0:01

BBC Sounds, music radio

0:03

podcasts. Hello lovely curious-minded

0:05

people, welcome to Inside Science, a

0:08

program that was first broadcast on

0:10

the 27th of February 2025. I'm

0:13

Victoria Gill, and today we are

0:15

going to delve straight into a

0:17

scientific revolution, artificial intelligence, and its

0:20

transformational role in science. Over the

0:22

next half hour, we're going to

0:24

unravel the power that's promised and

0:27

the threats that are posed by

0:29

this transformational technology. We'll investigate whether

0:32

AI really solved a major microbiological

0:34

mystery that had taken human scientists

0:36

years to crack in just a

0:38

few days, and we're examining the

0:41

bizarre world of AI fakery and

0:43

how it's making its way into

0:45

scientific papers. To navigate all of

0:47

this we have technology expert and

0:49

science communication lecturer Gareth Mitchell in

0:51

the studio. Hello, Gareth. Hello, nice to

0:53

be here. It's very nice to have you.

0:55

And this is right up your street, isn't

0:58

it? Oh, very much. Oh, I'm having such

1:00

a good program. Yes. You know, because I'm

1:02

so fascinated by machine learning. I've been reporting

1:05

on it for years. I know you have

1:07

as well, Vic. And I'm fascinated by the

1:09

way science happens. And so the story we're

1:11

talking about today brings both of those bumping

1:14

into each other. amazing consequences. You know for

1:16

good we hope but you know I've spoken

1:18

to some scientists who are already sounding a

1:21

bit skeptical about it as well so this

1:23

has put like a cat among the pigeons

1:25

a little bit so yeah it's a story

1:27

that matters for the world of science but

1:29

by default it means it matters for the

1:32

rest of us. Yeah absolutely let's get right

1:34

into it then Gareth you have been speaking

1:36

to a scientist who had a bit of

1:38

a bewildering experience when he tested a new...

1:41

AI tool is that right? Yeah absolutely yes

1:43

so this was at Imperial College where

1:45

I went to speak to Dr. Tiago

1:47

Costa and he and his team have

1:50

been testing a tool that's being developed

1:52

by Google and this is called the

1:54

co-scientists so I suppose the clues in

1:56

the name a little bit about what

1:59

this is about. And so far

2:01

this tool is being aimed specifically

2:03

at biomedical applications. And Tiago and

2:05

his team, they study bacteria, but

2:07

also these little things called phages.

2:10

And I hope you don't mind

2:12

a bit, but I might put

2:14

you on the spot here. Do

2:17

you know? as a top correspondent

2:19

and presenter. What are phages? Top

2:21

correspondent, I'm going to take that.

2:23

Okay, I mean, without referring at

2:26

all to my script and any

2:28

preparatory conversations I've had to do.

2:30

So phages are like little tiny

2:33

viruses that can infect bacteria, right?

2:35

Indeed, yes. And they do so

2:37

through these little tails. And it's

2:39

through the tail that genetic material,

2:42

DNA, in the phage, goes into

2:44

the bacteria. And there's some excitement

2:46

around them. I mean, scientific, they're

2:49

very interesting, but potentially, therapeutically as

2:51

well. So these phages matter. And

2:53

we're talking about phages quite a

2:55

lot in this interview, or about

2:58

two here. So, yeah, you have

3:00

the phage, has some DNA in

3:02

it. And there's a mechanism for

3:05

that DNA going into the bacteria.

3:07

So you spoke to Tiago about

3:09

his experience with co-scientists. So let's

3:11

hear more from that conversation. Phage

3:14

is a entity made by proteins.

3:16

This very big large ball with

3:18

the DNA inside and a tube

3:21

through which the DNA gets injected

3:23

into the bacteria. This non-infectious phages

3:25

are lacking a tail. Okay, and

3:27

the tail is what is important

3:30

for a phage to bind the

3:32

surface of a bacteria and inject

3:34

the DNA. What we found was

3:37

that the DNA of this non-infectious

3:39

phages were embedded in different bacterial

3:41

species and we didn't know why

3:43

and how it was reaching there.

3:46

So you were trying to find

3:48

out how this process works. get

3:50

down to. So you need to

3:53

sit down at some point and

3:55

think we think it might be

3:57

this thing as I understand and

3:59

then in science you have to

4:02

then go and test your hypothesis.

