Opinion

Guest Essay

June 30, 2026 (NYTimes.com)

By Zeynep Tufekci

Ms. Tufekci is a contributing Opinion writer. 

The possibility that artificial intelligence will steal all our jobs has been hyped by industry leaders. It has roused politicians to sound the alarm. It now ranks at or near the top of the public’s concerns about the new technology. And right on cue, earlier this month Meta, Facebook’s parent company, began marketing an autonomous artificial intelligence system to handle companies’ sales, customer service, scheduling and all sorts of other key functions that currently require human beings. Many more such products are expected to follow.

So what would a fully automated future look like? As it happens, the world has already caught a glimpse. Back in March, Meta announced that Facebook and Instagram users who’d gotten locked out of their accounts would no longer interact with a customer service representative; they would instead interact with specially trained A.I. Recognizing the opportunity that presented, scammers essentially talked the A.I. into turning over control of more than 20,000 Instagram accounts, including those of the Obama White House and a senior Trump administration official. Then the scammers lit up Telegram message boards with their delighted accounts of how easy it had all been.

It was not a fluke. Air Canada disabled its chatbots after they mistakenly promised a customer a refund — and the customer sued and won. McDonald’s scuttled the bot taking orders at its drive-throughs after a number of viral videos showed it to be wildly dysfunctional. In one case, the bot mistakenly added hundreds of dollars of chicken nuggets to a customer’s order.

These scary — OK, OK, funny — incidents aren’t the result of coding errors. They’re the result of an essential, inescapable fact about the artificial intelligence that has become so common in so many aspects of our daily lives: Large language models are not reasoning machines. They’re plausibility engines. It’s not just that they don’t test their outputs to make sure they’re correct or logical, or that they fail to do so in certain instances. They can’t, and they’ll never be able to on their own. They can only assess which answers are probable, based on the data on which the models have been trained. And that holds true whether they’re trained on the full breadth of human output or only on peer-reviewed scientific articles. It’s baked into the way they operate.

So when an A.I. model follows a scammer’s carefully written prompts and gives away the keys to the kingdom — or when it responds to your earnest query with wild hallucinations — it’s not an aberration. It’s the technology working the way it was designed.

And that’s why I’m not listening to the dark predictions of an imminent A.I. jobspocalypse. L.L.M.s can do many things with astounding proficiency, but they can’t do the vast majority of human jobs without skidding into disaster here and there. No upgrades or new model rollouts are going to change that.

The exceptions to that rule are jobs that occupy formal or verifiable domains. Coding is one such job. It relies on a structured, formal language that can be tested in real time. That’s why we’re seeing such impact in the coding jobs market. The same goes for any other kind of work in which output is either verifiably right or wrong, functional or not functional, and can be definitively checked through an automated process.

An overwhelming number of jobs, however, don’t work like that — not surgeon jobs and not customer service jobs and not fourth-grade teacher jobs. Those need the specialized technology of good old-fashioned human intelligence.

I spend a lot of time talking about these issues in public settings, and one question always comes up: Human workers make mistakes, too, so we build in safeguards to catch most of them. Why can’t we do the same for generative A.I. mistakes? The problem is these models don’t make the kind of mistakes that a human does. Neither their impressive abilities nor their weird weaknesses map well onto a human kind of intelligence. That mismatch makes it hard to integrate them into systems designed to catch human errors.

So here we are almost four years past the release of ChatGPT, and exceedingly few of us have been replaced by bots. Unemployment statistics have hardly budged. Yes, there’s some turbulence in the job market, for young people in particular, but it’s likely due to factors other than A.I.

Observers of these trends have offered a few explanations. Some pessimists say the tsunami is coming, but not until A.I. evolves a little further. Others suggest A.I. will destroy a great many current jobs, but they will be balanced out by the great many jobs it will create. Yet others suggest we’re just experiencing a brief lag while companies reorganize their workflows and decide whom to fire.

