AI is Changing Science: Here’s What Journalists Need to Know

 

Antonio José Paz Cardona, a senior editor for Mongabay Latam, is not a tech reporter. For the last decade, the Bogotá-based journalist has focused on covering environmental issues in Latin America, such as wildlife conservation and deforestation. But over the past few years, something new has crept into the stories he’s been writing and editing: AI.

Scientists using AI in ecology “really caught my attention,” he says. “I [wanted] to see how AI is helping scientists.”

Paz Cardona organized a special collection of reports dedicated to AI in conservation. His team unearthed six stories from across Latin America, from researchers using AI to fingerprint the songs of individual birds in Argentina to an AI-powered app for identifying stolen seashells at Costa Rica’s airports. “There were a lot of things that impressed me,” says Paz Cardona. He collated the findings in a feature surveying the promises and pitfalls of AI in conservation research, including the environmental impacts of data centers.

Paz Cardona’s experience encountering AI in his beat isn’t an outlier. AI adoption among researchers rose from 57 percent in 2024 to 84 percent in 2025, according to a survey by Wiley. And an August 2026 preprint reported that almost 90 percent of biomedical papers published in December 2025 and archived in PubMed show signs of AI assistance. Companies like OpenAI and Google are developing bespoke tools for scientists. Just under four years after the release of ChatGPT, researchers are using AI not only to analyze and produce data but as digital research assistants for everything from writing papers to planning and executing entire studies.

Amid this rapidly shifting landscape, journalists need to cover the influences of AI on science accurately and responsibly. That means learning what AI is—and it is more than one thing—and understanding the basics of how it works. Different AI technologies have different strengths and weaknesses, and their environmental and social impacts can be worlds apart. To understand whether AI is actually enhancing science, reporters will have to rely on their beat-specific knowledge and contend with private companies’ red tape.

“AI is touching everything,” says Matthew Hutson, a freelance science journalist who’s covered AI for decades. “So people who weren’t AI reporters or even science or tech reporters before are [now] writing about AI.”

 

Know Your AI(s)

Many of the metaphors that hover around AI—like neurons and learning and hallucinations—sound biological. But under the hood, AI is just code. And lots of math.

“It’s not magic,” says Eva Wolfangel, a science and technology journalist who covers AI for Germany’s DIE ZEIT and other outlets. You don’t need to be able to do the math yourself to grasp AI’s key principles. Instead, “try to understand on the meta-level what AI and machine learning [are] doing,” she says. Knowing the different forms of AI and their key functions will help you frame findings appropriately for your readers.

The explosive rise of ChatGPT since its release in late 2022 made AI a dinner table topic. But AI isn’t new in science. Researchers have long used machine learning to find patterns hidden in complicated data for classifying or categorizing things. Such models are known as discriminative AI. There are many nuances, but here’s how they work in general: Researchers present a model with sample data during a training session, and the model makes guesses about patterns in the data, such as whether an image contains a cat or a dog. The model then gets feedback, often by comparing its output with how humans labeled the data. Over many rounds of training, the model improves its accuracy. Then, when shown new data, the model uses what it has learned to classify them.

Many discriminative AI models, like those used by the ecologists in Paz Cardona’s story to identify birds and seashells, are artificial neural networks—a type of model composed of interconnected digital “neurons” that each perform a simple computation. These neural networks compute by passing information from neuron to neuron and learn by adjusting the strengths of those connections. This loosely mimics biological neural networks, which strengthen or weaken synaptic connections as they learn.

Freelance science writer John Pavlus recalls covering a neural network model called BERT for Quanta back in 2019. BERT was trained on text data and developed the ability to identify grammatical sentences and select sensible answers to questions. Soon, models would arise that could not just identify and analyze sentences but generate them, too.

Such models, the most powerful of which are very large, complex neural networks, are called generative AI. Generative models learn the patterns in training data to produce new data. This process typically requires more training data and computational resources than discriminative models. Large language models (LLMs) like GPT (a family of models including the ones behind ChatGPT) and Claude, advanced relatives of BERT, are a particular kind of generative AI. As their name suggests, the signature ability of LLMs is language. While discriminative AI can classify a poem as happy or sad, an LLM can write an entire poem.

Knowing about different kinds of AI gives journalists a leg up in their reporting—even if technical details don’t wind up in their final stories, Pavlus says. “You’ll have a little bit of extra muscle power for just knowing what to ask, what to question.”

