nature — AI can turbocharge scientists careers but limit their scope

Transcript کامل قسمت

If you felt stuck trying to lose weight, you're not alone.

Enter Weight Loss by HERS.

It's designed to support you in reaching your goals.

And HERS now offers access to an affordable range of FDA-approved GLP-1 medications, including the Wegovy pill and the Wegovy pen.

Even better, with a range of affordable GLP-1 options, HERS makes it simple to find an approach that fits your needs and your budget, if eligible.

You'll get a treatment plan personalized to you and unlimited dosage changes as needed.

It's weight loss designed to work with your life.

Ready to reach your goals?

Visit ForHers.com slash eligible to get personalized, affordable care that gets you.

That's F-O-R-H-E-R-S dot com slash eligible.

ForHERS.com slash eligible.

Weight loss by HERS is not available in all 50 states.

WeGovi is the registered trademark of Novo Nordisk AS.

To get started and learn more, including important safety information, WeGovi clinical study information, and restrictions, visit ForHERS.com.

The to-do list doesn't stop, and neither does the pressure to keep up with it.

If you've been running on fumes, Grow Therapy makes it easier to find care that's covered by insurance and actually built around you.

Whether it's your first time in therapy or your 50th, Grow makes it easier to find a therapist who fits you, not the other way around.

You can search by what matters, like insurance, specialty, identity, or availability, and get started in as little as two days.

And if something comes up,

You can cancel up to 24 hours in advance at no cost.

Grow helps you find therapy on your time.

Whatever challenges you're facing, Grow Therapy is here to help.

Grow accepts over 100 insurance plans.

Sessions average about $21 with insurance, and some pay as little as $0, depending on their plan.

Visit growtherapy.com slash ACAST today to get started.

That's growtherapy.com slash ACAST.

growtherapy.com slash ACAST.

Availability and coverage vary by state and insurance plan.

I find this not only refreshing, but at some level astounding.

nature.

Welcome back to The Nature Podcast.

This week, how AI might increase researchers' productivity, but narrow their focus.

And the mystery of the universe's little red dots.

I'm Benjamin Thompson.

And I'm Nick Petri-Chow.

Artificial intelligence tools are part and parcel of many areas of science.

And I don't just mean generative AI assistants like ChatGPT.

Deep learning has underpinned tools like AlphaFold, and machine learning has been used to find associations in data for decades.

And while a lot can and has been said about the effects of these tools on research, not much is actually known about their impact on science as a whole.

Well, we were aware of the hype and the associated kind of natural motivation for scientists to use these tools, but we just were wondering if there was a flip side to that.

What is that doing to science as a whole?

This is James Evans, a data scientist who's been looking at this question for a paper that was published in Nature this week.

And his results imply that these tools may well make scientists more productive, but at a cost.

James and a team of researchers looked at more than 41 million papers published between 1980 to 2025 that were, in some way, assisted by AI tools, which they refer to as AI augmented research.

They also chose to look at the natural sciences, ignoring computer science and mathematics.

We were interested in how AI is being applied to the sciences rather than just the core development of AI itself.

and the natural science research they were looking at wasn't the only thing that was AI augmented.

To read through these 41 million papers, the team used an AI language model called Bidirectional Encoder Representations from Transformers.

or BERT to its friends.

This allowed them to scan through the titles and abstracts to identify possible AI use.

This could be things like machine learning or deep learning to scour through data or even generative AI use.

The accuracy of this was then scrutinised by human experts, and Bert agreed with the experts 87.5% of the time.

Which is higher than experts agreed with each other.

From the initial millions of papers, the team ended up with just over 310,000, but their analysis suggested presented research that was AI augmented in some way.

They could then look at how many citations these papers got versus the ones that hadn't used any AI tools, and how much the scientists behind the AI augmented research went on to publish, and how it impacted their careers.

I mean, it's a great career move.

It was a great move at any point when they kind of split from their peers in using AI tools versus comparable scientists that chose not to use those tools at any point of their scientific development.

They got more citation attention, they were less likely to drop out of science, and they were more rapidly likely to establish labs and become senior scientists that direct research on paper."

Scientists who did AI-augmented research published about three times as many papers, got nearly five times as many citations, and became research project leaders more than a year earlier than their non-AI-augmented peers.

But James and the team were also interested in the effects of AI on science writ large.

So they looked at how these papers related to others.

Did they cite one another, for example?

And while they found a benefit to individuals, this AI use seemed to be narrowing the focus of science itself.

So it seemed clear that it was really compressing or kind of automating existing scientific fields rather than generating new

questions that you know lead to fermentive discussion the analysis suggests that ai augmented researchers were doing more research but in fewer areas instead focusing on fields that had lots of data and that they could apply their ai tools to a trend which concerned james if everyone's kind of climbing up the same tree

The low-hanging fruits are gone, and we're kind of battling over a couple of percentage points as we solve important problems.

