Saving AI from itself

This spring, a group of industry leaders put out a dire warning that AI could pose an existential threat to humanity if the proper safeguards aren’t in place. 

Nikhil Krishnaswamy, an assistant professor of computer science, said he thinks it’s more likely that tech execs are “being so wowed by their product that they’re perhaps blind to what it can’t do,” and that while some concerns are valid, the technology isn’t at a level where it can eliminate humans entirely. 

“In life-critical scenarios, like launching a military attack, we’ve built in a ton of safeguards into AI processes to ensure a bad actor can’t trigger some disastrous event,” he said. “If AI is in the system, we need to continue to build safeguards that protect us.” 

Krishnaswamy said at this stage, humans are involved in every part of the AI process, which means we need to take a close look at what it is and isn’t allowed to do. Secondly, he said everyone’s voice needs to be in these conversations. 

“We can’t just allow a few companies to decide what’s best for AI development and deployment,” Krishnaswamy said. 

He added that one of his key concerns is that conversations around existential threats are obscuring some of the very real issues that are occurring when AI is used unwisely, such as spreading misinformation or even propagating racism in the medical system. 

Qahri-Saremi echoed these concerns, saying that generative AI has the ability to create human-like, convincing claims along with an innate tendency to “hallucinate.” 

“As these technologies are increasingly adopted for content generation on social media, such as Snapchat’s My AI chatbot, their hard-to-detect hallucinations can further increase users’ likelihood of falling victim to false information,” he said. 

AI’s potential risks to humanity are a concern, but another concern is how AI might harm itself as its use becomes more prevalent. 

Qahri-Saremi said recent reports suggest a drop in the accuracy of ChatGPT outputs. While there are numerous potential culprits, one that stands out more than others: “AI cannibalism.” 

This can occur when generative-AI models like ChatGPT are trained on other content created by artificial intelligence rather than humans. 

“Human creativity has fed the generative AI and shaped its learning,” Qahri-Saremi said. “However, as humans are more dependent on generative algorithms for their creative works, an important source of AI learning is fading away, and eventually AI algorithms will have to feed on other AI-generated contents, which can be affected by biases, falsehoods and a lack of context.” 

Qahri-Saremi said this is important because it shows that people are needed to be kept in the loop to make these technologies work. 

“Instead of being concerned about whether AI will replace humans, the important question is perhaps how to determine the best ways for AI-human collaboration to take shape,” he said.