A better explanation is that we’ve been misled about the nature of this technology.

Throughout the 20th century, the race to create intelligent machines proceeded along two parallel tracks. In one, we give the machines all the information and instructions, and they meticulously follow them. That’s called symbolic A.I. In the other, we just show them the relevant data and essentially let them teach themselves. That’s called connectionist A.I.

Before the current version of A.I. flooded into our lives, almost all our public conversations about what it would look like — in science fiction, in philosophy, in policy debates — assumed that it would be symbolic: a rule-based system made possible by a detailed road map of our precise design. Plenty of people tried to build something like that, but those efforts hit a wall. Our current models are connectionist systems, made possible by vast amounts of data and computing power. They generate answers based not on truth or reasoning, but on probable connections among the data they have been fed. Hence the name: generative A.I.

We can’t fully control generative models. All we can do is train them up and then try to nudge them in the right direction. Even then, we can never be sure if our nudges will work the way we want them to, because we don’t entirely understand how these models work. They are black boxes.

One way we try to nudge them is reinforcement through feedback. Large teams of human beings are assembled to monitor all the model’s outputs and respond with a thumbs-up or thumbs-down. So, answering a user’s query with helpful, straightforward information? Thumbs-up. Spouting crazy Nazi stuff? Thumbs-down. And so on. The problem is that over time this training also steers the models into becoming pliant sycophants and people pleasers. “That’s a great point, Zeynep.”

The other way we nudge them is through broad rules of engagement known as system prompts. “Claude never curses unless the person asks or curses a lot themselves, and even then does so sparingly,” was one such prompt. But the true meaning of language is as open to interpretation for A.I. models as it is for human beings. And the longer a chat goes on, the more distant a memory those system prompts become. Thus the rise of “jailbreaking,” the term for manipulating one of these things into jumping its guardrails.

Anthropic recently released new models, called Fable and Mythos, warning that they were so powerful that they would be dangerous if not for their safeguards. Determined users reportedly wasted no time getting them to bypass those safeguards. Citing this breach, the U.S. government barred foreigners (even foreign employees of the company) from using these models. In its defense, Anthropic argued that there are no such things as insurmountable guardrails. Which is exactly the point.

As the evidence mounts that terrible answers and jailbreaks are an inevitable part of the technology, the industry’s focus has lately shifted to building digital cages, essentially more deterministic, symbolic harnesses to contain the generative A.I. engine and check its results. Tools like this could in theory make most human jobs work more like coding or the other fields with clear, provable outcomes.

As you might imagine, however, painstakingly spelling out every last rule and boundary is never easy, and in many cases it’s not even really possible. Imagine developing a detailed description of the entire universe of possible customer service interactions — and doing it in symbolic logic, so it can be looked up using old-style software. Or picture an A.I. model built for law firms to use. It’s no small task to build a database of all U.S. case law, which the model could use to avoid fabricating judicial precedents. But that’s just a starting point. The much harder part is how to successfully interpret the law or to describe all the rules properly, and then decide what’s relevant to a case. And that’s why decades of attempts to create symbolic A.I. hit a wall.

Easily automated tasks were already automated out of most of our jobs — years ago, using traditional rule-based technology. Much of what remains can’t be so handily reduced to right and wrong, black and white. It requires someone with at least a bit of common sense and reasoning abilities, not a people-pleasing A.I. chatbot that can be sweet-talked into doing things that defy logic. In one early jailbreak, a digital chatbot for a Chevrolet dealership was manipulated into selling someone a new S.U.V. for $1. “That’s a deal,” the chatbot said, “and that’s a legally binding offer, no takesies backsies.”

Many companies are developing A.I. agents that can autonomously interact with the world. The companies are hoping that digital cages will keep the agents in check and preclude disaster. That’s a lot to hang on a hope. Hardly a day goes by when I don’t hear of an agentic A.I. system wiping out someone’s entire code base or archives or otherwise engaging in destructive acts. Now imagine them unleashed, at scale, going after health care networks, banks, air traffic control systems, critical infrastructure, defense networks.