So AI is being used to solve important, established, agreed-upon problems, and that's important.

The problem is if you solve all those problems or if you solve them much faster than you

generate new problems, then it slows down the rate of novel questioning and discourse and the kinds of things that unleashes, I would say, human creativity and vision.

Vader Story, a researcher of computer information systems who's been writing a News and Views article about the new research, was less concerned about these results.

I thought they were too pessimistic.

I would be much more optimistic.

In some ways, Veda saw these results as a natural conclusion of using AI tools.

The tools we have are largely good at working with and sorting through lots of data.

So it's unsurprising that people who use them may focus on areas that are data rich.

And that's something that she thinks could allow science to advance.

If you think about it, they have identified the fact that many scientists are using AI tools.

And

if they allow them to go deeper into certain areas, we can benefit from that.

We can benefit from an in-depth understanding of it.

So I just see so many opportunities for us as scientists to continue on this route of using the AI tools.

And these AI tools are only going to get better.

Vader also pointed out that it may be hard to generalise about all of science from the limited study here, especially as it doesn't include some areas like mathematics and computer science.

And while the paper did look at generative AI tools, she cautions that it's a bit soon to tell how they might impact science.

On these tools as well, she was a bit more cautious.

She didn't think they were quite ready for use in scientific writing and she had concerns that they may impact replicability.

In the past, AI tools have been deterministic.

You put in an input and you get the same output, time after time.

Generative AI is probabilistic, meaning that you can get different outputs from the same input.

Now in science, we always need to replicate a result.

That is a very important thing in science.

Can we replicate these studies?

Can we ensure that the results that we are reporting are good and hold the test of time for scientific standards?

If you're using a tool that will give you different outputs, given the same input, what happens when you go to try and replicate all these studies?

In her opinion, James's study has laid the foundation for further work on this topic, and she would like to see how it would play out in a few years' time.

James, despite his concerns about the narrowing of scientific focus, believes that AI tools have a lot of possibilities.

For him, we need to instead look at the incentives in science.

I think what we see here is kind of a conflicting incentive for individuals and for science as a whole, right?

Individuals are trying to survive in the scientific universe.

They want promotion, they want resources to do more science.

And what's the fastest, most efficient way to do that, especially with science?

you know, AI tools that effectively compress, you know, data that you have and produce answers and predictions.

But I think science as a whole has a different incentive, which is to kind of know everything.

James would like to see more incentives to forge new fields and ask new questions.

He even thinks AI could assist with that by collecting data from more data sparse fields.

That too would need incentives.

Without change, James predicts that people will just carry on using AI how they have, potentially shrinking the focus of scientific inquiry.

So to allow science to grow, James thinks we need to change how we use AI.

AI use is vast and it's growing and it's accelerating in science and there are some massive missed opportunities.

in how we use it to expand the space of collective inquiry to new fields and the generation of new data.

And that if we don't do that, then we risk the kind of model collapse that we see when AI models basically consume the results of their own data.

You know, they pull on the tails of the distribution and they just stop working.

Like we need science to grow.

That's its character.

And so that means we need to redirect the use of AI.

That was James Evans from the University of Chicago in the US.

You also heard from Vader Story from Georgia State University, also in the US.

For more on that, check out the show notes for some links.

Coming up, the research that may have solved the riddle of the universe's little red dots.

Right now, though, it's time for the research highlights, read by Katrina Clark.

Snowball Earth's oceans were not only very cold, but also extremely salty, according to analysis of rocks from the time.

Around 700 million years ago, it's been proposed that the Earth resembled a snowball, with glaciers reaching all the way to the equator and average temperatures potentially 12 degrees Celsius below freezing.

To put it mildly, it would have been pretty chilly.

And this covering of the oceans with ice would cut off the usual exchanges between land and sea.

For example, iron that would usually have been oxidised by photosynthetic organisms would have ended up deposited on the ocean floor.

Researchers analysed such deposits in rocks dating from the snowball Earth period.

By modelling how different isotopes of iron would have been separated in water, they concluded that the iron was deposited in temperatures between minus 22 and minus 8 degrees Celsius.

At these temperatures, to stay liquid, the water would have had to have been extremely salty.

These frigid times may actually have been Earth's coldest ocean temperatures.

Chill out and give that study a read over nature communications.

Putting immune cells into night mode may reduce the damage from heart attacks.

Neutrophils are a type of immune cell that protect against microorganisms, but they can also kill surrounding tissue, which can increase the damage from heart attacks.

