Is ChatGPT useful yet for answering astronomy questions?

In summary, ChatGPT is a search engine that can be used to find the answers to astronomy questions. It seems inevitable that internet searches to answer astronomy questions will someday involve AI assistance. However, ChatGPT is not reliable enough yet for that, and still nowhere close. Some people have experimented with ChatGPT to see if it is better than simply using search engines to answer astronomy questions, and it seems to be a convincing human-sounding search engine. However, it is very hard to tell if the answers are correct, unless you know the answer yourself!
  • #1
Ken G
Gold Member
4,897
538
TL;DR Summary: It seems inevitable that internet searches to answer astronomy questions will someday involve AI assistance. Is ChatGPT reliable enough yet for that, or still nowhere close?

The issue for me at this point is how to distinguish the errors ChatGPT currently makes, versus the errors that are widespread on the internet that an unwary user could fall victim to even without AI assistance. Has anyone experimented with ChatGPT to see if it is better than simply using search engines to answer astronomy questions? Personally, I've only tested questions that I know regular internet searches will likely give wrong answers, and ChatGPT gives all those same wrong answers (questions like if fusion is necessary for hydrostatic equilibrium, or why are higher mass main-sequence stars more luminous, ChatGPT would have no hope of getting those right given all the misinformation out there). What happens if ChatGPT is used instead of, say, posting to a forum like this? I'll start with an example, using a question on this forum:
Me:
Could a brown dwarf ever have been a T Tauri star in its past?
ChatGPT:
Yes, a brown dwarf could have been a T Tauri star in its past.
T Tauri stars are young, pre-main-sequence stars that are still undergoing gravitational contraction and have not yet reached the point of sustained fusion in their cores. Brown dwarfs, on the other hand, are objects that are too small to sustain fusion in their cores, and they emit most of their energy as heat from their contracting cores.
However, T Tauri stars and brown dwarfs are thought to be similar objects in the early stages of their formation, and they may share many of the same physical characteristics. In fact, some brown dwarfs have been observed to have disks of gas and dust around them, similar to those seen around T Tauri stars, which suggests that they may have formed through a similar process.
Therefore, it's possible that a brown dwarf could have been a T Tauri star in its past, when it was still undergoing gravitational contraction and had not yet reached the point of sustained fusion in its core. Over time, as the brown dwarf continued to contract, it would have become denser and hotter, until it reached a stable state where it could no longer sustain fusion in its core.

So a decent answer in some ways, though technically wrong, and with a very weird ending!
 
Physics news on Phys.org
  • #2
Ken G said:
The issue for me at this point is how to distinguish the errors ChatGPT currently makes, versus the errors that are widespread on the internet that an unwary user could fall victim to even without AI assistance.
That's why we require mainstream peer-reviewed journal references or mainstream textbook links in technical threads at PF. That is the best solution we have found so far. :wink:
 
  • Like
  • Wow
Likes bagasme, Tom.G, russ_watters and 2 others
  • #3
[Mentor Note -- Since this thread start does not involve any technical question, it has been moved from the Astrophysics forum to the General Discussion forum]
 
  • Like
Likes Tom.G
  • #4
Answering? No problem.
Correctly answering? That's a whole 'nother kettle of fish.
 
  • #5
Ken G said:
The issue for me at this point is how to distinguish the errors ChatGPT currently makes, versus the errors that are widespread on the internet that an unwary user could fall victim to even without AI assistance. Has anyone experimented with ChatGPT to see if it is better than simply using search engines to answer astronomy questions? Personally, I've only tested questions that I know regular internet searches will likely give wrong answers, and ChatGPT gives all those same wrong answers....
So is ChatGPT an AI (or ai?)? Seems to me that it is just an eloquent search engine. It doesn't answer questions it just finds the answers to questions.
 
  • Like
Likes Lord Jestocost
  • #6
It does try to answer questions. Actually it sounds pretty convincingly human, rather remarkable that way, but it's very hard to know if the answers are correct, unless you know the answer yourself! Let me give an example. If you ask is, "why are high-mass main-sequence stars also high luminosity stars", this is what you get:
"
High-mass main-sequence stars are also high luminosity stars because their high mass allows for strong gravitational forces that generate very high temperatures and pressures in their cores. This leads to a rapid and intense fusion of hydrogen nuclei, releasing a large amount of energy in the form of light and heat.