4:04

But Google came to you didn't

4:06

they at one point and said

4:09

we've been working on something we

4:11

want you to help us try

4:13

it out. Is that how it

4:15

happens? asked this to basically to

4:18

understand whether the scientific hypothesis that

4:20

the AI system was generating were

4:22

valid. And the only way to

4:25

validate those hypotheses is actually go

4:27

to the lab and test whether

4:29

those hypotheses are correct or not.

4:31

Simple. system looked like? I'm just

4:34

trying to think is it a

4:36

bit like ChatGPT? Well actually we

4:38

didn't have direct access to the

4:41

system because this was still being

4:43

underdeveloped by Google so what we

4:45

did was to send them what

4:47

our scientific question was and a

4:50

few lines of introduction and a

4:52

few references. So what were you

4:54

typing into the Google co-scientists to

4:57

kind of get the hypothesis that

4:59

you wanted? So it was a

5:01

very simple question. was how these

5:03

phages, which are non-infectious, are able

5:06

to infect bacteria from different species

5:08

and integrate themselves into their DNA.

5:10

What was the mechanism? So after

5:13

a few years of research, we

5:15

understood that what those non-infectious entities

5:17

were doing, they were ejecting the

5:19

tales from different phages. in order

5:22

to inject their DNA, which was

5:24

encapsulated in this ball-shaped structure on

5:26

top of the tail. So we

5:29

knew the answer to the question

5:31

before we challenged the AI system.

5:33

So the question was, like, how

5:35

are these pages getting their DNA

5:38

into the bacteria, even though they

5:40

don't have the equipment, the mechanism,

5:42

the means, really, to do anything?

5:45

Yes, that was exactly the question.

5:47

And then, two days later, the

5:49

AI system came up with five

5:51

different hypotheses. So the algorithm spits

5:54

this out for you after 48

5:56

hours. What was your reaction when

5:58

you saw this? Well, it was,

6:01

we were very surprised because basically

6:03

the main hypothesis was a mirror.

6:05

of the experimental results that we

6:07

have obtained and took us several

6:10

years to reach. So it was

6:12

a shocking, I would say. Is

6:14

it possible though, just from the

6:17

literature you'd already put out, because

6:19

you'd be putting out a preprints

6:21

and stuff as you go along,

6:23

presenting it at conferences, that this

6:26

had... kind of crept in as

6:28

it would do into Google's data

6:30

set and so really it was

6:33

just reciting your own research back

6:35

to you or not. No, so

6:37

this this preprint was was kept

6:39

secretive for a while. Why? Because

6:42

we thought our opportunity to patent

6:44

this technology which we did so

6:46

that so that's the reason why

6:49

this did not become public until

6:51

the patent was filed. And during

6:53

this process, we will start to

6:55

talk with Google. There's no way

6:58

that Google will have access to

7:00

our preliminary data. while they were

7:02

generating the scientific aquatics. Some of

7:05

this seems a little bit opaque

7:07

to me and I don't know

7:09

if Google have paid you to

7:11

do this with search either. No,

7:14

there was zero funding, so this

7:16

was a benefit for both, but

7:18

zero pounds involved in this. As

7:21

we all know, anybody who's used

7:23

any of these LLLMs and chat

7:25

bots, well no, they hallucinate. They

7:27

can come up with stuff that

7:30

seems plausible, but is... wrong. So

7:32

one of the of the hypotheses

7:34

that has generated, we have never

7:37

thought about that. And from the

7:39

preliminary data that we already have,

7:41

it looks very, very likely that

7:43

that's a novel mechanism of DNA

7:46

transfer. Okay, so it not only

7:48

is validated, our experimental work, but

7:50

it has generated a novel hypothesis

7:53

that we have never thought about.

7:55

And the preliminary data we have,

7:57

it's looking, it looks promising. Were

7:59

any of these hypotheses that are

8:02

generated complete nonsense? I would say

8:04

nonsense, but I would say that

8:06

are less relevant to the question.

8:09

Do not answer directly the question.

8:11

And this is why it's so

8:13

important. the human brain to critically

8:15

evaluate those hypotheses and not take

8:18

it as any of the hypotheses

8:20

as the final answer or the

8:22

finitive answer to the scientific question.

