Artificial intelligence | Can we really control AI?

There are increasing numbers of expert outings about the risks of artificial intelligence (AI) as it promises to transform our future. So much so that teams from all over the world are pushing the machine to its limits to limit overflows. “Hell is paved with good intentions,” warns a researcher.

Lined by machine

What if AI came into our lives without warning?

For example, a business owner could buy an AI assistant and give it simple tasks, like writing emails and suggesting purchases.

Impressed, our CEO could decide after a few months to “promote” AI.

“Instead of composing emails, AI would now have control of the inbox,” writes Dan Hendrycks, director of the Center for AI Security and a graduate student in AI at UC Berkeley, in a scenario that he recently elaborated for the magazine Time.

Rather than suggesting purchases, the AI ​​could then make them, and pay for them by having access to company bank accounts.

Our CEO is not crazy: he would carefully monitor the work of this virtual assistant. But months would pass, and the AI ​​would become more efficient than human employees. And companies that use AI too little will fall behind.

“The CEO could then decide to have the AI ​​design the next model of a product line, or plan a new marketing campaign. »

The result: large swaths of the economy would be run by AI corporations, and humans would realize too late that they had lost control.

In interview with The PressDan Hendrycks explains that this “Darwinian” scenario, where the AI ​​takes up more and more space because it is more talented than the human, is less futuristic than it looks.

In this race, how do you avoid being overtaken by the machine?

The red team method

The advances in AI are mind-boggling. They also have the power to scare.

Since the launch of ChatGPT last November, many experts have sounded the alarm bells. Threatened jobs, loss of control, increase in social inequalities: the list of risks and questions raised by AI is growing every week.

Laurent Charlin, senior academic member at Mila-Quebec Institute for Artificial Intelligence, associate professor at HEC Montreal and holder of a Canada-CIFAR AI Chair, notes that even the AI ​​community has been shaken in recent months.

One of the most pressing issues is ensuring that models like ChatGPT are secure, he says. In short, don’t advocate for ethnic extermination, glorify suicide, or start providing the steps to follow to create the next pandemic.

OpenAI and Google, two of the companies behind these models, notably use the red team method (red teaming) in an attempt to control the responses given to users.

Concretely, a red team (red-team) is a group made up of international experts in several fields who are given privileged access to the model before it is released to the general public, and who try to push it to its limits. This makes it possible to identify, for example, racist or violent responses, or hallucinations of the model.

Roya Pakzad, a technology and human rights researcher affiliated with the CITRIS Policy Lab at UC Berkeley, California, is among the group of researchers who worked to analyze responses from ChatGPT 4 before its launch.

“During the tests, I used various strategies, such as emotional manipulation, adopting human characters, and language tests in English and Farsi,” says Ms.^me Pakzad in interview with The Press.

The Farsi responses contained more stereotypes and hallucinations (made-up events and names), she noted.

Unjustified concerns?

Wassim Bouachir, professor in the Science and Technology department of TELUQ and specialist in AI, notes that the wind of concern that is perceptible about AI is not justified, according to him.

We are still far from a general system of artificial intelligence, capable of learning as a human can. “We don’t see that today, specialists don’t see it in the future,” he said.

Laurent Charlin is also of the opinion that the questions of the dangers of an AI of the future are hypothetical. He notes that controlling today’s AI should be our priority.

“It’s a bit like you trained doctors, they went to school to make sure they had a positive effect on society, and they are given exams, they are also asked not to doing evil. I’m not saying that we should do exactly the same thing with AI, but there could be a system of evaluation, of formal accreditation, ”says Mr. Charlin, who foresees a classification system for AI. , a bit like the classifications of films according to the age of the public.

“For example, a model chatting with an adult would not be the same as a model chatting with a child. We could have a model in Quebec, but which would not be the same as a model in Israel, and which would reflect the fact that each society is different,” he illustrates.

Three great actors

For Dan Hendrycks, it is true to say that the dangers of AI are largely hypothetical, while everyone is mostly trying to see how AI can help them in their work.

“But at the rate research is advancing, we need to act now rather than waiting until it’s too late. Today’s AI is a bit like a lion cub: it’s cute and doesn’t have much impact on its environment. But lion cubs don’t stay babies for long, and that’s what worries me. Hell is paved with good intentions… ”

Hendrycks is among dozens of researchers and public figures, including Bill Gates and OpenAI founder Sam Altman, who recently signed the following public statement: “Mitigation of extinction risk [humaine] related to AI should be a global priority, along with other societal risks such as pandemics and nuclear wars. »

The time has come for countries to collaborate to create beacons to control AI, such as military-grade information security, internal audits, do and publish security research, says -he.

“There are three big AI players in the world: the United States, China and the United Kingdom. AI must not become a new arms race, as happened with nuclear weapons. If that happens, it could lead to a full-scale disaster. I think all three have to sit down and find a way to control the development of this technology. And the time to do it is now. »

Is it too late to control the AI? Probably not, but the days when it is still possible to do so are numbered.

A pirate in every cottage

Dan Hendrycks, director of the Center for AI Security and a PhD student in AI at UC Berkeley, is among the researchers who argue that humanity cannot afford to go wrong with artificial intelligence.

He wouldn’t be surprised, for example, if within a year AI was able to do hacking. “AI is already able to write code, so I don’t see why hacking would be so much more difficult,” he said in an interview. If this happens, everyone would have access to their personal robot hacker… We’re not talking about the extinction of the human race here, but it could be catastrophic for various vulnerable infrastructures around the world. »

Mr Hendrycks can imagine that an untrained AI could become an expert in virology within 10 years after reading all the published studies and material on the subject.

“There are currently 30,000 PhDs in virology, and an extremely small proportion of them could potentially create and release a biological weapon and create a pandemic. But if it is possible, for example, to unlock ChatGPT 4 and teach it how to do it, then the proportion of malicious people who could cause harm increases considerably. »

According to him, it is in the field of defense that it will be most urgent to regulate AI.

“If the opposite side has an AI that makes decisions in the blink of an eye, you must also have an AI that does the same thing, and even faster. What happens if the AI ​​sees a threat where there is none? Is this the end of our species? Maybe not, but we can see how things could get catastrophic very quickly. Hence the importance of supervising AI now. »