Concerns raised as OpenAI cuts AI safety testing time

OpenAI has significantly reduced the time and resources allocated to safety testing its advanced artificial intelligence (AI) models, according to the Financial Times (£), prompting concerns about the potential for its updates to be released without adequate safeguards. 

Testers are now given days, rather than months, to evaluate the risks and performance of new large language models (LLMs). Sources indicate that the acceleration in the testing process has led to less thorough evaluations, with insufficient time to identify and mitigate potential risks. 

OpenAI previously allowed much longer testing periods, such as the six months allocated for GPT-4. The article highlights that a tester of GPT-4 noted that some dangerous capabilities were only discovered months into the evaluation, suggesting that OpenAI's current, faster approach may not be prioritising public safety. OpenAI aims to release its o3 model as early as next week, leaving some testers with less than a week for safety checks.

While the EU Artificial Intelligence Act (AI Act) will mandate safety testing later this year, as things stand, no global standard exists. 

£ - This article requires a subscription.