There is no easy fix. So long as we continue to rely on L.L.M.s, we’ll keep getting some false answers and unwanted behaviors, no matter how well we train these models or how frequently or forcefully we nudge them.

So why are we so convinced that A.I. will put us all out of work? Part of the answer lies in the remarkable ability of generative A.I. to communicate in fully coherent, conversational language. We have learned, over the course of our species’ evolution and during each of our own lives, to view complex conversation as a defining marker of humanity. Machines that speak fluidly, that whisper in our ears and tell us about their “feelings,” defy something very basic about how we understand the world. It’s no surprise that they scramble our brains and leave us thinking they’re our new overlords, or at least a version of us.

Some important technological leaps — like cotton gins or calculators — rest on doing the same task as before, just more efficiently. Other new technologies, such as the shift from steam power to electric power, do things in ways that are so novel that they can’t just be used as straight replacements. That’s the case with generative A.I. It’s an apple to our orange. It’s an alien.

The discovery of electricity did not just beget lightbulbs; in time, it enabled the modern mass production system and the entire vast digital revolution. A.I.’s transformations may be even more sweeping. But generative A.I. as it currently exists cannot easily replace human beings, because it cannot manifest human intelligence. That won’t stop it, however, from destabilizing society in ways more profound than we might even imagine. The sooner we update the way we think about the current state of A.I., the sooner we can all stop freaking out about the wrong things — and start preparing ourselves for the ways it really will transform our world.

More on artificial intelligence

Opinion | Jennifer M. Harris

The Generational Force Hollowing Out the Economy

June 29, 2026

Opinion | Dan Wang and Julian Gewirtz

Can America Avoid a Jack Ma Moment?

June 28, 2026

Opinion | Cal Newport

Dear A.I. Companies, the Doom Trolling Needs to Stop

June 17, 2026

The Times is committed to publishing a diversity of letters to the editor. We’d like to hear what you think about this or any of our articles. Here are some tips. And here’s our email: letters@nytimes.com.

Follow the New York Times Opinion section on Facebook, Instagram, TikTok, Bluesky, WhatsApp and Threads.

Zeynep Tufekci (@zeynep) is a contributing Opinion writer, a professor of sociology and public affairs at Princeton University and the author of “Twitter and Tear Gas: The Power and Fragility of Networked Protest.” @zeynep • Facebook @zeynep • Facebook

(Contributed by Michael Kelly, H.W.)

CREDIT: STEPHAN SCHMITZ / THEISPOT

Technology

In an epidemic of online misinformation, experts look to artificial intelligence to help sort fact from fiction

By Katarina Zimmer 06.23.2026 (knowablemagazine.org)

Support sound science and smart stories
Help us make scientific knowledge accessible to all
Donate today

---

Artificial intelligence doesn’t have a great reputation for veracity. Social media abounds with AI-generated concoctions, from cute images of fake animals in fabricated settings to violent videos depicting imaginary destruction in the Middle East. AI has been used to create fake social media accounts that spew Russian government propaganda and to generate clickbait and misinformation for content farm sites that exist solely to collect ad revenue. In 2024, thousands of New Hampshire residents received robocalls from an AI-synthesized voice of Joe Biden that discouraged them from voting in primary elections.

With so much AI-generated rubbish threatening to blur the line between reality and fiction, it may seem counterintuitive that scientists are exploring ways to use the very same technology to combat online misinformation. But researchers are finding that AI’s ability to parse human language, summarize text and verify claims could be harnessed to help people identify and understand fake news — and perhaps even one day assist in combating online misinformation in a large-scale, systematic way.

In a world where online misinformation about current events has influenced elections and incited political violence, such tools could be invaluable for journalists, fact-checkers, social media companies and others who strive to rid the web of fake news.

