At the foundation of academic writing, there is a given: that you and the authors you are citing have made an effort to present factual arguments with citations when applicable.
ChatGPT doesn’t do that. And so I think it should almost never be cited.
—
I think of chatGPT as more of a writing coach. You tell it what you want to write about, and it puts together a sample response that, above all else, sounds convincing. If it happens to know relevant facts it will put them in (but not with citations). If chat GPT doesn’t know relevant facts, it uses made up facts as a placeholder, just to give an example of how a convincing response could be structured.
It’s then our job as an actual writer to take this sample, hunt down the relevant facts (with citations), and put forward our final argument.
—
Perhaps the only time it makes sense to me to cite chatGPT is if you are explicitly writing about it as an entity.
> ChatGPT doesn’t do that. And so I think it should almost never be cited.
I usually try to get ChatGPT to give me a source for its answer. At least for me in most cases it gives non working URLs. It's very frustrating. Or it was very frustrating, I've since stopped using it where I need to quote sources.
ChatGPT is not a useful source since its inability to tell you where it got the information from is no different than asking a family member and citing "Bob, my uncle"
If you have ChatGPT Plus, you get access to the Web Browsing model, which allows ChatGPT access to the Internet. Not only do you get an audit log for all the sites ChatGPT visit, you also get citations as well for any sources it does end up using in the generated text. Even when ChatGPT does fail, its audit log gives me links to several sources that I can consult instead.
The difference is that you aren’t the one coming up with the search terms, browsing through all the search results, and then selecting the specific results you want to explore further. You let the Web Browsing model do all that work for you.
I’ve used ChatGPT to generate value scenarios, possible use cases, and descriptions of use for internal white papers. If another colleague were helping with this, I’d credit them in the acknowledgements at minimum. If I had gotten a list from a published paper, I’d credit it — even if I changed it considerably.
I do this for three reasons: (1) giving credit; (2) letting readers find more details so they can better interrogate the paper; and (3) proving transparency of sources generally.
With that in mind, it seems I should credit ChatGPT, especially to meet the second or third reasons. And that crediting should include my prompts and, in an extended session, multiple prompts.
The hard part is the first reason credit. The way it works today, I can’t give credit to the pieces that help build the response I’m using in a very precise way. I’d love to see a better way of doing that.
For research uses, we built an Assistant on top of ChatGPT that uses our database of articles to ground its answers and give real references, might be useful for you or others on research papers at least: https://scite.ai/assistant
I have a very different experience with GPT-4. It always provides about 6-10 references for every response, and they work perfectly every single time. I realize the hallucination problem and check everything every time, but compared to 3.5 (which often got titles a bit wrong or mixed up some authors, etc), it's astoundingly good at providing reasonable and factual information most of the time when it gives references that work pretty much every time.
Definitely always check though. It's great for getting over staring at a blank page and getting writing on something where you probably already know the references that will be called anyway, such that you can edit out any subtle over reaches that are made.
I think if you significantly benefitted from ChatGPT in preparing the paper, then you should still cite it, just like you would cite a software package (e.g. scipy), even if in principle you could have done it by hand.
I didn't know that was standard, but where is the line? You would never cite e.g. Microsoft Word, right? Coincidently, that will soon have ChatGPT built-in.
APA for example says the following:
"... a reference is not necessary for standard software... Examples are Microsoft Word, Java, and Adobe Photoshop."
> Perhaps the only time it makes sense to me to cite chatGPT is if you are explicitly writing about it as an entity.
That specific situation is what TFA is discussing. It specifically says to discuss ChatGPT inside the paper itself, rather than confining that to the citation.
Text generated by ChatGPT that you have reproduced should be cited, if for no other reason than the fact that you are not the author. Taking credit for work that is not your own is unethical.
Generally speaking, plagiarism (as in, passing off other's words and ideas as your own) is not illegal except in cases where it intersects with copyright law. In the case of ghost writers, they transfer copyright to the publisher, so legally speaking there's no issue.
In academic writing, there's a standard that the authors are the sole authors of the text that is upheld by academic institutions, academic communities, and the academic publishing industry (but not the legal system itself). This same standard doesn't exist in the same way in non-academic publishing (or a lot of other media) so not crediting ghost writers is considered acceptable.
I would assume that most books written by CEOs, professional athletes, general celebrities, and the like are significantly cowritten with someone whether they were entirely ghostwritten or not.
Our university recently recommended the same thing, and I think this is a very bad idea for two reasons.
1. Not every sequence of words deserves to be cited. GPT-3 and ChatGPT are often confidently wrong about facts, why would you want to add a citation to this? When writing a paper, this needs to be fact-checked anyways, so why not add the original (actual) source?
2. It also breaks the citation graph. Imagine all papers now point to a catch-all reference from OpenAI (2023). Adding a citation is about saying where you got certain information from and this current format doesn't give enough to do that, it just points to the catch-all. With any other citation you can either look up the paper or—in the rare case of personal communication in a citation—ask the cited source directly. You can't ask chatGPT "hey, why did you say this in paper X" and expect a meaningful answer.
This is an exciting innovation! Making ChatGPT output a first-class and citeable source will massively increase the rate at which academic psychology generates research output, without significantly changing the quality of data or the soundness of conclusions drawn.
You are clearly making a joke, but like so many jokes I think there is some truth in it! It might even improve the quality or at least make it easier to spot false information.
Funny also how the APA article ends with
> We, the APA Style team humans, appreciate your patience as we navigate these unique challenges and new ways of thinking about how authors, researchers, and students learn, write, and work with new technologies.
So, they see this as an unique "challenge" only. Not as an opportunity.
In principle, citations should give credible sources of information to the reader and let them reproduce the result ultimately. However, this is impossible for ChatGPT because it generates answers probabilistically and its random number generator cannot be controlled. (Though perhaps this could be partially achieved by setting the temperature to zero through the API?)
I hope open-sourced models will achieve better performance soon.
The only time ChatGPT should be cited is when the purpose of your paper is to study the behavior of ChatGPT.
I think it should be morally okay to use ChatGPT in the research process and to improve the quality of your writing, e.g., "Make this more clear and flow better"
If ChatGPT must be cited, I think for maximum transparency the citation should include the model, date retrieved, prompt, and output in an appendix.
I really don't get why people are being so weird about this. People are writing papers about Llms, and citations aren't only for "knowledge" that you're referencing.
I could see psychology as a field exploring the idea that some of human cognition is LLM-like, which could make papers including examples of LLM output more common in psych fields.
I don't think it says what to do if you just use ChatGPT for rephrasing and grammar. I guess you just don't cite it, in the same way you wouldn't cite Microsoft Word.
ChatGPT doesn’t do that. And so I think it should almost never be cited.
—
I think of chatGPT as more of a writing coach. You tell it what you want to write about, and it puts together a sample response that, above all else, sounds convincing. If it happens to know relevant facts it will put them in (but not with citations). If chat GPT doesn’t know relevant facts, it uses made up facts as a placeholder, just to give an example of how a convincing response could be structured.
It’s then our job as an actual writer to take this sample, hunt down the relevant facts (with citations), and put forward our final argument.
—
Perhaps the only time it makes sense to me to cite chatGPT is if you are explicitly writing about it as an entity.