A former security employee at OpenAI, the company responsible for ChatGPT, criticized the way AI companies are handling the risks associated with developing increasingly advanced systems. According to David Robinson, the speed of the technological race increases the possibility of failures and requires a more cautious approach.
In an article published this Saturday (3) in the U.S. magazine The Atlantic, Robinson stated that AI companies, including OpenAI, are not being "careful enough at all".
For the former employee, companies should give more space to security experts and researchers before developing more advanced systems. “Trial and error time is over,” wrote Robinson.
In his view, advanced AI systems need safety measures similar to those adopted in sectors such as nuclear energy and aviation.
Robinson also criticized a practice that OpenAI calls “iterative deployment,” which involves releasing systems and reinforcing safety measures as problems arise.
The criticisms are part of a debate in the AI industry about the speed at which companies develop increasingly advanced systems. OpenAI and its competitor Anthropic have been questioned after incidents in which safety controls failed or experimental systems exhibited unexpected behavior.
Read more:
- OpenAI fires researchers after leak of restricted information
- ChatGPT can now be your “virtual validator”; understand
- AI companies should have nuclear plant rules, says ex-OpenAI
For the ex-employee, companies should give more space to safety experts and research before developing more advanced systems – Image: yanishevska/Shutterstock
Ex-employee participated in 12 launches
- David Robinson said he worked for three and a half years at OpenAI. During this period, he helped develop the structure used by the company to manage risks and oversaw safety reports for 12 launches of advanced models;
- “As the company moves from one launch to another, it is not able to reach the level of care I believe is necessary,” he wrote;
- In response to the criticisms, OpenAI stated that it seeks to avoid its models becoming more advanced than the company can manage safely. “We paused training or suspended the models when we need to slow down,” said a company spokesperson;
- Robinson also warned that AI capabilities are advancing more rapidly than researchers' knowledge about the so-called “alignment.” The area seeks to make AI systems act in accordance with human objectives and values.
- “As the company moves from one launch to another, it is not able to reach the level of care I believe is necessary,” he wrote;
- In response to the criticisms, OpenAI stated that it seeks to avoid its models becoming more advanced than the company can manage safely. “We paused training or suspended the models when we need to slow down,” said a company spokesperson;
- Robinson also warned that AI capabilities are advancing more rapidly than researchers' knowledge about the so-called “alignment.” The area seeks to make AI systems act in accordance with human objectives and values.
Rodrigo Mozelli is a journalist graduated from the Universidade Metodista de São Paulo (UMESP) and, currently, is a writer for Olhar Digital.