8:25

I think you say that with

8:27

genuine belief as somebody who is

8:29

paid to be a scientist and

8:31

doesn't want to be replaced any

8:34

time soon. Yeah, well, definitely this

8:36

system won't replace humans. That's for

8:38

sure. So it won't replace humans,

8:41

that's somewhat comforting. This is such

8:43

an intriguing story though, and we're

8:45

going to try and unpack exactly

8:47

what happened here. To help with

8:50

that, we have Maria Leah Carter

8:52

from Queen Mary University of London,

8:54

who is also a Turing AI

8:57

fellow. She is professor of natural

8:59

language processing and some of her

9:01

research focuses on the benefits and

9:03

limitations of AI in science. Hi,

9:06

Maria, welcome to Inside Science. Hi,

9:08

I'm really glad to be here.

9:10

Well, it's a pleasure to have

9:13

you and to hold our hands

9:15

through how exactly this works. So

9:17

many of us have at least

9:19

played with things like ChatGPT or

9:22

Google's Gemini. Is Google co-scientist something

9:24

very different from those chat bots

9:26

that we might be more familiar

9:29

with? How does it work? So

9:31

co-scientist is based on... technology, so

9:33

large language models itself is a

9:35

much more complex system than a

9:38

single large language model. But basically

9:40

how large language models work. They

9:42

are state-of-the-art probabilistic algorithms that have

9:45

been trained on large amounts of

9:47

data and they are trained to

9:49

generate outputs given a particular inputs.

9:51

Right. And inputs are prompts. So

9:54

they usually are texts, like a

9:56

phrase, a sentence, a question, or

9:58

more complex instructions. So when we

10:01

are asking a question to an

10:03

LLLM, we are providing it essentially

10:05

with a prompt to generate an

10:07

output. Right. I sort of see

10:10

it as the hovering up all

10:12

of this information and by listening

10:14

to it consuming it, when you

10:17

give them a prompt, they're working

10:19

out what the most probable... Next

10:21

word, answer to a question is

10:23

from all of that text and

10:26

info that they've consumed. Is that

10:28

fair if massively oversimplified? It's simplified

10:30

because they are making some really

10:33

complex associations. So they're not just

10:35

predicting the next word, but essentially

10:37

they're predicting what would be an

10:39

appropriate output given the prompt that

10:42

you've given. Right. And Google described

10:44

co-scientist as a multi-agent. systems, multiple

10:46

LLLM's, large language models. What does

10:49

that mean in practice? So that's

10:51

correct. So co-scientist itself is not

10:53

an LLEM, it's rather a coalition

10:55

of LLLM's, and they each specialize

10:58

in a different task. And so

11:00

what this multi-agent architecture means is

11:02

that you have... multipleLLMs that are

11:05

interacting with each other to generate

11:07

hypothesis that are evaluated and further

11:09

refined. And this kind of high-level

11:11

supervisor agent which coordinates all the

11:14

others, I would say, is the

11:16

main technical novelty in the system.

11:18

And I find something quite reassuring.

11:21

about that in a way Maria

11:23

because we have a lot of

11:25

discussions about artificial general intelligence and

11:27

you know this idea that we

11:30

before long we're going to end

11:32

up with an amazing brain and

11:34

you just say hey brain how

11:37

do you cure a range of

11:39

diseases and it'll just do it

11:41

clearly that's probably never going to

11:43

happen certainly not in the way

11:46

that I've described this Maria, from

11:48

what you're saying, when you have

11:50

multi-agents, it's just lots of LLLMs

11:53

and other machine learning systems that

11:55

specialize in certain things. One of

11:57

them might have generates some of

11:59

the initial hypotheses. One might run

12:02

the tournament that matches one off

12:04

against the other and works, which

12:06

is the best, like a ranking

12:09

kind of thing. In other words,

12:11

this is artificial, very specific intelligence.

12:13

It's going the opposite way from

12:15

this idea, artificial general intelligence, isn't

12:18

it? Yes, I think the trend

12:20

is to go for more sort

12:22

of smaller... more specialized systems than

12:25

having just one big system that

12:27

does everything. And sort of the

12:29

the idea here is that not

12:31

only is it LLLMs but they

12:34

also use external tools like web

12:36

searches and access to online databases.

