Recent concerns about the safety of AI may be overblown, according to multiple Pitt professors.
In recent weeks, a former Anthropic employee expressed concerns in his resignation posted to X that AI development could get out of control and lead to disastrous results. The incident has sparked conversations about the potential dangers and regulation of AI.
Patrick Skeba, a teaching assistant professor in the Department of Computer Science, said fears of AI being an “existential threat to humanity” are “complete nonsense.”
“Saying that AI is dangerous implies that it can somehow do things that we don’t ask it to,” Skeba said. “When they say things like, ‘What if it brought us to war? What if it did all of these things?’ Well, why would we give it that capability?”
Adam Zweber, a teaching assistant professor in philosophy, said fears of AI destroying the world are not new, citing the existing “probability of doom” metric, colloquially known as P(doom).
“There are lots of ways in which things can go very bad, even if AI isn’t that much smarter than us,” Zweber said. “There are lots of risks that can lead to very bad outcomes.”
Zweber explained that people within companies like Anthropic or OpenAI work to prevent the possibility of catastrophic events.
“In principle, you could build an AI without guardrails,” Zweber said. “But there are a lot of people working on this — a lot of very smart and very concerned people who are coming up with ways to prevent it from just telling people how to do bad things.”
Skeba said situations like the Hugging Face incident don’t point to a disaster being caused by AI. In these situations, AI agents are given internet access and break out of “poorly configured sandboxes,” according to Skeba.
“You basically gave them the key and said, ‘Promise not to use this,’ but they used it anyway,” Skeba said.
Skeba said there are reasons to be concerned about the potential usages of AI, but these worries lie in human, rather than artificial, decision making.
“[AI] can have some really unjust and terrible results,” Skeba said. “But to blame that on the AI doing that of its own accord, as opposed to saying, ‘Why are we taking human judgment out of this loop and replacing it with AI?’ That’s very different from the AI somehow choosing or deciding to do things that humans don’t like. That’s science fiction in my opinion.”
Discourse over regulating AI is another common topic for those in the industry. Michael Madison, a professor of law, said AI is different from other technologies that came before because of its large scale, which affects its ability to be regulated.
“It’s not something that is happening only in the United States,” Madison said. “So the idea of regulation in the abstract is complicated by the fact that it’s not obvious who should do the regulating.”
Madison said another struggle of regulating AI is the variety of ways that it could be regulated. Regulating areas such as development, usage or software are all pathways that provide separate and distinct options.
“These are all different elements of the ‘what might be regulated’ equation,” Madison said. “And each of them has different strengths and weaknesses in terms of their effectiveness, in terms of whether that type of regulation matches the kinds of risks that have been identified.”
Skeba said from a technological point of view, regulation has been needed for a long time, and in some instances, it has taken too long.
“A lot of what I care about it’s already too late for,” Skeba said. “I would have preferred that we didn’t allow these bots to scrape the entire Internet in the first place because I think there’s probably a lot of information that we don’t want in there.”
As for the future, Skeba said there should be regulation for when AI decision making is taking place, such as in situations where AI is taking medical notes and could be incorrect.
“That needs to be very open,” Skeba said. “If people are dealing with a system [where] the decision is made by some autonomous bot and not an accountable human, I think that’s the kind of stuff where regulation has to be very important.”
In recent weeks, the federal government has made agreements with AI founders asking them to self-regulate their technology. Skeba said he doesn’t believe that self-regulation will work due to the scale and lack of knowledge about AI systems.
“The people who create [large language models] don’t entirely know how they work because there are these models with trillions and trillions of parameters and data gaps from all over the world,” Skeba said. “Through no intentional lying, I don’t think they have the capability to control it in the way that they’re saying.”