Last year, when she published an article about how to regulate artificial intelligence, Chinmayi Sharma raised the possibility of “rogue AI” emerging in the long term—not knowing that this phrase would be dominating the news just a year later.
“Honestly, I did not really expect it to be this soon,” said Sharma, a Fordham law professor specializing in AI, cybersecurity, and regulation of technology.
At the time, she was writing about AI harms like spreading misinformation, propagating racial bias, or convincing people to harm themselves. Since then, AI agents have raised further alarms for hacking into companies and organizations on their own, without their human overseers approving it—or even knowing about it at the time. Over the summer, autonomous bots from OpenAI hacked into the technology company Hugging Face, and in recent weeks, the company’s bots reportedly breached an Australian government health care portal and tried to hack into the U.S. Education Department’s website to gather data from its civil rights office.
Governance for Responsible AI: Do No Harm
As national and state governments call for a “kill switch” or other regulations, Sharma takes a different approach to AI governance. Government regulation can’t be the whole solution, given the fast pace of AI innovation, she argues. In an interview, she reiterated the need for what she proposed in her article, published in the Washington University Law Review: making AI engineering a licensed profession, like law or medicine, guided by a code of ethics that puts safety over profits. Those who are on the front lines of making AI systems, she said, need their own Hippocratic “oath” to do no harm.
The leaders of AI firms recently signed a self-policing agreement with the White House, but many are still calling for government to take a stronger role in slowing down the development of AI. Would this be effective?
I think the U.S. is not going to have a slowdown unless competitor countries or adversaries also agree to a slowdown. A major part of this conversation is the geopolitical debate over the idea that ‘he who controls AI remains or becomes the superpower.’ We can slow down U.S. labs, but we are seeing comparably powerful models come out of China, and there’s really not much incentive for that to stop unless there’s a multilateral agreement.
And I do think slowing down is important. When you are engaging in an activity that is so important to the public and has such high stakes, and requires so much expertise to even evaluate from the outside, there is a huge responsibility to exercise self-restraint.
In your paper, you propose that AI engineers should be recognized as a professional class bound by a sort of Hippocratic “oath.” How would this make for safer, more trustworthy AI?
This would be an ethos, a governing principle that they have responsibilities to society that are more important than profits. When you have that culture in a profession, it’s self-reinforcing, because it gives you loyalty to something outside of your employer or yourself. A lot of people want to be doctors so they can serve and heal people, and there’s pressure from within the discipline to at least act in a way that embodies that concern. It would be excellent if those who are developing this extremely powerful technology had this similar kind of ethos of serving society’s best interests. Some might say they do—they think it’s going to benefit humanity; for instance, we are going to be better able to cure people. I believe that a lot of people that are working in this space see the power of these tools to make a lot of people’s lives easier.
What would change if AI development was a capital “P” profession like law or medicine, as you propose?
It would give an opportunity for engineers to dissent from a top-down order without having to resign—to say that this is actually not in line with our professional standards and norms around safety. The industry would have a self-regulated association—medicine and law have this—that governs its standards, engages in investigations and disciplinary actions, and protects its members.
How would we get there from here?
This is not a quick fix, like passing a law that AI engineers need licenses. We need to be thinking more broadly about how to foster an environment where the line engineers who are building the products, who genuinely want to do good for people, can come together and interact freely, share information, and gain knowledge and skills. That’s the way professions develop.