The rate of energy production in a star's core is proportional to its mass raised to a high power, roughly to the power of 3.5. Therefore, a star with a mass that is significantly larger than the Sun will produce energy at a much higher rate than a smaller star, resulting in a much higher luminosity.

The luminosity of a star also depends on its surface area, which is proportional to the square of its radius. High-mass stars are much larger in size than low-mass stars, and therefore have a larger surface area to emit light from, contributing to their high luminosity.

Overall, the combination of high mass and large surface area results in high-mass main-sequence stars being very luminous objects."

So that sounds pretty good, doesn't it? But it contains both wrong elements and irrelevant elements, which someone sincerely asking the question would have no way of knowing. Then again, many of the same errors are in seemingly reliable websites, so it's not really ChatGPT's fault. Nonetheless, one could see trying it out on a question being asked on the forum, to get the discussion started.
 
  • Like
Likes FactChecker
  • #7
berkeman said:
That's why we require mainstream peer-reviewed journal references or mainstream textbook links in technical threads at PF. That is the best solution we have found so far. :wink:
Yes, I can see the issue. Believe it or not, ChatGPT will suggest references if you ask it to-- but like everything else, it only understands what a reference looks like, not what it actually is, so usually the references have to be interpreted and corrected a bit! You always have to make some corrections, with everything it does, but if you can do that, you might find some usefulness.
 
  • #8
Ken G said:
It does try to answer questions.
But it doesn't know the answers, it searches the internet for the answers, right? A search engine.
Actually it sounds pretty convincingly human, rather remarkable that way...
So an eloquent search engine then.

I feel like you think that it is an intelligence that you can bounce ideas off of. I don't think it is.
 
  • Like
Likes Bystander and Lord Jestocost
  • #9
russ_watters said:
But it doesn't know the answers, it searches the internet for the answers, right? A search engine.

So an eloquent search engine then.
Yes, it is basically an eloquent search engine. It removes a lot of the responsibility from the user, which is not necessarily a good thing but it's what will make it attractive. One-stop shopping, but buyer beware.
russ_watters said:
I feel like you think that it is an intelligence that you can bounce ideas off of. I don't think it is.
It's a language model that is trained on a huge database and finds patterns in what it scans through. Is that what intelligence does? Yeah, pretty much! But how well it does it, that's the rub. It had to give up on logical consistency early in its programming, this is the web after all! An interesting thing you can do is correct its mistakes, and it will modify its answers in response. It is basically programmed to humor you.

An interesting quote from B.F. Skinner:

The real question is not whether machines think but whether men do. The mystery which surrounds a thinking machine already surrounds a thinking man.​

 
  • Like
Likes russ_watters
  • #10
It depends on the definition of intelligence. I have a policy to never argue about definitions.

I believe that ChatGPT could earn a BA in the humanities (!) at most colleges. Maybe any college. I also expect that within a few years an AI will be able to earn a degree in any discipline at any college. Nevertheless one may feel certain that many will say such an AI is not intelligent. Before long an AI will be able to graduate summa cum laude in any subject. Critics will not change their view.
 
  • Like
Likes russ_watters and FactChecker
  • #11
How long until they are teaching at the colleges?

Consider this:
Me:
Write me a 5 minute lecture on the Hertzsprung-Russell diagram.

ChatGPT:
"Sure, here's a 5 minute lecture on the Hertzsprung-Russell diagram:
The Hertzsprung-Russell diagram, or H-R diagram for short, is one of the most fundamental tools in the study of stars. It's a scatter plot of stars' luminosities against their surface temperatures, or equivalently, their spectral types. It was first developed independently by Danish astronomer Ejnar Hertzsprung and American astronomer Henry Norris Russell in the early 20th century.
The H-R diagram is important because it reveals a great deal about the nature of stars. For example, it shows that most stars in the universe are main-sequence stars, which are stars that are burning hydrogen in their cores. These stars follow a well-defined band on the H-R diagram, with more massive stars being hotter and more luminous than less massive ones. The position of a star on the main sequence is primarily determined by its mass, which controls its core temperature and therefore its rate of energy production.
The H-R diagram also shows that there are other types of stars besides main-sequence stars. For example, there are giant stars, which are much more luminous than main-sequence stars of the same temperature. Giants have exhausted the hydrogen in their cores and have expanded and cooled as a result. There are also white dwarfs, which are very hot and very faint. White dwarfs are the end states of stars like the Sun, which have exhausted all their nuclear fuel and have collapsed to a very small size.
Another interesting feature of the H-R diagram is the instability strip, which is a region where pulsating stars are found. These stars change their luminosity and temperature periodically, making them very useful as standard candles for distance measurements.
The H-R diagram has also been used to study the evolution of stars. By comparing the position of a star on the H-R diagram to theoretical models of stellar evolution, astronomers can determine its age and mass, as well as its current and future evolution.
Overall, the Hertzsprung-Russell diagram is a powerful tool for understanding the nature and behavior of stars. Its simple yet elegant representation of the relationships between luminosity, temperature, and evolutionary stage has proven to be a cornerstone of modern astrophysics."

