AI is crying out for regulation, while virologists doing gain-of-function research take the opposite tack. Why?

Over the past few months, prominent tech leaders have been raising alarms about the dangers of AI, and politicians are following suit. Just last week, the Senate held hearings on how to regulate AI. The tech industry itself is calling for regulation: just a few days ago, Microsoft’s CEO testified before Congress and asked the federal government “to govern AI at every part of its lifecycle.”

One of the founders of AI, Geoffrey Hinton, just left his high-level position at Google so that he could criticize AI without any constraints from his employer. And a couple of weeks ago, I asked the AI program ChatGPT if we should trust AI. No way, it told me.

This is all kind of surprising. AI experts seem to be saying “stop us before we do any harm.” It’s also kind of refreshing: usually the private sector wants the government to stay out of its affairs.

Now contrast all this with the behavior of virologists on a completely different technology: gain-of-function research on deadly pathogens. As I’ve explained before, gain-of-function (GoF) research takes a deadly pathogen, such as the influenza virus or the Covid-19 virus, and modifies it to make it even more deadly. Many scientists, including me, find this work both alarming and of little benefit, and we’ve been calling for it to be regulated for a decade now.

However, unlike AI experts, many virologists are opposed to any hint of regulation of their GoF work. On the contrary: just recently, 156 leading virologists jointly authored an opinion piece that lauded the many wonderful benefits of GoF, and pooh-poohed any risks.

Don’t worry your pretty little heads, these virologists seem to be saying to the rest of the world. We know what we’re doing, and it’s not that risky. Plus it’s great! Not to put too fine a point on it, but I disagree.

What’s caught my attention this week is not just the contrast in their willingness to be regulated, but the question of how one might imagine doing it, in both cases.

Simply defining what we mean by “AI” today is probably impossible. The number and types of programs that incorporate some form of artificial intelligence is vast and already affects our lives in many ways. The recent alarm bells were caused by one particular type of AI, known as large language models (LLMs), which have the ability to fool people in a new way. For several years now, more alarm bells have sounded (justifiably so) over “deep fakes,” images or videos that appear real but that are completely made up. These use completely different technology.

So even if we agree that AI needs to be reined in, no one can really say with any precision what that would mean.

Now let’s look at gain-of-function research on pathogens. One of the biggest objections that some virologists have made, on many occasions, is that there’s no way to define just the harmful research, so we really should just leave it all alone.

For example, the recent commentary by 156 virologists said that “gain-of-function approaches incorporate a large proportion of all research because they are a powerful genetic tool in the laboratory.” This is nonsense. It’s equivalent to saying “hey, this is science, and you don’t want to ban all science, do you?”

They also defend GoF by trotting out examples of research that were beneficial, such as the recent rapid development of Covid-19 vaccines. As was pointed out recently in the biology journal mBio, this is a red herring: it’s just not that difficult to define GoF “research of concern” and distinguish it from other, much more mundane virology and bacteriology research.

In fact, biologists have already done this, in a recent set of proposed new guidelines for regulating GoF research. As Hopkins researcher Tom Inglesby put it, “if you are going to make a more transmissible strain of Ebola, then you need to have the work reviewed by the U.S. government.”

So why do the AI scientists say “please regulate us” while many virologists say “leave our gain-of-function work alone”? It’s not because it’s too hard to define one or the other–if it were, the AI experts wouldn’t even consider regulation as a possibility.

No, it seems that it’s all about money. AI is thriving in both academia and industry, with tremendous growth ahead. The people calling for regulation just aren’t worried about money. They know that AI will continue to thrive, and they are calling for regulation because they seem to have genuine concerns about the threat that AI poses to society.

On the other hand, the world of gain-of-function research is very small, and almost entirely dependent on government funding. Although I’m sure they’ll deny it, these scientists are worried that they’ll lose their grants if even a small portion of GoF research is shut down. They may also be worried about more direct threats to their finances: the conflict-of-interest statement on that recent article by 156 virologists goes on for 731 words. (That is one of the longest conflict-of-interest statements I’ve ever seen on a scientific article.)

I decided to ask an AI (ChatGPT) these questions. When asked about regulating GoF, it replied with a long answer that concluded,

“Ultimately, the decision to regulate gain-of-function research involves weighing the potential risks and benefits. Striking the right balance requires collaboration between scientists, policymakers, and relevant stakeholders to establish guidelines, promote responsible research practices, and implement appropriate oversight mechanisms.”

ChatGPT’s answer about regulating AI was similar, concluding:

“Regulation can play a crucial role in ensuring that AI systems are developed and deployed responsibly... The specific nature and extent of regulation will likely depend on the application and level of risk associated with AI systems. Striking the right balance between regulation and fostering innovation is essential to ensure that AI technology benefits society while safeguarding against potential risks and ethical concerns.”

