OpenAI may loosen safeguards if competitors launch potentially dangerous AI models

OpenAI Adjusts Safety Protocol Amid Competitive Pressures

OpenAI has revised its Preparedness Framework, the internal mechanism guiding AI model safety assessments and safeguard implementation throughout development cycles. The updated policy notably includes provisions allowing potential relaxation of safety measures if competing firms launch high-risk systems without equivalent protections.

This modification underscores growing market forces compelling AI developers to accelerate deployment timelines. OpenAI faces mounting scrutiny regarding allegedly diminishing safety prioritization to enable swifter releases, including delayed publication of safety assessment documentation. Recent legal proceedings saw 12 former employees supporting Elon Musk's case against OpenAI, warning that planned corporate restructuring could further erode safety standards.

Balancing Competition and Safety

Anticipating criticism, OpenAI emphasized any policy adjustments would maintain rigorous protection levels. "While we may adapt requirements if competitors deploy high-risk systems lacking comparable safeguards," stated Tuesday's blog update, "we commit to thorough risk verification, transparency about modifications, and maintaining superior protective measures."

The revised framework highlights OpenAI's increasing reliance on automated evaluation tools alongside traditional human-led testing. Company representatives assert these automated systems enable maintaining release pace without compromising assessment quality.

Controversial Testing Practices Emerge

Reports contradict OpenAI's official stance, with insider accounts revealing compressed safety review periods. According to Financial Times sources, testers recently received under one week to evaluate an impending model - drastically shorter than previous timelines. Additional concerns emerged regarding safety testing occurring predominantly on pre-release model variants rather than public deployment versions.

OpenAI quietly reducing safety commitments.

Omitted from Framework changes:

• No safety tests required for finetuned models

— Steven Adler (@sjgadler) April 15, 2025

Redefined Risk Classification System

The updated framework introduces streamlined model categorization focusing on two capability thresholds:

• High Capability: Models amplifying existing severe harm vectors
• Critical Capability: Models establishing novel severe harm pathways

"High-capability systems mandate risk-mitigating safeguards pre-deployment," clarified OpenAI. "Critical-capability models require additional developmental phase protections."

This marks the framework's first substantial revision since 2023, reflecting evolving industry dynamics and internal operational changes at OpenAI.