That really isn't bad, it might already be useful as a teaching assistant. What will it be in ten years?
 
  • Like
Likes FactChecker
  • #12
It lies and makes up references (google for fake doi returned by GPT), so no.
 
  • Like
Likes Bystander
  • #13
Hornbein said:
It depends on the definition of intelligence.
That depends on the distinction between syntax and semantics: Whether one makes an association between "intelligence" and "understanding of meaning/semantics" or not. Fact is that digital computers merely use syntactic rules to manipulate symbol strings, but have no understanding of meaning or semantics.
 
  • #14
Borek said:
It lies and makes up references (google for fake doi returned by GPT), so no.
It would be fairer to say it "approximates" references. We are used to exact references, but you can still use approximate citations to find the real ones, with a little poking around.
 
  • Like
Likes russ_watters
  • #15
Lord Jestocost said:
That depends on the distinction between syntax and semantics: Whether one makes an association between "intelligence" and "understanding of meaning/semantics" or not. Fact is that digital computers merely use syntactic rules to manipulate symbol strings, but have no understanding of meaning or semantics.
Yet, the distinction between semantics and syntax is not so obvious, it may be more a difference in complexity versus kind. Consider a mathematical proof, for example: we tend to imagine that we don't "understand" a theorem until we know why it is true, and that requires proving it. But of course, proving a theorem entails reducing it to a purely syntactic manipulation of a set of axioms! Indeed, it can be said that a central goal of mathematics is to remove all semantics and replace it with syntactics, on the grounds that "meaning" is nonrigorous, while symbolic manipulation is rigorous. So how can removing the meaning from a mathematical structure be the highest penetration into the meaning of that structure? I would say the connection between syntax and semantics is still not well understood, possibly even paradoxical.

To elaborate on that point, if we imagine what our brains are doing as we achieve "semantic understanding", we must admit all that is apparent is there are voltages and currents and so forth inside the cells of our brains. That all looks like purely syntactic manipulation of biological states, that could be translated into symbols and relations between those symbols: i.e., syntax. What exactly "breathes life" into that process, and gives us a sense that our thoughts "mean something", is a very deep question, which I believe is at the heart of the B.F. Skinner quote above.
 
  • #16
Ken G said:
How long until they are teaching at the colleges?

Consider this:
Me:
Write me a 5 minute lecture on the Hertzsprung-Russell diagram.

ChatGPT:
"Sure, here's a 5 minute lecture on the Hertzsprung-Russell diagram:
The Hertzsprung-Russell diagram, or H-R diagram for short, is one of the most fundamental tools in the study of stars. It's a scatter plot of stars' luminosities against their surface temperatures, or equivalently, their spectral types. It was first developed independently by Danish astronomer Ejnar Hertzsprung and American astronomer Henry Norris Russell in the early 20th century.
The H-R diagram is important because it reveals a great deal about the nature of stars. For example, it shows that most stars in the universe are main-sequence stars, which are stars that are burning hydrogen in their cores. These stars follow a well-defined band on the H-R diagram, with more massive stars being hotter and more luminous than less massive ones. The position of a star on the main sequence is primarily determined by its mass, which controls its core temperature and therefore its rate of energy production.
The H-R diagram also shows that there are other types of stars besides main-sequence stars. For example, there are giant stars, which are much more luminous than main-sequence stars of the same temperature. Giants have exhausted the hydrogen in their cores and have expanded and cooled as a result. There are also white dwarfs, which are very hot and very faint. White dwarfs are the end states of stars like the Sun, which have exhausted all their nuclear fuel and have collapsed to a very small size.
Another interesting feature of the H-R diagram is the instability strip, which is a region where pulsating stars are found. These stars change their luminosity and temperature periodically, making them very useful as standard candles for distance measurements.
The H-R diagram has also been used to study the evolution of stars. By comparing the position of a star on the H-R diagram to theoretical models of stellar evolution, astronomers can determine its age and mass, as well as its current and future evolution.
Overall, the Hertzsprung-Russell diagram is a powerful tool for understanding the nature and behavior of stars. Its simple yet elegant representation of the relationships between luminosity, temperature, and evolutionary stage has proven to be a cornerstone of modern astrophysics."