12:38

So it's kind of a quite

12:41

a complex pipeline that consists of

12:43

a lot of different components. One

12:45

thing that came out from the

12:47

interview with Tiago is that this

12:50

co-scientist came out with one or

12:52

two hypothesesacies. that we're not anything

12:54

to do with existing literature or

12:57

any training data that was already

12:59

there. These seem to be absolutely

13:01

novel and I was rather stunned

13:03

by that, Maria. Is that something

13:06

that you might have expected from

13:08

such a co-scientist as well? So

13:10

I think that this is possible.

13:13

So, I mean, a novel hypothesis

13:15

can be created by essentially generating

13:17

outputs that synthesizes and stems from

13:19

indirect associations, right? So that you

13:22

have these kind of... complex context

13:24

that you wouldn't necessarily put together,

13:26

but by having this kind of

13:29

iterative hypothesis generation and refinement, it

13:31

can happen. Right. So I can

13:33

see there how those tools by

13:35

coming up with things that scientists

13:38

hadn't spotted could help scientists make

13:40

progress quickly, but what are the

13:42

pitfalls of this approach? So what

13:45

is presented here is a very

13:47

very complex pipeline. It's very resource

13:49

intensive. This is not really... discussed

13:51

very much in the paper, which

13:54

by the way is more of

13:56

a showcase paper than actually revealing

13:58

a lot of the technical details.

14:01

But you know, the fact that

14:03

it's so resource intensive means that

14:05

it's not something that would be

14:07

very readily replicable by scientists, so

14:10

they would have to sort of,

14:12

you know, be working with someone

14:14

like Google to do this. So

14:17

what do you think of the

14:19

potential dangers of scientists overestimating and

14:21

over relying on... on technology like

14:23

this? My main concern is becoming

14:26

complacent and not really understanding how

14:28

hypothesis are generated, understanding the reasoning

14:30

process itself, you know, having the

14:33

background knowledge. I mean, at the

14:35

moment, there are quite a few

14:37

limitations to this system in particular,

14:39

but assuming that these can be

14:42

very much improved in the future,

14:44

I think this is a concern

14:46

I have. Yeah. Well, Maria Liakata

14:49

from Queen Mary University of London,

14:51

thank you very much indeed for

14:53

taking us through that. So it

14:55

seems kind of to come back.

14:58

Gareth to this idea that co-scientist

15:00

isn't a scientist, it's not a

15:02

human, it's a tool, and understanding

15:05

the limitations and how that tool

15:07

works is really important for science.

15:09

Yeah, hugely, and by sheer good

15:11

fortune I happened to spend a

15:14

day with the whole room full

15:16

of early career scientists yesterday, so

15:18

I asked them what they thought,

15:21

and I was reassured that there

15:23

didn't seem to be too much

15:25

potential complacency setting in, if anything,

15:27

they were really quite cautious about

15:30

systems like this. they were worried

15:32

also about their data, you know,

15:34

because they'd have to give a

15:36

lot of their data to feed

15:39

into the AI and prompt in

15:41

order to get whatever results out.

15:43

and they had worries about their

15:46

ownership of that data and then

15:48

what might come out of the

15:50

machine the other side. And also

15:52

that for them, hypothesis generation is

15:55

a key really important professional skill

15:57

for any scientist, especially in early

15:59

career scientists. It is lovely that

16:02

you have a machine that will

16:04

help you with your hypothesis, but

16:06

unlike your labmate who might rip

16:08

with you about your hypothesis and

16:11

give you a few ideas, you

16:13

can't say, well, that's a genius

16:15

idea. Where did that idea come

16:18

from? You can't ask a black

16:20

box how it did that. So

16:22

I think they were healthily skeptical

16:24

about it. On the other hand,

16:27

there's huge pressure in science to

16:29

get out there and publish. So

16:31

if you can use tools that

16:34

are going to speed up your

16:36

pipeline that your rival lab isn't

16:38

using, you know, you can see

16:40

some temptation there as well. Yeah,

16:43

which is, you know, where these

16:45

kind of pitfalls interact with all

16:47

of that promise, right? So I

16:50

should say that Google told Inside

16:52

Science that co-scientists is still in

16:54

development, is currently available via their

16:56

trusted tester program, and that's a

16:59

program that research institutions can apply

17:01

to join. Now though,

17:03

AI has also started making its

17:05

presence felt in scientific papers, specifically

17:07

generative AI, which can create new

17:09

content based on all the information

17:11

it's been trained on, including text

17:13

or images. Garif, you have some

17:15

striking examples for me. Yes, I

17:18

do. This one, I'm blushing just

17:20

thinking about this. It was a

17:22

paper, there are the dry title

17:24

really, of cellular functions of spermatogonial

17:26

stem cells. It appeared in the

17:28

journal Frontiers in cell development and

17:30

biology. And there was an image

17:32

that went viral on social media.