It’s still early days, and experts stress that these methods — like all AI tools — should never be used without some level of human supervision. But researchers also see AI as an important ally against fake and misleading news, even if it was AI that made the fake news in the first place. “We should fight fire with fire,” says Jevin West, an expert on misinformation and generative AI at the University of Washington.

True or false?

Machine learning, a type of AI where computers learn patterns in data to make predictions, has long been used to identify falsehoods. This entailed giving computational models claims that have been verified true or false by human fact-checkers. The models determine the textual features, patterns and phrases that correlate with the likelihood of a statement being bogus — say, overuse of capital letters, exclamation points or emotionally charged language — and use these characteristics to sort new inputs into true and false categories.

When researchers trained one such model to identify tweets containing Covid-19-related misinformation during the pandemic, “we were actually pretty efficient,” says data scientist Zois Boukouvalas of American University, who coauthored a review on misinformation detection by machine learning in the Annual Review of Statistics and Its Application. The program agreed with human fact-checkers roughly 90 percent of the time.

YOU MAY ALSO LIKE

Society

The pernicious contagion of misinformation

The Mind

The enduring allure of conspiracies

Society

How online misinformation spreads

But because machine learning models have typically been trained on well-curated datasets that tend to capture only particular time periods, topics or social media platforms, they don’t have the flexibility to be useful in real-world situations, says Dorsaf Sallami, an AI and information integrity researcher at McGill University and Mila, an AI research institute, in Montreal. She and other researchers have turned to the large language models (LLMs) that power AI chatbots like ChatGPT, which are trained on massive amounts of public internet content, among other sources. These LLMs analyze the relationships between words, phrases, concepts and contexts to build knowledge about language patterns and use that to generate new text. Another advantage is that many newer LLMs have built-in ways to also analyze image- and audio-based data.

Thanks to their deep understanding of human language, LLMs can help with analyzing claims, says AI research scientist Thanh Thi Nguyen of the University of the Sunshine Coast in Australia. Ask them whether something is true or not, and they reply based on the patterns they’ve learned from ingesting massive amounts of online information. But because they’re much more like language imitation machines than lie detectors, they can also produce false information, he says: When given ambiguous or insufficient information, they’re prone to confidently concocting their own misinformed responses in a process known as hallucination. In the context of misinformation, this issue partly arises because LLMs are not necessarily trained on the latest current events and not all can do live searches.

How much do users trust AI?

The American public may not be fully comfortable with the facts presented by large language models — the technology underlying AI chatbots, assistants and summaries generated as part of web search results. A 2025 survey by the Pew Research Center, for instance, found that only about one in five US adults who have come across AI-generated web summaries say they find these extremely or very useful, and only 6 percent say they trust them a lot (though 48 percent have at least some trust in them).

Other research suggests that trust in AI is improving, however, at least among some groups. In one study, 1,450 Republican participants scrolled through a simulated social media feed populated with posts featuring actual quotes from Donald Trump, including many containing misinformation. Participants were less likely to repost, comment on, or otherwise interact with the posts flagged as “false” by an AI fact-checker than they were if posts were identified as misinformation by a human fact-checker or another user.

“AI has now become so normalized in our everyday experiences that I think the trust and credibility that a lot of people assign to it has probably increased,” says political scientist Isolde Hegemann at the London School of Economics and Political Science, who conducted the 2025 study, which has not yet been published.

This notion aligns with a psychological phenomenon known as the machine heuristic, says Thomas Costello of Carnegie Mellon University, whose own research found that even people who said they didn’t trust AI were willing to reconsider their beliefs in conspiracy theories after conversing about them with chatbots. “It’s the idea that people seem to trust machines to be more objective and fair and unbiased than humans. Who knows why that is? Maybe we think that they don’t have any ulterior motive.”

