats that could appear if more powerful models were released without further safeguards.

Brockman tied the adjustments to broader concerns about the societal effects of increasingly capable AI. He said OpenAI is balancing speed with caution and is willing to accept delays in feature rollouts and new model iterations while safety work continues.

OpenAI’s process includes intensified internal testing, expanded human feedback loops and more staff assigned to high‑risk projects. Training runs and deployment plans are being scheduled with extra time to allow repeated safety checks. When teams identify potential risks, development pauses until mitigation strategies are tested and approved; Brockman added that this pattern has become routine.

Brockman noted the company is consulting outside researchers and policymakers to broaden oversight and bring additional perspectives into evaluation processes. He said those conversations inform decisions about when and how to deploy new capabilities and help OpenAI understand public expectations and potential regulatory requirements.

OpenAI has previously paired product development with safety work, using iterative releases and deployment controls to limit harmful outputs. As models have grown more capable, concerns about hallucinations, misinformation, bias and misuse have increased and alignment research has expanded inside the company and across the field.

The company says it aims to release useful, safe AI while continuing research to reduce risks. Brockman’s comments describe how safety work is shaping the pace and sequencing of OpenAI’s most advanced projects.