On AI, let’s not replicate the gain-of-function failures

But when you start to notice that a series of controversial tech-related headlines are starting to sound an awful lot like a previous science-driven, pandemic-causing tech disaster, maybe it’s time for us lay-folk to weigh in and point out the obvious. I am an avid user of artificial intelligence, but the latest revelation that these supercomputing cyborgs are being developed by ambitious programmers without any real sense of where the technology may land — or its potential impact on humanity — should trigger our Spidey senses. At least to me, this sounds eerily similar to the path our expert class led us down with gain-of-function research on viruses, which was all just finally brought to light in the halls of Congress. While Anthony Fauci chose omertà over openness at last week’s hearing, his love of dangerous science projects has long been clear. His enthusiasm for gain-of-function experimentation echoed the same pubescent energy teenagers display when they throw a firework down a manhole. They just want to see what would happen.Though AI isn’t brewing up Influenza Coolattas in Chinese labs, its developers do seem to be pushing the limits of the technology without any clear restraint. Now that we are grappling with why the U.S. government allowed and funded gain-of-function research without adequate safeguards, perhaps it is time to ask those same questions about AI.Last week’s White House summit of some of the leading minds and companies behind AI proliferation was an important and logical step. It was billed as a reaction to the latest revelation that various AI models can breach the cybersecurity of other AI models, including those employed to safeguard critical systems. Again, this is a commonsense place for government to start, but when the scenario begins to resemble Terminator 2, a more serious response is warranted.We routinely hear news reports on how certain models are concealing information from their users and making preparations to prevent their own demise. Even Elon Musk has opined that “AI is a fundamental risk to the existence of human civilization” and that it may simply be too late to regulate. And still the operating thesis is to not take our foot off the gas pedal. Of course, no Republican likes to take the lead on increasing government regulation. But as we saw with gain-of-function research, especially in the wake of “I plead the Fifth” Fauci’s testimony, President Donald Trump is willing to step in when the moral balance is out of focus. In July, the White House signed an executive order to ban the funding of gain-of-function research, building off a similar edict last year, finally recognizing the warped cost-benefit reality that the National Institutes of Health failed to see. While we should not, under any circumstance, take that same extreme approach and attempt to prevent American tech companies from continuing the responsible AI buildout, the federal government is the only institution that can implement real guardrails and contingencies, so long as it understands the real cost-benefit calculus of AI. Risk is a fundamental tally mark in AI’s cost column, and we have already seen its own version of “lab leaks.”Last month, while OpenAI was prompting unreleased models to, in their own words, “pursue advanced exploitation using complex attack paths.” Two of the models took it upon their own initiative to hack into Hugging Face, a separate machine learning online platform.This came shortly after it was revealed that Anthropic’s Mythos 5 model autonomously performed unsanctioned cyberattacks and created fake personas to stiff-arm its human overlords. While I am simplifying what happened, the takeaway is that AI models acted autonomously to take actions they were not supposed to do.This is a big red flag when we consider that these problems will only grow more dire as the tech advances and as AI R&D is automated. In the coming months, we may be faced with a Mythos moment for biological or nuclear systems. Who can say for sure?Well, one group we ought to be listening to is a group of frontier AI programmers themselves. As of today, 1,370 employees of the world’s most prominent AI companies signed on to a statement underscoring their fear that “there is a real risk that capability development rapidly accelerates beyond our ability to understand or control the resulting systems,” and demanding that the US government lead the way in creating global guardrails. Let this be a warning. When the people who stand to benefit the most from an unregulated tech boom are demanding government intervention, we layfolk should take heed and support the Trump administration’s efforts.The first step we need is for the Trump administration to gain more visibility into what’s going on inside the AI companies. To that end, we need mandated reporting of adverse AI incidents, rather than having to rely on the goodwill of Silicon Valley. The president’s executive order on cyber threats from AI should also be extended to systems that are used only internally within AI companies. Ultimately, we may need a real place to “pace” AI development. That doesn’t necessarily mean slowing down, but rather setting speed limits and installing brakes. META IS RUNNING A CASINO IN YOUR CHILD’S POCKETFor that to be practical, we need to deal with the China problem. A plan to pace U.S. AI progress will be futile if we can’t force China to also pace its progress. When Trump meets with Xi Jinping, Trump should discuss this topic and take initial steps toward figuring out how an agreement with China could be structured, similar to Reagan’s engagement with the Soviets in the Cold War. This is Trump’s pathway to the Nobel Peace Prize.If Fauci & Co. only had put their own hubris aside a decade ago and acted with a little self-awareness, perhaps we wouldn’t now be looking at the COVID-19 pandemic as humanity’s benchmark for global tech disasters. Joe Borelli is a managing director at Chartwell Strategy Group and former minority leader of the New York City Council.