That really isn't bad, it might already be useful as a teaching assistant. What will it be in ten years?
I find it wordy. I reddened words and phrases that I do not feel are necessary.
I guess it spits out its first draft without edits, nor re-writes.
 
  • #17
It is a bit padded, yes, I'll bet if you ask it to write a paper for a literature class or some such thing it does that too-- but so do students trying to reach a page limit! What you can do with ChatGPT is suggest improvements, and that's how it can be used effectively. That's what you cannot do with simple google searches, you could easily uncover an essay on the topic and use that as your lecture, or take the essay and modify it to your needs, but ChatGPT lets you tell it how to modify it and it will try to do it. What I've found doing this, you reach a kind of point of "diminishing returns" where it becomes easier to just modify it yourself, so that's probably the mode you would use it-- you'd start with its answer, ask it to make some changes, and then take it from there yourself. Maybe not all that much different from what we do now with google searches, but there's no sifting through different sites. For better or for worse-- it saves time, but with a painful loss of accountability to the sources. It should definitely use citations like Wiki does, though right now it doesn't really "get" citations (as mentioned above).
 
  • #18
I do wonder just now if ChatGPT and Grammarly should get hitched up.
An excellent student/writer ..... that should make.
 
  • #19
One can imagine a day where, instead of only a standard TL;DR summary to get things started, the standard becomes the summary, the question, and the AI response, to get things started... and sometimes finished.
 
  • #20
Ken G said:
It would be fairer to say it "approximates" references. We are used to exact references, but you can still use approximate citations to find the real ones, with a little poking around.
Nope, occasionally it generates random doi's and claims these are existing papers, nothing approximate about it.
 
  • Informative
Likes nsaspook
  • #21
It turns out that just isn't the right way to make use of AI programs like ChatGPT. These programs are only able to manipulate language patterns, they have a very poor sense of "true" and "false", and are not adept at what we would call mathematical logic! (Try making sense of the internet using mathematical logic...) As such, they can see patterns associated with what we call a "citation", but they cannot create true and precise citations. This means the links and doi numbers they give are of no value at all, since they must be precise to be useful (and this is something future versions will definitely want to address, perhaps inventing some kind of "protected" information blocks.) In the mean time, never use an AI for precision.

What they can do, however, is point you toward potentially useful articles, if you recognize that their citations are "approximate" patterns of language. For example, if you want to find a good review article on brown dwarfs, you could always use many of the available search engines and do just fine, it's not an ideal purpose for ChatGPT. But if you did want to do it within the context of a ChatGPT session, perhaps because you are organizing your entire search experience that way, you could ask it for a citation, and get something like this:

ChatGPT:
Here's a recent review article about brown dwarfs:
Title: "Brown Dwarfs: Recent Advances and New Frontiers" Authors: Jacqueline Radigan, Aleks Scholz, and Katelyn Allers Journal: Annual Review of Astronomy and Astrophysics Link: https://doi.org/10.1146/annurev-astro-081920-042315
This review article covers recent advances in our understanding of brown dwarfs, including their formation, properties, and potential as probes of the universe. It also discusses new frontiers in brown dwarf research, such as the search for cool and substellar companions to nearby stars, the study of brown dwarfs in star-forming regions, and the use of brown dwarfs as benchmarks for testing models of low-mass star and planet evolution.
I hope you find this article informative and helpful!

