AI safety experts have long warned of the possible negative implications of unchecked development of increasingly intelligent AI, and Dan Hendrycks, an AI safety expert and director of the Center for AI Safety, recently published a paper that outlines a range of potential doomsday scenarios from weaponization and power-seeking behavior, to data bias and privacy breaches. The paper emphasizes that safety should be considered at the forefront when it comes to AI systems and frameworks are being designed.
The potential risks the paper highlighted include: weaponization of AI systems, power-seeking behavior, malicious data poisoning and bias, data privacy breaches, resource misuse, emergent behavior, automation of malicious tasks, and input space bypass attack. Although the risks might be thought low probability, Hendrycks warns that they cannot be dismissed and our institutions must address them to prepare for the larger risks that may arise in the future.
Other experts in the field, such as Elon Musk and Sam Altman, CEO of OpenAI, have echoed similar sentiments and recently co-signed an open letter that calls to suspend the creation and training of AI models that are more powerful than GPT-4 in order to address the current AI arms race. Although the frustration of the open letter was acknowledged, Altman addressed the fact that it was missing technical nuances and OpenAI has no plans to create GPT-5.
Making sure that safety requirements are not overlooked during AI development processes may be difficult to implement, considering the competition between developers to build the most powerful AI models. However, Hendrycks stresses the importance of taking the time to ensure that safety measures are taken and that AI is built responsibly. “You can’t do something both hastily and safely,” he said.
Safety remains an integral part of the development of AI systems and is necessary to ensure that the risks remain low and AI systems bring about more good than harm. Developing with safety in mind require a combination of technical measures, policymaking and a long-term view on AI development, following certain protocols and processes, and making sure there are mechanisms to alert and address mistakes as they arise.