 

LLMs Everywhere

Having at least a basic familiarity with AI is becoming essential, as researchers are now using generative AI, especially LLMs, in almost every field of science.

Social scientists are synthesizing large volumes of data on everything from education to health outcomes to inform life course research. Qualitative researchers are using LLM chatbots to conduct interviews at scale. And cognitive scientists are using LLMs as a kind of alien lab rat to study the brain—especially language cognition and development, since there is no animal model for language. Researchers are even using LLMs as stand-ins for humans in behavioral and social experiments, though not without problems. And in biology, scientists are training such models not on human language but on the “letters” of protein and DNA sequences to produce programs for analyzing genomes.

But the impact of LLMs on science runs deeper than any one project. Because the models can work with text, data, and code flexibly across many contexts, scientists are incorporating them into practically every stage of research. “Scientists use AI to help write papers, to write code or brainstorm ideas, or to design experiments,” says Hutson. A 2025 survey found that more than half of scientists are already using LLM-powered for peer review. In theoretical physics and mathematics, researchers are using LLMs to work through ideas, generate conjectures, check proofs, and search for counterexamples.

In some cases, the increasing capabilities of LLMs are triggering feelings of existential dread among researchers. Joshua Sokol reported for Science in June 2026 that some astronomers worry the technology will “lead to nothing less than the death of astrophysics as a human endeavor.” Similarly, Pavlus reported for Quanta Magazine in 2025 that LLMs “broke an entire field” of research—natural language processing—and forced scientists to find new research questions. Mathematicians, too, are grappling with the future of their field in the wake of OpenAI’s announcement that its model solved the Navier-Stokes problem, an enduring mathematical problem with a $1 million prize attached, quickly followed by the news that the same model has solved hundreds more.

The Navier-Stokes result in particular was the work of LLM-based AI “agents,” which perform tasks autonomously by writing and executing instructions for themselves and other linked programs. “Agents have really blown up in late 2025, early 2026,” says science and tech journalist Peter Hall, who covered a key technology that allows agents to interact with other programs for MIT Technology Review in 2025. In science, AI agents can automate tasks, from writing code (now, even for quantum computing) to performing literature reviews to running entire research pipelines. Google, for instance, has an agentic system called Co-Scientist that helps researchers identify promising hypotheses and test them. These systems are often positioned as solutions for science, but they raise new problems too—including ignoring experimental evidence, as science and tech journalist Kathryn Hulick reported for Science News earlier this year.

The many ways scientists are using AI poses a challenge to journalists. “The thing that’s very hard to tell, even for a lot of AI reporters, is what is actually novel about the application of AI,” says Hall. “This is what you really want to get at.” If scientists are using AI for something specific to their field, like the researchers in Hall’s 2025 story on an AI tool for predicting solar storms, ask what AI brings to the table, he says. Hall has a computer science background, so he says finding experts in space weather to put the new AI result in context was critical for writing that piece. He suggests drilling into the details: Why now? Why use AI and why this model? Why hasn’t this been done before?

 

Critical Questions

Asking these questions, such as what type of model researchers used, helps journalists and readers alike understand what AI means for research, including potential environmental impacts. “I have to differentiate now between generative AI—that is, ChatGPT—and all the rest of the field, which is a big field that has a lot more in it,” says Hulick.

Training and operating LLM behemoths like Claude and ChatGPT requires specialized data centers with problematic appetites for energy and water. But many other AI models, like the brain-inspired “spiking neural networks” behind the low-power robot eye Hulick reported on in 2025, use far less energy. Without this clarification, a reader could easily attribute an AI-assisted discovery to an LLM running at a data center, when it was actually a lean model running on a cellphone hooked up to a solar panel.

Another important thing to clarify is the nature of a model’s training data and what it learns to spit out. “What is the input? What is the output? What is happening in between?” says Hutson. LLMs can regurgitate their training data, raising concerns about originality and plagiarism. And biases can sneak in via training data, too. For example, the astronomical amount of text LLMs train on is almost entirely in English—and mostly U.S. English. “There’s a very American-centric bias in everything machine learning,” says Wolfangel. “Surprisingly often we will find out that scientists don’t think a lot about these things—or not enough.”

That particular bias might not matter if you’re reporting on an LLM-assisted theoretical physics calculation, but it could for a story about a social science experiment that used LLM agents as a stand-in for humans. And bias isn’t just a problem with LLMs: All AI models learn patterns, and pick up potential biases, in their training data. Back in 2021, before the ChatGPT era, Wolfangel reported on racial bias in AI systems used to approve health insurance applications in the U.S. So ask your sources about any inherent biases in the training data that could be consequential for your story and whether they’ve been accounted for.