Me:

I see that those authors write about brown dwarfs, but the three of them have never collaborated on a review article on the topic. Can you suggest a review article that was written with one or more of them in the author list?ChatGPT:
I apologize for the confusion. Here's a review article on brown dwarfs that includes at least one of the authors I previously mentioned:
Title: "Brown Dwarfs as Astrophysical and Fundamental Objects" Authors: France Allard, Derek Homeier, Adam Burrows, and Isabelle Baraffe Journal: Reports on Progress in Physics Link: https://doi.org/10.1088/0034-4885/76/1/016901

Note that all the links and doi numbers are still useless, as they need to be precise, and note that I had to assist in the process of finding a useful citation. Even the final citation is not actually a true article (again, the issue with "truth" rather than "language pattern"). But I can still use the approximate citation, by googling the title and author list, upon which I receive a link to a useful paper: https://arxiv.org/ftp/arxiv/papers/1903/1903.04667.pdf
If you look at that paper, you see it does not have the same title, and involves only two of the authors I googled, yet it is a white paper on brown dwarfs that does have a lot of useful information about them. Going a little farther down the google results, we also see: https://www.annualreviews.org/doi/abs/10.1146/annurev.astro.38.1.337
which is probably exactly what we are looking for, though it involves none of the original authors.

Now, of course you might ask, what good was ChatGPT if we end up having to peruse a google list anyway, and the answer is, not much-- except within the context of a session with ChatGPT that we are already examining other aspects of brown dwarf physics. So it might be that you are skeptical (for good reason) of part of the answer to some question you got, and you want to poke around some solid resources to investigate. You still don't need ChatGPT for that, but if you enjoy doing it this way rather than another way, it is still possible. The thing to remember is that you always have to involve yourself in the process, ChatGPT finds patterns but they are never guaranteed to be the truth, as with all its answers. That might sound like a significant limitation, and you'd be very right, but still if you try it, you might find that it is still quite a useful way to target specific answers that are harder to target with unguided search engines-- as long as you are willing to do some legwork to check the information.
 
  • #22
A corollary to all this is that at the moment, a language model like ChatGPT is best at tasks that involve pure language manipulation, like writing a poem or a creative story. If you want facts, ChatGPT does not distinguish those from any other language pattern, so it can sometimes craft the appearance of something correct that is not in fact correct at all. Nevertheless, so far in using it, I have found that it is not a bad first cut at any question of interest, even one that involves high-level reasoning, it's just very far from guaranteed to be true. It could be said that I have found the same thing about web articles from any number of sources, even university course websites and NASA popular articles, though humans are generally better able to stress the importance of avoiding outright contradictory statements.
 

1. Is ChatGPT able to accurately answer complex astronomy questions?

Currently, ChatGPT is still in its early stages and may not be able to provide accurate answers to complex astronomy questions. Its responses are based on the data it has been trained on, which may not cover all aspects of astronomy. Therefore, it may not be the most reliable source for answering such questions.

2. Can ChatGPT provide up-to-date information on astronomy?

ChatGPT is constantly learning and updating its knowledge base, but it may not always have the most recent information on astronomy. It relies on the data it has been trained on, which may not include the latest discoveries or developments in the field. It is best to consult other sources for the most up-to-date information.

3. How does ChatGPT handle conflicting information in astronomy?

As with any AI-based system, ChatGPT may struggle with conflicting information in astronomy. It may provide different answers to the same question depending on the data it has been trained on. It is important to verify the information provided by ChatGPT with other reliable sources.

4. Can ChatGPT provide explanations for its answers in astronomy?

Currently, ChatGPT does not have the capability to provide explanations for its answers in astronomy. Its responses are based on the data it has been trained on and may not necessarily include the reasoning behind its answers. It is best to consult other sources for detailed explanations.

5. Is ChatGPT useful for beginners in astronomy?

ChatGPT may be useful for beginners in astronomy as it can provide basic information and definitions. However, it may not be able to answer more complex questions or provide in-depth explanations. It is recommended to use ChatGPT in conjunction with other sources for a better understanding of astronomy.

Similar threads

  • General Discussion
3
Replies
102
Views
7K
Replies
190
Views
9K
  • Computing and Technology
7
Replies
212
Views
8K
  • STEM Educators and Teaching
Replies
33
Views
3K
  • Astronomy and Astrophysics
Replies
3
Views
1K
  • Astronomy and Astrophysics
Replies
21
Views
1K
  • General Discussion
Replies
4
Views
666
  • Astronomy and Astrophysics
2
Replies
49
Views
2K
Replies
10
Views
2K
Back
Top