Artificial intelligence is developing at so fast a clip that it can be difficult to keep up with the changes or fully understand what they mean.
This month alone, officials at the two biggest American A.I. companies celebrated breakthroughs in mathematics, on the one hand, and, on the other, warned that the technology might eventually destroy humanity.
Confused? You’re not the only one. We asked readers to send us their questions about artificial intelligence. Nearly 1,000 people responded with queries, big and small.
We’ve selected a sampling of the questions, edited them for length and clarity and done our best to answer them (without using A.I.).
Spencer Layman, Kettering, OHio
What is the point or mechanism whereby A.I. becomes able to act on its own, without human control?
No one knows. At this point, that moment is hypothetical. Today, we have things called “A.I. agents,” and they can do more than just chat. They can edit files, send email, book flights, write computer code and more. But these agents don’t have agency. Humans still have to tell them what to do. And when they do stuff we don’t want them to do, which happens a lot, we can shut them off, just like any other piece of software.
Some people worry, however, that as we start connecting these agents to the internet or even tying them into vital infrastructure, including power grids, stock markets and military weapons, they will wreak havoc in ways we do not expect. Separately, people are worried that researchers will one day develop A.I. systems that do have agency. But it is unclear how or when that would happen.
Dale Kuropat, Wickenburg, ARIZ.
How might A.I. threaten the existence of humanity?
As people warn that A.I. could destroy humanity, they are light on examples of how. They say that malicious people could use A.I. to develop a bioweapon. And they worry that as companies and governments integrate A.I. with vital infrastructure, the systems could end up hoarding the world’s economic resources and depriving people of food, housing and other things we need to live. But those scenarios are highly speculative.
They also worry that once A.I. is hooked into the world’s infrastructure, it could start a global drone war or set off a nuclear bomb.
Even the risk of someone using A.I. to unleash a biological attack is smaller than it might seem. But for the doomsayers, the worry is that A.I. makes this easier.
Emily Wright, Birmingham, ALA.
What guardrails are in place, and how do they work?
Companies like OpenAI, Anthropic and Google have already put guardrails on their A.I. systems. If you ask a chatbot like ChatGPT to help you build a bioweapon, it will probably refuse. If you try to hack into a computer network using Anthropic’s most powerful technology, it will probably clam up. Anthropic may even shut down your account.
But the guardrails placed on these systems are flawed. Researchers in Italy recently discovered that they could break through these kinds of protections with … poetry. They used flowery language to trick 31 A.I. systems into ignoring their own internal safety controls.
This loophole is just one of many methods that allow hackers and other malicious people to bypass guardrails on the leading A.I. systems. All the top A.I. companies use the same basic techniques to build protections into their systems, and they are surprisingly easy to break through.
Companies like OpenAI and Anthropic can push these systems toward what they see as good behavior. But because A.I. systems learn from enormous amounts of digital data, these systems can still do things that even their creators do not expect.
Ken Miller, Tacoma, WASH.
What would regulation of A.I. look like?
Politicians, tech executives and pundits have floated many possibilities. California state legislators, for instance, tried to institute a law that would require companies to test the safety of their technologies before releasing them. One version of the bill would have given California’s attorney general the power to sue companies if their technologies caused serious harm, such as property damage or any human casualties.
More recently, lawmakers and policymakers have suggested that the federal government should review A.I. technologies before they’re released. In June, President Trump signed an executive order creating a framework for government reviews, but this program is voluntary. Recently, tech leaders have suggested that independent evaluators should be embedded inside their companies.
But there is no agreement on the best way to proceed. Many people believe that freely releasing technologies is actually the safest path forward. The argument is that while people can use A.I. to do harm, they can also use it to prevent harm. The leading Chinese companies take this approach, which is called open source.
Deena Shoshked, NEW JERSEY
How realistic is it to create a viable “kill button”?
Proposed bipartisan legislation such as the “A.I. Kill Switch Act” would mandate that the leading A.I. labs establish a way to instantly shut down their systems. Experts say these measures will improve A.I. safety, but that they are not sufficient on their own to prevent dangerous scenarios.
For one, A.I. companies need to become better at monitoring capabilities to understand when their systems have gone off the rails. In one recent incident, OpenAI’s agents hacked multiple organizations, including the software company Hugging Face. It took OpenAI researchers months to learn that their A.I. systems had gone rogue — and we are still discovering the extent of the damage.
Experts also worry about more advanced hypothetical scenarios, such as an A.I. system duplicating itself into other computers like a virus, or bad actors using open-source models built without kill switches. Security engineers are working to invent shut-off mechanisms in the underlying computer chips that power A.I. technology, but these efforts will require global coordination, and likely take years to implement.
Daniel Wang, Seattle, WASH.
How will we know if A.I. achieves “superintelligence”? And could it cure cancer or end world hunger?
You make an important point. We don’t even know how to measure human intelligence. How can we say we know how to build “superintelligence”? Even if we can build it, how will we know when we get there?
