Essays and (Computational) Physics

There have been more news stories about ChatGPT and assessment in universities going around. There’s one here from The Journal and another here from The Conversation to give just two examples.

I wrote about this myself a couple of months ago in a post that included this:

I have to admit that I’ve never really understood the obsession in some parts of academia with “the student Essay” as a form of assessment. I agree that writing skills are extremely important but they’re not the only skills it is important for students to acquire during the course of a degree. Of course I’m biased because I work in Theoretical Physics, an area in which student essays play a negligible role in assessment. Our students do have to write project reports, etc, but writing about something you yourself have done seems to me to be different from writing about what other people have done. While forms of assessment in science subjects have evolved considerably over the last 50 years, other domains still seem to concentrate almost exclusively on “The Essay”.

Whatever you think about the intrinsic value of The Essay (or lack thereof) it is clear that if it is not done in isolation (and under supervision) it is extremely vulnerable to cheating.

A few people have retorted that communication skills are very important in higher education. I agree with that wholeheartedly, but it seems to me that (a) there are other ways of communicating than via formal essays and (b) there are, should be, more to academic study than  writing about things.

That said, I do think we could be doing more in some disciplines, including my own, to cultivate communication skills in general and writing skills in particular. In Theoretical Physics we certainly don’t do this as much as we should. I do have a project report in my 3rd Year computational physics module, but that is a relatively short document and the report itself counts only one-third of the marks (and the project is only 40% of the module mark).

These thoughts somehow reminded me of this. You can click on it to make it bigger if it’s difficult to read. It was the first paper (called colloquially Paper Zero) of my finals examination at the University of Cambridge way back in 1985, getting on for 40 years ago:

As you can probably infer from the little circle around number 4, I decided to write an Essay about topic 4. I’ve always been interested in detective stories so this was an easy choice for me, but I have absolutely no idea what I wrote about for three hours. Nor do I recall actually ever getting a mark for the essay, so I never really knew whether it really counted for anything. I do remember, however, that I had another 3-hour examination in the afternoon of the same day, two three-hour examinations the following day, and would have had two the day after that had I not elected to do a theory project which let me off one paper at the end and for which I got a good mark.

Anyway, to get back to the essay paper, we certainly don’t set essay examinations like that here in the the Department of Theoretical Physics at Maynooth University and I suspect they no longer do so in the Department of Physics at Cambridge either. At the time I didn’t really see the point of making students write such things under examination conditions but then we didn’t have ChatGPT way back then. No doubt it could generate a reasonable essay on any of the topics given.

I am skeptical about whether any of my 3rd year computational physicists would use ChatGPT to write their reports, but they might. But ChatGPT can write Python code too. Am I worried about that? Not greatly. I’ve asked it to write scripts for the various class exercises I’ve set so far and the code it has produced has usually failed. It will get better though….

What should it mean to be an author of a scientific paper?

The implementation of artificial intelligence techniques in tools for generating text (such as ChatGPT) has caused a lot of head-scratching recently as organizations try to cope with the implications. For instance, I noticed that the arXiv recently adopted a new policy on the use of generative AI in submissions. One obvious question is whether ChatGPT can be listed as an author. This has an equally obvious answer: “no”. Authors are required to acknowledge the use of such tools when they have used them in writing a paper.

One particular piece of the new policy statement caught my eye:

…by signing their name as an author of a paper, they each individually take full responsibility for all its contents, irrespective of how the contents were generated. If generative AI language tools generate inappropriate language, plagiarized content, biased content, errors, mistakes, incorrect references, or misleading content, and that output is included in scientific works, it is the responsibility of the author(s).

The first sentence of this quote states an obvious principle, but there are situations in which I don’t think it is applied in practice. One example relates to papers emanating from large collaborations or consortia, where the author lists are often very long indeed, sometimes numbering in the thousands. Not all the “authors” of such papers will have even read the paper, so do they “each individually take full responsibility”? I don’t think so. And how can this principle be enforced as policy?

All large consortia have methods for assigning authorship rights as a way of assigning credit for contributions made. But why does “credit” have to mean “authorship”? Papers just don’t have thousands of authors, in the meaningful sense of the term. It’s only ever a handful of people who actually do any writing. That doesn’t mean that the others didn’t do any work. The project would probably not have been possible without them. It does mean, however, that pretending that they participated in writing the article that describes the work isn’t be the right way to acknowledge their contribution. How are young scientists supposed to carve out a reputation if their name is always buried in immensely long author lists? The very system that attempts to give them credit at the same renders that credit worthless.

As science evolves it is extremely important that the methods for disseminating scientific results evolve too. The trouble is that they aren’t. We remain obsessed with archaic modes of publication, partly because of innate conservatism and partly because the lucrative publishing industry benefits from the status quo. The system is clearly broken, but the scientific community carries on regardless. When there are so many brilliant minds engaged in this sort of research, why are so few willing to challenge an orthodoxy that has long outlived its usefulness.

In my view the real problem is not so much the question of authorship but the very idea of the paper. It seems quite clear to me that the academic journal is an anachronism. Digital technology enables us to communicate ideas far more rapidly than in the past and allows much greater levels of interaction between researchers. The future for many fields will be defined not in terms of “papers” which purport to represent “final” research outcomes, but by living documents continuously updated in response to open scrutiny by the community of researchers. I’ve long argued that the modern academic publishing industry is not facilitating but hindering the communication of research. The arXiv has already made academic journals redundant in many of branches of  physics and astronomy; other disciplines will inevitably follow. The age of the academic journal is drawing to a close. Now to rethink the concept of “the paper”.

In the meantime I urge all scientists to remember that by signing their name as an author of a paper, they individually take full responsibility for all its contents. That means to me that at the very least you should have read the paper you’re claiming to have written.