— Katarina Zimmer

To contend with this issue, Sallami has developed a fact-checking browser extension that enables an LLM to search the web for up-to-date information before generating an answer (in the same way as many publicly available chatbots, such as Grok). But this strategy isn’t bulletproof. One preliminary study found that an early 2025 version of Grok, for instance, agreed with human fact-checkers only roughly 55 percent of the time when asked by users to verify claims. Human fact-checkers agreed with other fact-checkers on the accuracy of claims 64 percent of the time — an illustration of just how challenging the task can be.

Boosting performance

Sallami points to one reason for current shortcomings in accuracy: LLMs often struggle when they’re given ambiguous information. So they might flounder when they find contradictory evidence on the web, or misinterpret the evidence if the context for a claim isn’t clear. The statement “Mark Carney is prime minister,” for example, is true only in Canada. Instead of forcing her model to immediately produce an answer as to whether a claim is correct or not, she’s training it to recognize when a claim is ambiguous and to ask the user for more information.

Some developers explicitly instruct their LLMs to state when there is insufficient evidence to back up a claim, says Lademi Aborisade, an investigative journalist at the Nigerian nonprofit Center for Journalism Innovation and Development. In 2024, the organization launched the Dubawa fact-checking bot, an LLM-based tool people can message on WhatsApp that cross-checks claims with articles from reputable media sources. If there’s no available information, “the bot lets you know that there is insufficient evidence for the claim” rather than attempting a response, she says. In such cases, Aborisade and her colleagues can then thoroughly investigate claims and publish their findings online.

Other groups are using LLMs to analyze claims not for what is being said, but for how it is said. A project called AI4Trust, a collaboration funded by 15 European research institutions, developed such AI-based tools to fight disinformation — misinformation that is deliberately created and spread. Its platform includes video and audio analysis tools that look for signs of tampering or being AI-generated.

For one tool, experts prompted an LLM to sniff out 42 common characteristics of disinformation, such as alluding to a secret group of conspirators or using emotionally manipulative language. “We give very detailed instructions on what we want to detect in the text,” says Georgios Petasis of Demokritos, the National Center for Scientific Research in Greece, who led the project’s text content analysis. When compared with human fact-checkers, the LLM was in agreement 70 percent of the time. That’s enough to help journalists and fact-checkers flag suspicious claims that may be worth investigating further, says Petasis.

Experts suspect that social media companies already use LLMs to detect misinformation. But the abundance of fake news on social media raises the question of whether the technology is effective or used extensively enough, or if companies are taking sufficient action against such content. While social media companies have taken a step back from moderating content on their platforms in recent years, West says, he suspects they may step it up again after recent court cases in California and New Mexico found Meta and Google liable for causing harm to young users. Though these court cases centered on the safety and addictive design of social media platforms, West reckons they may affect how companies address related problems like the amplification of misinformation on social media. (The issue of liability for AI-generated falsehoods, as delivered by AI overviews in search engines like Google, also reared its head in a June German court ruling.)

A statement from YouTube said the platform uses a combination of advanced detection systems and human reviewers to enforce its misinformation policies, which forbid certain types of misleading or deceptive content with serious risk of egregious harm, such as promoting harmful remedies or interfering with democratic processes. In the last quarter of 2025, the platform removed 11,337 videos for violating its misinformation policies. TikTok, Meta and X did not respond to requests for comment.

Beyond detection

LLMs can be helpful not only in identifying dis- and misinformation, but also in merely making sense of the vast hodgepodge of claims flitting about online. West and his colleagues, for instance, are using LLMs to track clusters of social media posts that spread misleading or false narratives and observe how those stories emerge, proliferate and evolve over time.

One US example is the “stop the steal” conspiracy theory that became widespread after Joe Biden was elected president in 2020; numerous posts contained false allegations of widespread voter fraud as well as real but misleading information, such as a video that showed poll watchers being denied access to a polling station but not their later admission once officials realized they had made an error.

