OpenAI cuts safety tests in ‘reckless’ AI push
OpenAI has slashed the time it will spend testing its most powerful artificial intelligence (AI) models, raising concern that the $300bn company is prioritising speed over safety in the global AI arms race.
Once known for its rigorous approach to AI risk, OpenAI is now giving safety evaluators just days – sometimes even less than a week – to test new large language models (LLMs), according to a report in the Financial Times.
That’s a stark contrast to the six months afforded for GPT4 in 2023.
“This is when we should be more cautious, not less”, said one current tester. “It’s reckless’.
The reduced testing window comes as OpenAI races to stay ahead of competitors like Google, Meta, and Elon Musk’s xAI.
Internally, some researchers say the firm is no longer rigorously testing models for risks like biological weapon creation – tests which were once central to its safety strategy.
Unlike in Europe, where the EU’s new AI Act will soon require extensive risk assessments and transparency for advanced AI systems, there is still no binding regulation in the UK or US.
Instead, OpenAI and other major players have signed voluntary commitments to allow limited testing by external safety institutes.
‘They’re not being transparent’
Critics have warned that this is a dangerous mismatch in the high stakes global race.
“There’s no law saying they have to tell the public about what their models can really do”, said former OpenAI researcher Daniel Kokotajlo. “And they’re all in a rush.”
The contrast is especially stark between Brussels and Washington.
While the EU AI Act includes provisions for ongoing post-market monitoring and independent audits, US policy relies heavily on corporate self-governance.
The UK has backed a ‘pro-innovation’ approach, leaving safety largely in the hands of firms like OpenAI.
OpenAI has insisted that it is still prioritising safety through automated tools and ‘near final’ model testing.
But, former staff members have argued that the tech giant is cutting corners and downplaying risk: “They’re not being transparent. And that should worry everyone”.