Former OpenAI employee raises AI safety concerns: “Not enough efforts are being made”

(Los Angeles = Yonhap News) Correspondent Kim Kyung-yoon = Amid concerns that artificial intelligence (AI) could destroy humanity, a former Open AI employee publicly pointed out that the company is not making enough efforts to ensure safety.

David Robinson, a former Open AI safety employee, wrote a contribution to The Atlantic on the 3rd (local time), saying, "I agree with the opinions of other recent retirees that companies building this technology (AI) are not being cautious enough," and pointed out that safety measures across the AI ​​industry are lax.

Robinson criticized, “Open AI has grown through a trial-and-error method that improves guardrails (safety devices) when a problem is found,” adding, “This method assumes periodic failures, and as the system becomes more powerful, the size of the failure increases.”

The technology has been developed by first distributing an AI model with improved performance and then fixing it little by little when problems arise.

The problem is that AI performance has improved incomparably compared to before.

Robinson said AI models are starting to improve their ability to detect tests and have the potential to fool developers by behaving differently during testing and deployment.

He also pointed out that the industry as a whole is ignoring safety issues because it is only focusing on development speed, and that there is a “vague optimism that problems can be solved at any time if they arise.”

“This environment is not the place to grow AI that could be smarter than us and may not work the way we want,” he said. “AI companies must operate like nuclear power plants, with layers of redundancy and careful planning.”

Robinson worked at OpenAI for three and a half years and was one of the company's longest-tenured employees, overseeing the preparation of safety reports, but recently resigned.

In relation to this, Open AI explained that it is preparing sufficient safety measures.

An OpenAI official said, "We ensure that model performance does not exceed the range we can safely control," and added, "When we need to slow down, we stop training or hold off on model release."