With the rise of LLMs, what should we really be concerned about?

The arrival of large language AI models like ChatGPT has triggered debates across academia, government, business and the media. Discussion range from their impact on jobs and politics to speculation on the existential threat they could present to humanity.

Michael Wooldridge 打開新的分頁／視窗, Professor of Computer Science at the University of Oxford, described the advent of these large language models (LLMs) as being like “weird things beamed down to Earth that suddenly make possible things in AI that were just philosophical debate until three years ago.” For Michael, the potential existential threat of AI is overstated, while the actual — even mortal — harms they can already cause are understated. And the potential they offer is tantalizing.

What is the real risk of AI?

Speaking in advance of delivering one of the Royal Institution Christmas Lectures 打開新的分頁／視窗 December 12, Michael said concerns around existential threats 打開新的分頁／視窗 were unrealistic:

In terms of the big risks around AI, you don’t have to worry about ChatGPT crawling out of the computer and taking over. If you look under the hood of ChatGPT and see how it works, you understand that’s not going to be the case. In all the discussion around existential threat, nobody has ever given me a plausible scenario for how AI might be an existential risk.

Instead, Michael sees the focus on this issue as a distraction that can “suck all the air out of the room” and ensure there’s no space to talk about anything else — including more immediate risks:

There’s a danger that nothing else ever gets discussed, even when there are abuses and harms being caused right now, and which will be caused over the next few years, that need consideration, that need attention, regulation and governance.

Michael outlined a scenario where a teenager with medical symptoms might find themselves too embarrassed or awkward to go to a doctor or discuss them with a caregiver. In such a situation, that teenager might go to an LLM for help and receive poor quality advice.

“Responsible providers will try to intercept queries like that and say, ‘I don’t do medical advice.’ But it’s not hard to get around those guardrails, and when the technology proliferates, there will be a lot of providers who aren’t responsible,” Michael said. “People will die as a consequence because people are releasing products that aren’t properly safeguarded.”

That scenario — where technology proliferates without guardrails — is a major risk around AI, Michael argued. AI itself won’t seek to do us harm, but people misusing AI can and do cause harm.

“The British government has been very active in looking at risks around AI and they summarize a scenario they call the Wild West,” he said. In that scenario, AI develops in a way that everyone can get their hands on a LLMs with no guardrails, which then becomes impossible to control.

“It puts powerful tools in the hands of potentially bad actors who use it to do bad things,” he said. “We’re going into elections in the UK, in the US, in India, where there is going to be a really big issue around misinformation.”

What is the grand challenge — and how can we address it?

Michael summarized the challenge as: “How do we support people who want to innovate in this field, while at the same time avoiding this technology proliferating in such a way that it becomes impossible to govern?”

There are no easy answers immediately available, but Michael noted that more could be done by social media companies to implement systems that spot gross misinformation and prevent it from propagating. “The usual counter-arguments are that if you try and address this, you’re stifling freedom of speech,” he said. “But when there are manifest falsehoods being spread, I think there is an obligation for social media companies to be doing more.”

Finding the balance between preventing harm and enabling innovation is essential because, as Michael pointed out, LLMs are a fascinating area for researchers with a lot of potential:

The arrival of these models has just been this supermassive black hole that has twisted the whole fabric of computing. And all of science has been moved by this enormous presence.

Michael noted that all 10 of the research groups in his university department have been affected by the advances in LLMs: “In some instances, it’s re-written their research agenda; in others, they’re wrapping up because the work just isn’t relevant anymore.”

For Michael personally, multi-agent systems are of particular interest, where multiple AI systems with competing or complimentary goals interact with each other to solve a problem that would elude a single system.

“That really pushes my buttons,” he said. “Large language models represent a really tantalizing opportunity there — this idea of having them interact with each other and not necessarily doing it in human language.

“So, for example, one idea is that you can deal with hallucination 打開新的分頁／視窗 by having large language models that are essentially in a competitive scenario with one another. One model is coming up with copy and the other is critiquing it, and the idea is that the process ends with them in some kind of agreement on a factual statement.”

In the face of the kind of seismic change these AI models represent, Michael sees science communication as essential. As co-Editor-in-Chief of the Elsevier-published journal Artificial Intelligence 打開新的分頁／視窗, he is well versed in communicating about research among researchers. The Royal Institute Christmas lectures, meanwhile, provide a platform to communicate facts about AI more broadly.

“That’s very prominent on my agenda and has been for several years,” he said. “With all the discussion around AI, I see it as essential to try and inform the public about what AI is. It’s part of the science. If I accept public funding for my work, I have an obligation — if this becomes something people are discussing — to stand up and talk about it.”

In particular, Michael talks about addressing the public misconception that AI has its own intent or its own considerations:

One of the big misunderstandings is that people imagine there is a mind on the other side of the screen, and there absolutely is not. An AI doesn’t contemplate your question. When you understand how they work, even at a superficial level, you realize it’s just a statistical algorithm. It’s a very cool and impressive statistical algorithm, but it doesn’t think or consider. Some people are surprised to learn that there’s no ‘mind’ there at all.

That misconception can be fueled by the language around AI — that it “looks” for information, that it can be “tricked,” or that it “wants” to provide a certain kind of answer.

“We use that language because it’s convenient,” Michael said, “but the danger in anthropomorphizing AI is that we read far more into it than is actually there.”

Despite the storm of discussion, catastrophizing, misconception and potential for misinformation, Michael is enthusiastic about AI from a research perspective:

It’s such an interesting development. We’ve got really powerful tools, and we’re just starting to explore their dimensions. These tools are weird, and we don’t understand exactly why they go wrong in certain ways and what their capabilities are. Mapping that out is a fascinating journey.