17:34

I remember this. Machine learning generated

17:36

images. And it was of a

17:38

rat, okay. And we can all

17:40

remember, you know, in our school

17:42

textbooks, maybe a biological picture of

17:44

a rat where you can see

17:46

it with its, you know, an

17:48

artist's impression of, you know, the

17:50

fur, the whiskers, and it looks

17:52

quite cute, but then maybe a

17:54

bit of it's being cut away

17:56

as a kind of a kind

17:59

of dissection. Yeah. And that was

18:01

the case with this beautiful, penis.

18:03

But the only problem was that

18:05

the penis that was also shown

18:07

in that dissection cutaway style where

18:09

you could see all the vessels

18:11

and the tissue inside. It was

18:13

nearly as thick as the rat

18:15

was wide and was so large

18:17

it extended outside the frame. It's

18:19

about twice the size of the

18:21

rat's body. About twice the size

18:23

of the actual rat and some

18:25

very purerile people on the internet

18:27

thought it was incredibly shareable and

18:29

red places all around and the

18:31

journal did retract the article and

18:33

apologize. I mean it was a

18:35

fairly well-known journal as well wasn't

18:38

it? And it's some of the

18:40

labelling on that diagram made no

18:42

sense whatsoever. Like Sarah Goma cell

18:44

I think it was and synctolic

18:46

stem cells. It also did point

18:48

to rats and got that... correct

18:50

and said this was indeed a

18:52

route. But this is just the

18:54

tip of the iceberg. AI generated

18:56

images are making an increasingly frequent

18:58

appearance in published scientific papers. So

19:00

is that a problem? Joining us

19:02

now is image integrity analyst Jana

19:04

Christopher. Hi Jana, welcome to the

19:06

program. Hello and thanks for having

19:08

me. It's absolute pleasure. Now it's

19:10

your job to check images in

19:12

research papers before they're accepted for

19:14

journal publication. Is that right? AI

19:17

must have... must have really changed

19:19

things for you then? Yeah, with

19:21

AI, obviously things have changed. We

19:23

were talking there, you know, somewhat

19:25

tongue in cheek about an image

19:27

that was very obviously and ridiculously

19:29

AI produced, but how difficult generally

19:31

in your job is it to

19:33

distinguish an AI produced image from

19:35

a genuine image? I mean in

19:37

a nutshell is very difficult and

19:39

it can almost be impossible. So

19:41

at this point, AI image generator

19:43

is still... occasionally make mistakes. If

19:45

you have an image of a

19:47

human or an everyday object that's

19:49

AI generated, then you might find

19:51

flaws like extra digits or body

19:53

parts that don't align or things

19:55

like that. But when you're talking

19:58

about histology images, for example, it

20:00

becomes much more difficult. So, Jana,

20:02

can you just explain what you

20:04

mean by a histological image? But

20:06

by that I mean a microscopy

20:08

image showing a piece of tissue

20:10

on a cellular level so that

20:12

you can see all the details.

20:14

Right. So some of these more

20:16

complex scientific images, you're delving into

20:18

kind of more detail and more

20:20

complexity and so it's difficult to

20:22

start. Yes, and it's hard to

20:24

know, you know, when there's like

20:26

a little glitch, it's hard to

20:28

know whether that's actually in the

20:30

tissue or whether that's something to

20:32

do with, you know, with the

20:34

AI generation. So there was an

20:37

interesting study actually with over 800

20:39

participants published in Nature in November

20:41

last year. that they studied the

20:43

ability of human subjects to discriminate

20:45

between artificial and genuine histological images.

20:47

And they found a clear difference

20:49

between naive and expert test subjects.