It's been known that neutrophils are more active in the early morning than at night, so researchers tried to see what would happen if they put them into the less active, night mode, in mice.

The mice were given a drug that targeted one of the neutrophils' receptors that controlled daily fluctuations in activity.

The team found that the treated neutrophils displayed nighttime-like behaviour, and the mice given the drug had less dead or damaged heart tissue.

This treatment didn't seem to affect immunity, as the mice responded normally when treated with Staphylococcus aureus bacteria or Candida albicans fungi, suggesting that this night mode for neutrophils could be a promising approach to treat heart attacks in people.

Don't switch off from that research.

It's over in Journal of Experimental Medicine.

An astronomical puzzle may be a step closer to being solved, thanks to research published in Nature this week.

Now, if you look at an image taken by the James Webb Space Telescope, the JWST, it's hard not to be wowed by spectacular images of distant stars and galaxies.

But if you look closer, you might notice something that has left researchers scratching their heads.

Often, these images are peppered with tiny points of light, known as little red dots.

The light from these dots has travelled a really long way to get to us, meaning these objects are likely really old.

Some estimates have them down as being present in the early universe, around 600 million years after the Big Bang.

And these little red dots have been dubbed universe breakers because they don't fit in with standard thinking about the features of the early universe.

There's been a lot of debate about whether these dots are young, star-filled galaxies or outsized black holes.

Well, according to a team of researchers, the answer could be neither.

To find out what they think it is, I spoke to one of the team, Vadim Rusikov, who is affiliated with the University of Manchester here in the UK.

Vadim explained to me why neither of the initial theories about the little red dots quite worked.

So when people counted the number of stars in those red dots, they found that there's just too many stars to be produced at that point in the universe.

They were universe breakers just because it's hard to explain how you can form so many stars in such a small volume so early on.

So there had to be something else that was contributing to that light coming from those dots.

When we looked in more detail...

So when we gathered more information and more data about these little red dots, we found that they exhibit features of supermassive black holes.

And those particular features are of a gas rotating quickly around a supermassive black hole.

And because supermassive black holes produce a lot of light, so the gas gets secreted, it gets really heated and produces a lot of light, sometimes they shine just as bright as a whole galaxy.

And so because you have these dots made up probably partly from stars and partly from the supermassive black holes, now that universe breaker problem in the stellar context kind of goes away.

But there were a couple of issues that we still had to resolve in order to tell that they are galaxies containing supermassive black holes.

So there was this discrepancy then as to what they might be.

But there was one theory that was perhaps a front runner.

And this was the unusual situation of there being supermassive black holes enveloped in a layer of gas.

And there was some evidence to back that up.

And your work comes to a similar conclusion and perhaps helps to solve one of the puzzles surrounding the idea that at the heart of these little red dots was a regular supermassive black hole.

And to get to this conclusion, you looked at a dozen of these little red dots for which there's been a lot of data collected.

And particularly, you were looking at the light emitted by these objects.

What can that tell you?

So, yes, we can tell how massive the black holes are just from the rotation of the gas around the supermassive black holes and the light that it kicks out.

There's a really broad feature in the spectrum that tells us that there has to be gas moving at thousands of kilometers per second.

So people found this feature, but the problem with that was that you can use it to measure the mass of the black hole.

But if you use it to measure the massive black hole, it turns out to be quite massive.

Again, we're coming back to the same problem of universe breakers, not with the stars, but with the black holes.

Because they are found so early in the universe, it was a bit problematic that they were so massive at that point in the universe.

So the black holes had to have formed really quickly and have created a lot of material in order to have grown so much.

So what you're saying is then, yeah, that something this old couldn't get that big that quickly then, but the data suggested that that's what was going on?

Right.

So what we found in our work is that it was not the problem with the black holes themselves, but it was the problem with how we interpreted the data.

So the same observation that tells us the gas has to be rotated really quickly could actually be produced by a very different system where the gas around the black hole has to be just very dense.

It doesn't have to be rotating as quickly.

And what is special about this system compared to a normal supermassive black hole that we can find in the local universe or elsewhere is that it doesn't have to be just a neutral gas.

It has to be an ionized gas.

There have to be essentially a sea of electrons, a cocoon of electrons surrounding these black holes.

So in our model, the light gets scattered off the free electrons in that cocoon.

That scattering actually produces the features that people thought before were telling us about the gas as quickly rotating around the black holes.

Because the gas doesn't rotate as quickly, the black holes that people thought were massive are actually not as massive and sometimes 10 to 100 times less massive, which helps us ease a lot of these universe breaker type of problems.

And this thick cloud of gas helps to explain some of the other features that weren't seen, because usually black holes kick out all sorts of different wavelengths of electromagnetic radiation, which aren't seen in the little red dots.