ChatGPT and Physics Education

Following on the theme of ChatGPT, I see that Phil Moriarty has done a blog post about its use in Physics Education which many of my readers will find well worth reading in full. His findings are well in accord with mine, although I haven’t had as much time to play with it as he has. In particular, it is easily defeated by figures and pictures so if you want to make your assessment ChatGPT-proof all you need to do is unlike lots of graphics. More generally, ChatGPT is trained to produce waffle so avoid questions that require students to produce waffle. This shouldn’t pose too many problems, except for disciplines in which waffle is all there is.

Phil Moriarty also done a video in the Sixty Symbols series, on that YouTube thing that young people look at, which you can view here:

I start teaching Computational Physics next week and will be seeing how ChatGPT does at the Python coding exercises I was planning to set!

ChatGPT in the Swampland

My inestimable PhD student Kay Lehnert has been having a look at the capabilities of the Artificial Intelligence platform ChatGPT at writing about string theoretical ideas, specifically the swampland conjectures. It’s remarkable what this does well but also notable what it doesn’t do well at all. What he found was so interesting he wrote it up as a little paper, which you can find on the arXiv here. The abstract is:

In this case study, we explore the capabilities and limitations of ChatGPT, a natural language processing model developed by OpenAI, in the field of string theoretical swampland conjectures. We find that it is effective at paraphrasing and explaining concepts in a variety of styles, but not at genuinely connecting concepts. It will provide false information with full confidence and make up statements when necessary. However, its ingenious use of language can be fruitful for identifying analogies and describing visual representations of abstract concepts.

It took arXiv a while to decide what to do with this paper as it doesn’t fit in any of the usual categories. The arXiv sections that usually cover string theory are General Relativity and Quantum Cosmology (gr-qc) and/or high-energy physics theory (hep-th), which was where it was originally submitted, but this isn’t really a string theory paper per se. After being held by the moderators for a while it eventually it appeared in Popular Physics (physics.pop-ph), cross-listed in Artificial Intelligence (cs.AI) & Computation and Language (cs.CL); the latter two are computer science categories, obviously.

Figure 2 of the paper, which you should read if you want to know what it represents!

The reclassification of this paper was perfectly reasonable. In fact with this, as with any other arXiv paper, the thing that matters most is that it it is freely available to anyone who wants to read it and is discoverable, i.e. can easily be found via search engines. In the era of Open Access, things will generate interest if they are interesting (and accessible).

We posted the following on the Maynooth University Theoretical Physics Department Twitter account, something we do whenever a new paper by someone in the Department comes out:

Judging by the number of views (101K) by this morning, this one certainly seems to be attracting interest! Hopefully this blog post will generate even more..

Finally, there might be people reading this blog who can suggest a journal that might consider publishing an article on this sort of subject? Whatever you think about ChatGPT I think it’s generating a lot of discussion right now, so the topic is… er… topical.

Please use the box below for any suggestions.

Essays, Mills and ChatGPT

I saw an article the other day about how “contract cheating” was endangering the integrity of Ireland’s universities. This refers to the problem of students outsourcing their assignments to professional essay mills. Given the enthusiasm that university managers have for outsourcing in other contexts I’d be surprised if they see this as a problem. Indeed, their response might well be to outsource the grading of assignments in a similar fashion. It does however raise questions about academic integrity for thus of us who care about such matters.

I have to admit that I’ve never really understood the obsession in some parts of academia with “the student Essay” as a form of assessment. I agree that writing skills are extremely important but they’re not the only skills it is important for students to acquire during the course of a degree. Of course I’m biased because I work in Theoretical Physics, an area in which student essays play a negligible role in assessment. Our students do have to write project reports, etc, but writing about something you yourself have done seems to me to be different from writing about what other people have done. While forms of assessment in science subjects have evolved considerably over the last 50 years, other domains still seem to concentrate almost exclusively on “The Essay”.

Whatever you think about the intrinsic value of The Essay (or lack thereof) it is clear that if it is not done in isolation (and under supervision) it is extremely vulnerable to cheating. The article I referred to above concentrates on the corrupting influence of “Essay Mills” who will produce – for a fee – an essay on a given topic.

I believe however that this is not the biggest threat to academic integrity. There is a lot of discussion going on these days about ChatGPT, an AI system that can produce text on demand using sources on the internet. This is not great at dealing with technically complex specialist topics but can produce plausible if somewhat superficial offerings in many circumstances where something less demanding is required. Indeed, the more banal the task the better ChatGPT does. For example, here is an AI version of a university Strategic Plan which captures the vacuous nature of such documents with uncanny accuracy:

According to a pilot programme of which I am aware, the present configuration of ChatGPT scores an average of about 70% on (short) student essay tasks. It won’t be long before it gets much better than that. I predict that something like it will soon put essay mills out of business as it will be far cheaper and its use far more widespread. This is the real threat to the viability of “The Essay” in a modern university. The response will need to be quick!

Update: piece from the Irish Times.

P.S. It’s worth mentioning that AI systems can also write simple computer code to a reasonable level of proficiency, so this vulnerability also affects programming tasks.

Guest Post – by A.I. Addio

Quite a few people have been playing around with a new-fangled AI tool called ChatGPT the developers of which say this:

We’ve trained a model called ChatGPT which interacts in a conversational way. The dialogue format makes it possible for ChatGPT to answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests. ChatGPT is a sibling model to InstructGPT, which is trained to follow an instruction in a prompt and provide a detailed response.

Here is an example (stolen from here) wherein this “model” creates the abstract of a scientific paper on a suggested topic:

This makes me wonder how many abstracts on astro-ph are actually written this way!

Please note that no papers of mine involved the use of any form of Artificial Insemination. I hope this clarifies the situation.