20:51

So an example for an expert

20:53

would be an oncologist studying liver

20:55

cancer looking at images of liver

20:57

cells, right? So whereas naive participants

20:59

only classified about half of the

21:01

images correctly. the experts performed significantly

21:03

better with a hit rate of

21:05

about 70%. So I guess this

21:07

might give us hope that that

21:09

some of the incidences might be

21:11

picked out doing peer review. But

21:13

experience shows that peer reviewers rarely

21:16

pay that much attention to manuscript

21:18

figures, and they might actually need

21:20

to be prompted to do so.

21:22

Right. They're more focused on the

21:24

text. And you can see how,

21:26

if you're presenting data looking, say,

21:28

at the impact of a treatment

21:30

on cancer cells, for example, that

21:32

that's a real problem if that's

21:34

being faked by AI, how widespread

21:36

is the problem, do we know?

21:38

We don't really know how many

21:40

papers have AI-generated images in them.

21:42

simply because we lack reliable tools

21:44

to detect them. We're seeing an

21:46

exponential growth of academic articles published

21:48

per year over the last decade,

21:50

whilst the time spent on obtaining

21:52

the results and validating them and

21:54

peer-reviewing them has decreased significantly. We've

21:57

seen some journals trying to tackle

21:59

this directly. has actually banned publication,

22:01

the use of AI-generated images in

22:03

scientific papers, but how else do

22:05

we tackle this issue? What can

22:07

be done? In terms of guidelines,

22:09

most journals still permit the use

22:11

of gen AI and large language

22:13

models like ChatGPT to improve the

22:15

readability of their own writing, of

22:17

course. However, they are accountable for

22:19

the accuracy of their publication, and

22:21

any use of AI must be

22:23

disclosed. So do you think academics

22:25

need to be specifically trained in

22:27

spotting? Well, it's useful if scientists,

22:29

you know, if the readership is

22:31

able to spot these things, but

22:33

I suppose, you know, it's not

22:36

going to be possible without tools.

22:38

And so the publishers and journals

22:40

are really at the front line

22:42

of this, and they are responding

22:44

to what many regard an integrity

22:46

crisis in scientific publishing. And they're

22:48

building up their defenses by expanding

22:50

the integrity department, but we also

22:52

have a choice of image integrity

22:54

tools, all of which mainly look...

22:56

for image duplications and these detection

22:58

tools are also attempting to detect

23:00

AI generated images. Is that using

23:02

AI to detect AI? That's right,

23:04

exactly. That's right, and at this

23:06

point it has to be said

23:08

that they are very unreliable, unfortunately.

23:10

Well thank you very much, Jano

23:12

Christopher, from the Federation of European

23:15

Biochemical Societies. And thank you, Gareth

23:17

Mitchell, for taking us on this

23:19

fascinating deep dive into AI. I

23:21

think it's probably raised more questions

23:23

than answers. Certainly has. And our

23:25

rough images will stay with me

23:27

for a while. It's been a

23:29

real pleasure, though, but thank you.

23:31

Absolutely pleasure. Now, though, we're going

23:33

to end today by turning away

23:35

from AI-generated content and looking up

23:37

at the night sky, because January

23:39

and February, when it's not been

23:41

raining, have given us some star-gazing

23:43

highlights, have been visible. I'm joined

23:45

now by Catherine Haymans, Astronomer Royal

23:47

for Scotland. Hello Catherine. Hello Vic,

23:49

have you been looking up in

23:51

the night sky and admiring the

23:53

planetary parade? Well I had a

23:56

little glimpse. The night before last,

23:58

but the cloud has just been,

24:00

as per usual, really dominating my

24:02

view of the night sky. But

24:04

yeah, I did have a little

24:06

bit. You'll often find me, Vic,

24:08

shaking my fists at the clouds.

24:10

Yeah, well, I'm in the northwest

24:12

of England. You're in Scotland, maybe

24:14

not best, with the most stargazing,

24:16

friendly weather. But talk history, Catherine.

24:18

What exactly is a planetary parade?

24:20

It's such a planetary parade. It's

24:22

such a lovely phrase. It is a

24:24

lovely phrase isn't it? As I said

24:27

for the last few months we've had

24:29

six planets up in our night sky

24:31

and in the last in this week

24:33

Mercury is joining the pack, so that

24:35

completes your planetary bingo card. You've got

24:37

Mercury, Venus, Earth. We're sat on, of

24:40

course, Mars, Jupiter, sat, Uranus, Neptune, all

24:42

up in the night sky, about half

24:44

an hour after sunset. You don't need

24:46

to stay out really late at night.

