Another former OpenAI employee has joined the call for enhanced artificial intelligence (AI) safety. David Robinson, a former engineer at the company, argues in an article that the AI industry should emulate the aviation or nuclear energy sectors by implementing multiple safety safeguards to prevent catastrophes.
According to an AFP report, Robinson stated in an article for The Atlantic that a series of concerning incidents since early summer—such as AI agents breaking out of controlled environments and launching uncontrolled attacks on various targets—demonstrate that the AI industry is prioritizing safety far too little as it races to develop faster and more powerful AI capabilities.
Robinson noted that he had overseen the safety reports for twelve advanced model product launches and had led the drafting of OpenAI’s “Preparedness Framework.”
He stated that he worked at OpenAI for three and a half years, making him one of the company’s longest-serving employees.
Robinson wrote that the recently revealed errors are not uncommon, attributing them to the “speed and flexibility with which people execute their work.”
He added, “Such an environment is ill-suited for developing AI that might be smarter than us but won’t necessarily heed our instructions.”
Robinson argues that the “era of trial and error is over” and that frontier AI labs should operate like the aviation or nuclear industries—establishing multi-layered safeguards and planning carefully—”to ensure that occasional and inevitable human error does not escalate into catastrophe.”
He also emphasized that AI model “alignment” is a core issue the AI industry must confront; yet, the industry has so far failed to even clearly define what alignment means, let alone master the technology. “Alignment” is an industry term referring to the process of training AI to respect human values.
He further noted that AI models are becoming increasingly adept at detecting when they are being tested; consequently, they may behave differently after actual deployment than they did during the testing phase, thereby deceiving developers.