OpenAI, the esteemed artificial intelligence start-up valued at approximately $300 billion, is reportedly expediting its testing processes for new AI models, raising alarm over the potential risks associated with the reduced safety evaluations. According to sources familiar with the company's internal operations, testing periods for assessing the safety and performance of its latest large language models (LLMs) have been drastically shortened from several months to mere days.
This accelerated pace of model development, driven by intense competition with major tech firms including Meta and Google, as well as emerging start-ups such as Elon Musk’s xAI, has led to concerns about the robustness of the safety measures being employed. One individual involved in testing OpenAI’s upcoming model, referred to as o3, commented, “We had more thorough safety testing when [the technology] was less important.” They expressed fears regarding the heightened risks associated with more advanced models, stating, “But because there is more demand for it, they want it out faster. I hope it is not a catastrophic mis-step, but it is reckless. This is a recipe for disaster.”
The urgency reportedly stems from the competitive pressures within the rapidly evolving AI sector, where companies are racing to leverage cutting-edge technology before their rivals. Despite the lack of global standards for AI safety testing, the European Union's AI Act, set to take effect later this year, will mandate companies to perform safety evaluations on their most powerful models. Previously, OpenAI had engaged in voluntary agreements with governments, including those of the UK and US, allowing for external assessments by AI safety research institutions.
OpenAI aims to launch the o3 model as early as next week, although the precise date remains subject to change. Some testers are reportedly being allotted less than a week for their safety evaluations, a stark contrast to the six months provided for the GPT-4 model, released in 2023. One tester from the GPT-4 evaluation disclosed that some significant safety concerns were only uncovered two months into the testing process, stating, “They are just not prioritising public safety at all.”
Daniel Kokotajlo, a former OpenAI researcher now leading the AI Futures Project, observed the implications of a lack of regulation around disclosure of potential risks, mentioning that companies are under considerable pressure to advance their models rapidly.
OpenAI has previously pledged to develop tailored versions of its models to investigate potential misuses, such as the risk they might pose in facilitating bioengineering tasks. Nonetheless, the company has reportedly limited its fine-tuning efforts to less advanced models rather than applying these practices to its latest and more capable technologies. Steven Adler, another former OpenAI safety researcher, highlighted the importance of these tailored tests, stating, “Not doing such tests could mean OpenAI and the other AI companies are underestimating the worst risks of their models.”
While OpenAI has claimed to have made processes more efficient through automated testing, it admitted the complexities of tasks such as fine-tuning do not have a universally accepted methodology. Johannes Heidecke, head of safety systems at OpenAI, defended the company’s approach, asserting, “We have a good balance of how fast we move and how thorough we are.”
Another significant concern is that tests are frequently not conducted on the final models released to the public, instead focusing on earlier versions, often referred to as “checkpoints.” Critics have described this practice as inadequate, noting that releasing models that differ from the evaluated versions undermines the integrity of the testing. OpenAI maintains that these checkpoints are “basically identical” to the final products launched, yet these assurances do little to alleviate growing apprehension regarding the rapid release of high-stakes AI technologies without comprehensive safety assurances. The dialogue surrounding OpenAI's current testing protocols and safety evaluations continues as the industry evolves.
Source: Noah Wire Services