24:48

You don't need to get up early

24:50

in the morning, just eat your dinner,

24:52

go out, and tick all of those

24:54

planets off your card. The solar system on

24:57

parade. How often does this happen? How rare

24:59

is this happen? Yeah, so there are always

25:01

planets up in the night sky to look

25:03

at the earth, goes around the sun once

25:05

every year, so at some point during the

25:07

year you'll be able to see where the

25:09

planets are. But to have all seven up

25:12

in the night sky at the same time,

25:14

it isn't going to happen again until 2040,

25:16

so this is quite rare. But some of

25:18

my astronomy colleagues are a little bit grumpy

25:20

about the hype because actually it's not the

25:22

best time to see some of the planets.

25:25

So we've satin has been beautiful in our

25:27

night sky in our night sky up until...

25:29

a couple of weeks ago and now it's

25:31

getting really close to the sun and so

25:33

to see Saturn you're seeing it in the

25:35

glare of the sunset on the western horizon

25:38

as it's really hard to see Saturn at

25:40

the moment and Mercury the same it's only

25:42

just popped out from the glare of the

25:44

sunset and in a few weeks time it's

25:47

going to be much easier to see Mercury

25:49

but by then Saturn will have gone. So

25:51

we're kind of in this sweet spot right

25:53

now where we've got all seven up. actually

25:55

really hard to see mercury in Saturn at

25:58

the moment and also Uranus and Nietzsche. you

26:00

always need to telescope for anyway. But

26:02

Venus, really easy to see, super bright

26:04

in the West if you... think it's

26:06

an airplane but it's not moving, that's

26:08

Venus. You've got Jupiter right up above

26:10

your head at the moment just after

26:12

about six o'clock in the evening and

26:14

if you draw a line in your

26:16

mind between Venus and Jupiter from the

26:18

west across the Jupiter and then extend

26:20

that in an arc all the way

26:22

across to the east then you will

26:24

hit the red planet of Mars. Is

26:26

there a time that should we be

26:28

going out in the middle of the

26:30

night? Is this right after sunset when's

26:32

the best time to get the best

26:34

display? If you want all seven planets

26:36

at the same time, you've got a

26:39

very short window, about half an hour

26:41

after sunset. So head out about six

26:43

o'clock. I would advise people to download

26:45

an app. on your smartphone to cheat.

26:47

So there are lots of different star

26:49

apps. You can get one called Stellarium

26:51

that's free. And then you can just

26:53

point your smartphone up at the night

26:55

sky and it will tell you where

26:57

everything is. And that's a much easier

26:59

way of particularly Uranus and Neptune, which

27:01

you can't see with your own eye

27:03

anyway. But as the night goes on,

27:05

Mercury and Saturn will set. But you've

27:07

still got Venus Jupiter and Mars shining

27:09

bright. for a lot of the night.

27:11

And that will carry on throughout the

27:13

rest of March. And mercury is going

27:15

to get easier to see as well.

27:17

Wonderful, get your coats on, take your

27:19

smartphones and then once you've picked out

27:21

the planets, put the phones away and

27:23

just stare up at the night sky.

27:25

Thank you very much indeed Catherine. It's

27:27

been an absolute pleasure to talk to

27:30

you about the planetary parade. But that

27:32

is all the night sky wonder. But

27:34

that's an absolute pleasure to talk to

27:36

you about the planetary parade. But that

27:38

is all the night sky wonder and

27:40

disturbing AI-generated imagery that we have time

27:42

for this week. You have been listening

27:44

to BBC inside science with me, Victoriais.

27:46

Wales and West. Do you think you

27:48

know more about space? than

27:50

we do, head to

27:52

to .co .uk, search for

27:54

BBC Inside Science and

27:56

follow the links to

27:58

the Open University the

28:00

try the Open

28:02

University Space to try the

28:04

if you have any

28:06

questions or comments

28:08

for the Inside Science

28:10

if you do contact

28:12

us by email or comments

28:14

bbc .co .uk. Until

28:16

next time, thanks for

28:18

listening. on Inside Science at bbc.co.uk.uk.

28:21

Until next time, thanks for listening.