OpenAI Employees Claim AI Safety Promises Were Not Met

AI expert OpenAI promised the White House last summer that it would put fresh iterations of its technology through extensive security testing to make sure the AI wouldn’t harm, such as teaching people how to make bioweapons or assisting hackers in creating new kinds of cyberattacks. Three anonymous sources claim that this spring, several OpenAI security team members were under pressure to complete a new testing procedure quickly to meet the May launch date set by OpenAI leadership.

Pre-Launch Celebrations and Safety Concerns

The first significant test of the new safety architecture, involving human reviewers and outside auditors, was OpenAI’s most recent model, GPT-4o. Nevertheless, despite protests from staff members, the testing was condensed into a single week. An open letter, signed in June by a number of both current and past employees of OpenAI, demanded that AI businesses release their workers from confidentiality agreements so that they may alert the public and authorities to security dangers.

Days after GPT-4o’s release, Jan Leike, a former OpenAI executive, quit, citing the company’s emphasis on products over security. William Saunders, a different ex-OpenAI developer, also observed a trend of hurried security work. Although the assessments were finished within a week, an OpenAI readiness team spokesperson acknowledged that the timetable was “tight” and not the ideal method.

President Biden has encouraged Congress to pass new legislation to protect the public from the risks posed by artificial intelligence. “President Biden has made clear to technology companies the importance of ensuring their products are safe and trustworthy before making them available to the public,” said Robyn Patterson, a spokesperson for the White House. Almost a dozen businesses have willingly committed to the White House, including Anthropic, Nvidia, Palantir, Google DeepMind, and Meta. OpenAI is one of them.

Testing Protocols and Internal Discontent

OpenAI’s latest model, GPT-4o, was the first major test of the new safety framework, involving human reviewers and external auditors. However, the testing was compressed into a single week, despite complaints from employees. Several current and former OpenAI employees signed an open letter in June, demanding that AI companies release their employees from confidentiality agreements to warn regulators and the public about security risks.

Former OpenAI executive Jan Leike resigned days after GPT-4o launched, criticizing the company’s prioritization of products over security. Another former OpenAI engineer, William Saunders, also noted a pattern of rushed security work. An OpenAI readiness team representative admitted that while the evaluations were completed within the week, the schedule was “tight” and not the best approach.

Despite the strain on teams, according to OpenAI spokeswoman Lindsey Held, the business did not cut corners on security procedures. The corporation undertook significant internal and external testing and initially withheld certain multimedia elements for more security work to fulfill White House obligations.

The event clarifies OpenAI’s internal dynamics as well as the possible dangers of AI development moving too quickly. The preparedness project by OpenAI sought to apply scientific rigor to the analysis of catastrophic risks; nonetheless, the expedited testing procedure and internal dissatisfaction underscore the difficulties in striking a balance between safety and innovation. To prevent such problems in the future, the company’s leadership is determined to give testing additional time.

Photo by Andrew Neel