Overall, not bad advice. Now if only those virologists will listen. 

Should we allow AI to control the battlefield? AI itself thinks not.

.

Artificial Intelligence, or AI, seems to have finally arrived. With the introduction of ChatGPT last November, millions of people suddenly discovered that AI was far, far more than just a research activity. The range and sophistication of ChatGPT’s answers to questions across a wide range of disciplines is, frankly, pretty stunning.

AI is already in lots of places where you might not even realize it. Google Translate has been using AI for years, and it’s remarkably good, although nowhere near as good as a human translator. The technology that Pandora uses to customize your music is a type of AI, as is the technology behind Tesla’s self-driving cars. Within my own field, the program AlphaFold2 was a true breakthrough in scientists’ ability to predict the structure of proteins.

Along with these apparently beneficial developments, though, comes a great deal of concern. As AI gets better and better, can we continue to trust it to make decisions for us? It’s one thing if an AI recommends a song that we don’t like, which we can just ignore, but if AI is driving our cars, then mistakes might be much more costly.

And there are far riskier tasks that AI might be doing soon: what if the military starts using AI on the battlefield? One might be tempted to say no, of course they won’t do that, but what if another country starts using AI this way? Can we simply ignore that?

To be honest, I haven’t been worried, at least not yet. I have a little secret to confess: I did my Ph.D. thesis in AI, in the area called machine learning. I haven’t been doing AI research for the past 25 years, but I think I still understand at least a little of the basics behind the technology.

Thanks mostly to the attention that ChatGPT is getting, though, there’s suddenly a huge concern that AI represents a danger to society, although no one is exactly sure why. President Biden had a meeting just this week with the CEOs of Google, Microsoft, and OpenAI (the creators of ChatGPT) to talk about these concerns, where he told the execs that “What you’re doing has enormous potential and enormous danger.”

Of course, politicians have no idea what AI technology is–although I had to laugh at this quote from the White House press secretary: “The president has been extensively briefed on ChatGPT and knows how it works.” Um, no he doesn’t. I’m skeptical that the CEOs of Google and Microsoft know how it works either. However, you don’t need to understand the technology to understand the dangers.

Just recently, one of AI’s early pioneers, Geoffrey Hinton, has been speaking out to warn the public that AI has reached a point where “there’s a danger of something really bad happening.” Not surprisingly, other experts, such as German scientist Jürgen Schmidhuber, disagree.

The chair of the FTC, Lina Khan, says we can regulate AI effectively, and perhaps we can. But I’m not so sure.

Much of the concern in the press seems to be about the use of AI to spread misinformation. For example, AI can help create “deep fakes” that can be used to convince people that, say, an election was stolen even when it wasn’t. That’s a legitimate concern, but it’s happening even without AI.

A better question, which a family member just asked me, is should we be worried about AI making decisions on the battlefield? I thought I should take this question directly to the source, so I asked ChatGPT itself. The answer I got was remarkably clear and direct; I’ll quote it here in full:

“Yes, people should be concerned about AI making decisions on the battlefield. There are several reasons for this.

Firstly, AI lacks human empathy, judgment, and intuition. In a complex and dynamic environment such as the battlefield, human decision-making is often guided by experience, intuition, and an understanding of the broader context. AI lacks the ability to consider these intangible factors and may make decisions solely based on pre-programmed algorithms or data inputs, which could lead to unintended consequences or even catastrophic errors.

Secondly, the use of AI in military decision-making raises significant ethical concerns. For instance, an AI system may be programmed to prioritize the protection of friendly forces at the expense of civilian lives, leading to potentially devastating consequences for non-combatants caught in the crossfire. Additionally, AI systems may be susceptible to hacking or manipulation, which could result in unintended consequences or a loss of control over the decision-making process.

Finally, there is the risk that reliance on AI systems may lead to the erosion of human decision-making capabilities, potentially diminishing the overall quality of military decision-making and increasing the likelihood of errors or unintended consequences.”

So there you have it: ChatGPT itself gives us several reasons why it’s a terrible idea to let AI make decisions that might cost human lives. We should listen to what it’s telling us, because as it gets more intelligent (and it will), it might decide not to be so honest in the future. And I have to say that it’s reasoning here is both well-written and compelling.

One thing that worries me as well is that the very term “artificial intelligence” is much too broad. We should be concerned about giving any technology the power to control our lives, whatever someone calls it. But when an AI itself says “don’t trust me,” then I think we might want to listen.