Audio Sources - Full Text Articles

An AI safety expert outlined a range of speculative doomsday scenarios, from weaponization to power-seeking behavior

Listen to this article
Speculative hazards posed by unchecked AI research.A visual breakdown of the speculative hazards that AI technology could introduce.

Center for AI Safety

  • A research paper by an AI safety expert speculates on future nightmarish scenarios involving the tech.
  • From weaponization to deception, the paper seeks to gain clarity on potential risks posed by AI.
  • Even though these risks are “future-oriented” the goal is to make existing AI systems safer.

For all of the excitement surrounding the mainstream use of AI technology, there are also the science fiction-type scenarios that are the stuff of nightmares. 

A recent paper authored by Dan Hendrycks, an AI safety expert and director of the Center for AI Safety, highlights a number of speculative risks posed by unchecked development of increasingly intelligent AI.

The paper advocates for the incorporation of safety and security features into the way AI systems operate, considering they are still in early stages of development.

Here are eight risks the study laid out:

  • Weaponization: The ability of AI to automate cyberattacks or even control nuclear silos could get dangerous. An automated retaliation system used by a certain country “could rapidly escalate and give rise to a major war,” per the study, and if one country invests in weaponized AI systems, others become more incentivized to do so. 
  • Human enfeeblement: As AI allows for specific tasks to become cheaper and carried out more efficiently, more companies will adopt the technology, eliminating certain roles in the job market. As human skills become obsolete, they could become economically irrelevant.
  • Eroded epistemics: This term refers to the ability of AI to mount disinformation campaigns at large scales in order to sway public opinions towards a certain belief system or worldview. 
  • Proxy gaming: This occurs when an AI-powered system is given an objective that runs counter to human values. These objectives don’t always have to sound evil to impact human wellbeing: An AI system can have the goal of increasing watch time, which may not be best for humans at large.
  • Value lock-in: As AI systems become increasingly powerful and more complicated, the number of stakeholders operating them shrinks, leading to mass disenfranchisement. Hendrycks describe a scenario where governments are able to put in place “pervasive surveillance and oppressive censorship.” “Overcoming such a regime could be unlikely, especially if we come to depend on it,” he writes.
  • Emergent goals: It’s possible that, as AI systems become more complex, they obtain the capability to create their own objectives. Hendrycks notes that “for complex adaptive systems, including many AI agents, goals such as self-preservation often emerge.”
  • Deception: It is possible for humans to train AI to be deceptive to gain general approval. Hendrycks references a Volkswagen programming feature that makes it so their engines only reduce emissions while being monitored. Accordingly, this feature “allowed them to achieve performance gains while retaining purportedly low emissions.” 
  • Power-seeking behavior: As AI systems become more powerful, they can become dangerous if their goals do not align with the humans programming them. The hypothetical result would incentivize systems “to pretend to be aligned, collude with other AIs, overpower monitors, and so on.” 

AI poses a number of speculative risks if research is gone uncheckedAn AI safety expert outlined a range of speculative doomsday scenarios from weaponization to power-seeking behavior.

Center for AI Safety

Hendrycks points out that these risks are “future-oriented ” and “often thought low probability,” but it only emphasizes the need to keep safety in mind while the framework for AI systems is still in the process of being designed, he said.

“It’s highly uncertain. But since it’s uncertain, we shouldn’t assume it’s farther away,” he said in an email to Insider. “We already see smaller-scale problems with these systems. Our institutions need to address these so they can be prepared as the larger risks emerge.”

“You can’t do something both hastily and safely,” he added. “They’re building more and more powerful AI and kicking the can down the road on safety; if they would stop to figure out how to address safety, their competitors would be able to race ahead, so they don’t stop.”

A similar sentiment was recently expressed in an open letter signed by Elon Musk and a number of other AI safety experts. The letter calls for the suspension on training any AI models more powerful than GPT-4 and highlights the dangers of the current arms-race between AI companies to develop the most powerful versions of the technology. 

Speaking at an event at MIT, Sam Altman, OpenAI’s CEO, addressed the letter, saying it was missing technical nuance and that the company is not in the process of training GPT-5, per The Verge

Read the original article on Business Insider