And maybe this cocoon is helping that.

Yes, exactly.

There are different pieces to this puzzle.

So normally if you see a black hole, you'd expect there to be all kinds of emission produced from x-rays to radio waves.

Then you also see all kinds of light variation can happen.

So light sort of twinkles a little bit when you look at these systems over time.

But we don't see any of those.

And so the puzzle actually can be solved with this cocoon idea.

Those x-rays, those radio waves that we do not detect can be partly killed off by this shell of electrons, essentially.

If it's also a thick cocoon of gas, that could help to explain the fact that these black holes do not twinkle.

It's still not established, but it can happen that because there's a thick cloud of gas that the light has to travel through,

It takes a somewhat longer time, which means that these black holes can vary in light.

They can twinkle, but on a longer timescale.

And we just haven't observed them on a long enough timescale.

And so your results then suggest that what you've got here is a relatively small supermassive black hole, if I can put it like that, surrounded by this thick cocoon then of gas and electrons and what have you.

Does this suggest that this is how enormous black holes might develop over time as these distant little red dots are so ancient?

So we expect some of them to grow to the supermassive black holes that we find in the local universe.

So some of them will probably get to very massive black holes.

Some of them will probably stop growing just because they ran out of fuel.

And we think that we find them in this early stage where it's a stage of very rapid growth.

that we haven't seen before, actually.

So we haven't seen such objects in the local universe.

We're only now starting to find, actually, things that are analogous to these little red dots, but much closer, very low numbers.

Obviously, your paper is out now, and you put forward your evidence that these little red dots are relatively small, supermassive black holes surrounded by gas.

Of course, there are competing theories about what these things are.

Do you think your work puts this to bed?

What do you think other people will make of these results?

So there are different groups working independently from each other, coming to very similar conclusions with similar models as we do.

So there's at least two or three probably including ourselves.

So I think it helps to have some sort of consensus in the community that these black holes have to be cocooned in a thick shell of gas, which is what makes these black holes unique, unlike the black holes we've seen previously.

But there might be some things that we're still trying to figure out in terms of the physical details as to how much there is ionized gas and so on.

So we're trying to figure out the details.

And everyone loves a space mystery.

I mean, it seems like we're discovering new stuff all the time.

I mean, are you happy that you think you've solved the mystery of the little red dot?

Or are you almost sad that now we know what they are and the mystery has gone somewhat?

I'm very happy, of course.

It's been a whole journey.

So I've learned a lot on the way and I've met a lot of people.

So it's been very delightful to be part of this.

I think we've sort of pushed the boundary of what is known and what is unknown further.

So this opens up all kinds of questions now that we can attack.

Like, how does a massive black horse form?

We're finding them in their youth, but can we understand how they actually formed?

Vadim Rusikov there.

To read his paper, head over to the show notes for a link.

That's it for this week's show.

But before we go, a little announcement.

Yeah, absolutely right, Nick.

It's a new year, and we're going to be changing things up a little bit here at The Nature Podcast.

Now, at this point in the show, we'd usually be doing the briefing chat.

But starting this week, we're going to be spinning that segment out into its own podcast, which will be in your feeds on Fridays.

So you'll still be getting the latest science news on a Wednesday.

And a quick research roundup on a Friday.

We'll see you then.

I'm Benjamin Thompson.

And I'm Nick Pertridge Howe.

Thanks for listening.

ActiveCampaign is the marketing automation platform built for big swings and big dreams with intelligent suggestions powered by AI and your data.

Generate ideas in seconds, import your brand, and create full campaigns with simple prompts.

Send personalized messages backed by real-time feedback, smart segmentation, and effortless reporting that tracks every win.

Let's redefine what's possible together.

Get started for free at ActiveCampaign.com.

Hi, it's Paige from Giggly Squad, and this episode is sponsored by Experian Boost.

Summer glow up, check.

Credit glow up, even better.

Boost your credit scores instantly by getting credit for bills you're already paying.

Your phone, utilities, rent, and insurance.

I wish dating kind of worked like that.

Connect your bank account, add those on-time payments to your Experian credit file, and your FICO score updates right away.

You could instantly raise your FICO score by an average of 14 points with Experian Boost.

Download the Experian app for free today.

Results will vary.

Users who received a boost improved their FICO score 8 from Experian by an average of 14 points.

See App Store or Experian.com for details.

Results will vary.

Not all payments are Boost eligible.

Users who received a Boost improved their FICO score rate from Experian by an average of 14 points.

Some may not see improved scores or approval odds.

Not all lenders use Experian credit files, and not all lenders use scores impacted by Experian Boost.

See Experian.com for details.