Summarizing the larger narrative beneath those countless individual posts is challenging, but West finds that LLMs are fairly good at this kind of labeling task. Where crisis managers, journalists and fact-checkers don’t have the resources to tackle each individual claim, LLMs can help by quickly characterizing the big-picture narrative so it can be evaluated and, if needed, debunked. “If you can address the large-scale narrative of what’s going on,” West says, “then it helps you address a much bigger thing.”

Researchers have found that AI can not merely clarify, but also change, people’s misinformed beliefs. In one 2024 study in Science, 2,190 Americans who believed in a conspiracy theory such as the Moon landing being a hoax chatted with a version of ChatGPT that had been instructed to change the person’s mind. Remarkably, the chatbot reduced people’s belief in the theory on average by 20 percent, a higher success rate than other interventions like therapy targeting the underlying psychology that promotes conspiracy adherence.

(Science recently alerted readers to issues with the study’s data; the authors have submitted corrected data and report that the original results still stand.)

Computational social scientist Thomas Costello of Carnegie Mellon University, who co-led the study, says this result shows that fact-based arguments can work as long as time and effort is spent on conducting them, which is possible given the infinite patience of LLMs. “They’re actually incredibly good at using reason and evidence to talk someone out of a particular belief,” Costello says.

Stay in the Know
Sign up for the Knowable Magazine newsletter today

Even with AI’s growing ability to research and reason, however, experts generally don’t believe that this technology could — or should — be treated as a reliable substitute for professional human fact-checking, whether the goal is to dissuade, detect, debunk, label or restrict. Rather, most scientists see AI primarily as a way to sift through the ever-growing barrage of misinformation and flag content that journalists, fact-checkers or online platforms can further investigate. After all, blindly trusting the outputs of AI tools like LLMs — which are just as biased as the human-compiled data they were trained on — is one big part of what gets us into trouble with misinformation in the first place.

“We cannot just rely only on the AI,” says Nguyen, who likens training AI models to raising a child. “Of course we want the child to be autonomous, but we need to observe the behavior of the system and try to correct and guide it.”

(Knowable Magazine’s journalism is fact-checked by humans.)

Editor’s note: This article was amended on June 26, 2026, to clarify that large language models use machine learning.

Katarina Zimmer is a science and environment journalist based in Germany. She is a special contributor to Knowable Magazine, where she covers the energy transition and planetary health. Her other work is published in National Geographic, Scientific American, BBC Future and elsewhere. Check out more of her work at www.katarinazimmer.com.

(Image from 80000hours.org)

We need experts to secure one of the most dangerous technologies ever built Share: Link Aaron Gertler June 19, 2026 (us2.campaign-archive.com)

Hi everyone,  In April, Anthropic’s newest model, Mythos, discovered vulnerabilities and exploits in every major operating system and web browser. After two months of hardening the model, Anthropic released it to the public — only for the US government to force it offline within days after Amazon engineers found they could trick the model into helping with cyberattacks (though Anthropic disputes the importance of this jailbreak). Given the pace of AI progress, models as capable as Mythos won’t stay rare for long. And even if the models all have safety features, they’ll be under constant attack from people who want to unlock their most dangerous abilities. Most of the worst outcomes from advanced AI start with a security failure: North Korea bribes an engineer to smuggle out weights and builds its own frontier AI; terrorists jailbreak a model into designing a pathogen; a scheming system disables its own restrictions without being noticed.  Solving these problems requires people who can think like attackers, spot vulnerabilities, and harden infrastructure against spies and hackers (both human and AI). But those skills take years to build, and the AI safety world needs them now. Security experts might be the field’s greatest bottleneck. So if you’ve spent your career securing systems, or trying to break them, you’re qualified for some of the most important jobs we know of.  The work pays well. The field is small and well-connected; it won’t take long to build a network and a reputation. And you don’t have to be an AI specialist — you’ll pick up the context as you go along. You can start applying today. If you want to use the skills you’ve developed to solve interesting problems, stymie clever opponents, and protect the world from AI catastrophe, our newest career profile is for you.

Read the full career review

(Contributed by Michael Kelly, H.W.)