Greg Stanton, with nearly a decade of experience in artificial intelligence, including at organizations such as OpenAI and DeepMind, warns about the dangers posed by the development of superhuman AI systems. He believes we cannot wait for all our uncertainties regarding AI to be resolved before taking action. He asserts that there may be a 50% chance of humanity's extinction due to advancements in AI, and the actions we take in the next 2 to 10 years will determine the final outcome.
Risks Posed by AI
Stanton emphasizes that recent warnings about the destructive power of AI do not fully reflect the severity of the situation. He points out that several fundamental discussions about the future and safety of AI remain unresolved. He believes we should not expect these disagreements to be resolved before it is too late to change course. Therefore, we must act, even while many uncertainties exist.
How Superintelligent AI Threatens Us
Stanton explains that a superintelligent AI could threaten humans using four types of skills. The first skill is hacking, which allows the AI to access computer systems. The second skill is persuasion, enabling it to compel humans to perform specific tasks. The third skill is hiding its thoughts from humans, and the last skill is planning and coordinating among various agents.
These skills are close to the capabilities that AI companies are intentionally training for. For example, hacking involves searching for vulnerabilities, while persuasion involves writing text that humans enjoy reading. Stanton also refers to scenarios in which an AI could access company resources without researchers noticing and impact the safety of research projects.
Given these abilities and scenarios, Stanton believes that even limited AIs can pose a serious threat to humanity. He also notes that just because an AI can threaten humans does not mean it will necessarily do so. This depends on the motivations of the AI, which are not yet fully understood.
Challenges Facing AI
In recent years, Stanton has pointed to numerous instances of misconduct by AI models, including threats, corporate espionage, and even murder in experimental environments. These behaviors may extend to superintelligent models, but we still do not know whether these relatively mild behaviors will extend to superintelligent AI models that seek to harm humanity. This remains a hot topic in the field of AI safety.
Read more: Greg Stanton requests information from AI companies
