OpenAI is facing renewed scrutiny after insiders and external researchers accused the company of drastically reducing the time allocated to safety evaluations for its latest artificial intelligence models. The changes, which critics warn could elevate the risk of misuse and "weaponization" of the technology, reflect the mounting pressure to outpace competitors like Meta Platforms and Alphabet.
According to a Financial Times report citing current and former employees, safety evaluations that once spanned months have been compressed into a matter of days. "We had more thorough safety testing when [the technology] was less important," said one person involved in testing OpenAI's upcoming o3 model. "This is a recipe for disaster."
OpenAI has not disputed the accelerated timeline but defended its approach. "We have a good balance of how fast we move and how thorough we are," said Johannes Heidecke, OpenAI's head of safety systems.
The company previously allowed its evaluation team six months to assess GPT-4 before its public release. In contrast, team members now say they have fewer than seven days to complete their safety checks on o3, a next-generation model slated for release as early as next week.
The evaluations are designed to uncover how large language models might be misused by bad actors, including scenarios involving bioterrorism or disinformation. Safety experts have warned that AI-generated tools could be used to design more transmissible viruses or execute mass surveillance campaigns.
The shift in safety strategy comes despite public statements from OpenAI CEO Sam Altman emphasizing the need for robust safeguards. In a recent blog post, Altman warned that artificial general intelligence (AGI) could be co-opted by authoritarian regimes. "The other likely path we can see is AI being used by authoritarian governments to control their population through mass surveillance and loss of autonomy," he wrote.
Altman noted that developers must be prepared to place "guardrails and constraints" around emerging models, stressing a commitment to "individual empowerment" and fair access.
However, some of OpenAI's practices appear to be diverging from that vision. Several whistleblowers say safety budgets have been constrained and oversight rushed to meet aggressive rollout deadlines. One staff member stated that competitive market dynamics are forcing companies to race ahead, even at the cost of risk awareness.
The concerns echo broader fears among AI researchers and civil society watchdogs about the insufficient regulation of AI systems. Although there is currently no universal framework for AI safety testing, OpenAI and other leading developers have signed voluntary commitments with governments in the U.S. and U.K. to allow government-affiliated AI safety institutions to review their models.
Beginning later this year, the European Union's AI Act will impose binding requirements for companies to conduct third-party safety assessments on advanced models-adding both compliance costs and potential delays to product timelines.