It’s also crucial to frame AI-assisted results appropriately. As LLMs have become more capable and agents more autonomous, their “outputs” have expanded to include research findings themselves, such as an entire mathematical proof produced by generative AI models. But when researchers or companies announce that an AI made a discovery on its own, it’s worth poking at that claim, says Ananyo Bhattacharya, the chief science writer at the London Institute for Mathematical Sciences. Experts sometimes don’t appreciate how much their knowledge contributes to an AI’s seemingly autonomous solution. “It’s a useful tool in the right hands, but in the wrong hands produces a lot of garbage,” says Bhattacharya. Simply knowing which problems are worth solving often requires a specialist’s taste, he adds. Clarifying the role of researchers in AI-enabled findings can help keep hype in check.

Reproducibility is another point to push sources on. When scientists use AI to generate or analyze data, there are several factors to consider. For example, if researchers are using proprietary models like GPT or Claude, you “don’t know what’s going on behind the scenes,” says Hutson. Companies might quietly change the models, which could shift findings mid-experiment. Similarly, LLMs will answer the same prompts differently, even when using the same model, because of the inherent randomness baked into their design.

“There is a stochastic nature of the models, but ideally a good experimental design would account for this,” says Ben Brubaker, a computer science reporter at Quanta. Journalists can lean on outside experts to vet a study’s design when it comes to incorporating AI.

 

An Elusive Frontier

Finding expert sources and asking sharp questions about technical topics is nothing new for science journalists. But the increasing use of AI in science presents special reporting challenges.

Private industry—mainly a handful of companies like OpenAI, Anthropic, Google, and Meta—dominates the cutting edge of AI research because of its ability to fund it. If your reporting leads to sources at an AI company, they will likely refuse to answer questions, and if they do talk, there’s a financial conflict of interest at play. “Try your damndest to talk to people outside of that company, and even better, outside of industry,” says Pavlus. “It’s not like I think any individual researcher at Anthropic or DeepMind or OpenAI is suspect, but you’d have to be a little bit of a fool to think … their [companies’] incentives are not suspect.”

This isn’t just true for AI researchers. As AI companies increasingly push into science, it’s quite possible that a physicist, mathematician, or biologist you want to speak with could be working at an AI lab, adds Will Douglas Heaven, the senior editor for AI at MIT Technology Review. “You are going to need access.”

Accessing academics can be as simple as emailing or picking up the phone. But in tech journalism, it’s often more difficult to reach sources, especially those who work in frontier AI labs, Heaven says. He was able to interview the head of OpenAI for Science in part able to in part because he’d built a good working relationship with the communications team over many years. Even in such cases, contact with the media is usually on the company’s terms.

If you are granted access, it’s worth thinking about what that means, says Hulick. “Unfortunately, the only way to get quotes … is if they think there’s some chance [of] a good marketing outcome,” she says.

Still, there are workarounds. Pavlus recommends looking for researchers at universities with partial affiliations at AI companies (sometimes their email addresses appear on publications and preprints)—people with “one foot in, one foot out” might be more willing to talk. Meeting researchers in person can also be a great help, says Hulick. She likes to line up interviews with AI researchers in advance of attending MIT’s EmTech conferences—or just introduces herself in the halls. Once you’ve built up relationships with researchers, it’s easier to start going to them directly.

Companies with a lower profile can be good sources, too. “There’s an entire ecosystem of smaller companies, startups who are also starting to dabble in AI for science,” says Heaven. They are typically more willing to talk—and challenge big claims from the tech giants.

As AI companies focus more of their energies on science and math, there won’t be a shortage of stories to cover. But journalists should keep in mind that the mere use of AI in a discovery isn’t necessarily newsworthy.

If you’re a beat reporter covering AI in your own domain, you can lean on your knowledge, says Heaven. Right now, the “AI” part of headlines like “AI solved a big problem in science” might be the most attention-grabbing. “As the technology quickly matures over the next few years … that will become less surprising,” he says. “Reporters with backgrounds in biology, chemistry, and physics are going to be the people really holding these claims to account more than the straight tech journalists.”

In other words: Tell the science story. And to do that, keeping up with AI is going to be vital. “Artificial intelligence isn’t going to stop,” says Paz Cardona. “We’re talking about a topic that’s evolving continuously … we’re definitely going to have to keep coming back to it.”