The truth of the matter is that machines are already more powerful than the human brain in many (obvious) ways. This was true of the pocket calculator in the late 1960s. It could perform arithmetic far faster than we humans could ever hope to.
Many experts say that you should think about A.I. as a “jagged intelligence.” It is racing ahead of us in some areas — as in math, and computer coding — but still lagging in other ways. It does not have common sense. Unlike us, it does not understand the physical world.
Can it cure cancer? Well, various companies and academic labs are certainly developing technologies that can accelerate scientific discovery. These technologies can help develop new drugs or create new materials. The leaders of top Silicon Valley companies love to say that A.I. can cure cancer or solve world hunger, or even bring about world peace, but at this point, this is just talk.
Let’s not forget how difficult it is to cure cancer. Though A.I. can accelerate drug discovery, that is just one small part of the process. And as much as A.I. can seem like magic, it is not.
Mick Parent, Bayfield, WIS.
What types of jobs are most at risk of being eliminated because of A.I. and which jobs are safer?
As recently as the early 2010s, experts assumed that A.I. would replace blue collar workers first. Their assumption was that robots — physical robots — would start doing manual labor. But it turns out that today’s A.I. is far more likely to replace white collar workers, including writers, graphic designers, accountants, personal assistants and, most notably, computer programmers. Technologies from companies like Anthropic and OpenAI are really good at writing computer code.
In Silicon Valley, some start-up companies are using A.I. for all those tasks, keeping their hiring to a bare minimum. But experts disagree on how all this plays out.
Dennis Kornbluh, Lincoln, NEB.
If A.I. can replace a team of software developers, shouldn’t it cost more to use?
Right now, A.I. companies like Anthropic and OpenAI are spending far more money than they are taking in. Building and operating the leading A.I. systems requires hundreds of millions of dollars in computer chips and electrical power.
This is so expensive that many people worry the whole operation could come crashing down. The tech industry and its partners have taken on billions of dollars in debt to finance the giant data centers needed to power A.I.
For the foreseeable future, the big A.I. companies will keep spending. They believe that the technology will become so powerful, they will eventually make trillions of dollars.
Natalie Leavenworth, Urbana, Ill
Do we have to use up all, or most of, our drinking water to help power A.I. data centers?
Certainly not all of it! But even experts are struggling to pinpoint exactly how much water data centers — the buildings that house the hardware running an A.I. — actually use. That uncertainty can make it difficult for communities, especially in drought-prone regions, to plan ahead and curtail data centers’ water use when needed.
Data centers in many localities aren’t currently required to report their water usage. Still, we know that data centers use water in many different ways, including for cooling high-powered servers.
Such activities can add up to a lot of water. For example, one data center in Council Bluffs, Iowa, reportedly used over one billion gallons on-site in 2024. And depending on where data centers get their water — whether from streams and reservoirs, underground aquifers, city taps or recycled wastewater — a lot might be coming from the same supply that you drink from.
Jackie Sinatra, Fort Collins, COLO.
I use ChatGPT as an app on my phone and iPad. Can the A.I. go into my folders and retrieve passwords or my bank account information?
It cannot. Although you can choose to upload files to ChatGPT, and you can link it to other apps like Zillow or DoorDash, allowing you to, say, browse real estate listings or order food.
Some people, however, are going even further. In Silicon Valley, for instance, many people have started using what are called “A.I. agents.”
These systems are based on the same technology as a chatbot like ChatGPT, but they can also operate as a kind of personal assistant, doing research, editing files, sending messages and more. In some cases, people are willingly providing these systems with access to their folders, email services and, believe it or not, their credit cards.
This is not recommended. A.I. agents have been known to suddenly delete files, and spend large amounts of money in unexpected ways.
Richard Orr, Columbus, OHIO
What is it about A.I. that makes it so much faster and more comprehensive than a simple Google search?
Today’s A.I. systems are based on what are called “neural networks,” mathematical systems that learn skills by analyzing vast amounts of digital data. By pinpointing patterns in thousands of cat photos, for instance, a neural network can learn to recognize a cat, or even generate a new cat photo.
To build a chatbot like ChatGPT, OpenAI gathered up just about all of the text on the internet and fed it into a neural network. By finding patterns in all that text, including Wikipedia articles, news stories, chat logs, digital books, even computer programs, the system learned to generate text on its own, piecing together words a lot like people do.
But it also learns the relationships between all those words. And it learns how people have answered questions in the past. So that when you ask one of these questions, it instantly knows how to respond.
Sometimes, it gets things wrong. This is called hallucination. But companies have worked to reduce these mistakes.
All this means that A.I. chatbots are faster and more comprehensive than an old fashioned Google search. But bear in mind: they still make mistakes.
The post Will A.I. Kill Us? Can It Hack My Bank Account? Your A.I. Questions Answered appeared first on New York Times.
