AI was nice till it was not.

When I arrived at Oxford, I skilled what many would name a cultural shock. Beyond the flamboyant events and visual micro-aggressions of the gowned elite, my physics and math programs felt radically totally different than earlier than. The lectures have been considerably extra impersonal; the professors have been a bit hostile; the group was enormous, and everybody had one thing to show. This would have felt relatively regular if I had determined to proceed my schooling in Argentina, however, as a substitute, I had tailored to the American schooling mannequin—at a faculty like Northwestern no much less, which had a reasonably small physics program. Not solely that, I spent half of my undergraduate schooling surrounded by PhD college students, taking the identical lessons as them.

But a masters in England was totally different. It felt greater, extra ominous. Most folks opted to review and clear up downside units alone, and once we tried to work in teams there was an expectation to carry out because the “genius” sort. I had PhD presents from Harvard and Oxford straight from undergrad, and I felt like I needed to show I deserved it. It was partly my fault. I did really feel the necessity to brag so as to be taken significantly, nevertheless it backfired on me. It was awkward at first, however the surroundings shifted drastically mid Michaelmas time period. Pretty quickly our discussions and ego video games shifted from meaningless whiteboard debates to my classmates merely stating “I simply requested ChatGPT how one can do it.”

ChatGPT?

This was in late 2024. I had heard of AI up to now. Everyone was raving about all of it yr, however nobody I knew had used it for sophistication. I noticed folks complaining about how a lot water it used, how horrible it was for the surroundings and some folks even said that it would kill us all sometime. This felt like full nonsense to me. I recall after I was writing my undergraduate thesis in mid-spring. I opened instagram to procrastinate for a couple of minutes, and I noticed a well-liked “PhD influencer” (no matter meaning) say she put her title in ChatGPT, and it blasted quite a lot of details about her and the way nice she was. I believed it was a bit scary for a “web site” to drag up all of your info like that (and I discovered the endeavour a bit narcissistic), however I used to be deeply curious, so I gave the chatbot an opportunity. I keep in mind asking some query about scattering amplitudes, and it replied with nonsense. So, I pressed it and requested once more with extra element, and it mentioned:

“You ought to ask a theoretical physicist.”

And at that time, that was sufficient for me to shut the window and by no means open it once more.

That is, till a number of months later, I used to be learning at Oxford with my “buddies” on the time, they usually stored raving about how ChatGPT might do their homework. Problem units for a lot of the MMathPhys programs at Oxford usually are not graded. Real grades come from exams or your dissertation. A couple of lessons require you to submit homework, however ‘passing’ is sufficient. I believed it was a bit meaningless to ask a chatbot to do all the things for you when your complete function of the homework was to function examination apply. Yet, in January 2025, I discovered myself caught attempting to unravel one QFT query, so I requested the chatbot like my buddies beneficial, and, once more, it mentioned some bullshit. I didn’t have a look at it once more till a number of months later.

It was a bit odd how the concept of “homework” had appeared to decay. Before, not less than, we had the decency to have a look at Stack Overflow, and scout for comparable issues or downside fixing methods that would assist, however now abruptly folks weren’t even doing that. They simply requested ChatGPT to do all the things. It defeated the aim of doing homework to be trustworthy. Though, I feel we had misplaced that some time in the past. I all the time thought that past training for the examination, downside units have been nice to be taught what collaboration throughout downside fixing is like.

But in actuality, collaboration has hardly ever been a part of my actuality as a scholar. The final time I recall actually collaborating with fellow college students was in my freshman yr of faculty. I used to be taking multivariable calculus with a bunch of engineering college students. I knew two women, Andrea and Juliana, from orientation and we instantly shaped a small group chat to work collectively. We would attempt every downside set individually first, after which every of us would ship our solutions to the group chat and we might debate. For occasion, one woman would ask me “oh, that’s a intelligent manner to do that query. How do you know we needed to combine over this curve? Can you clarify it to me?” or inform me “Hey, I feel this query is flawed! You ought to use Lagrange multipliers this manner as a substitute.” And I might do the identical factor for them. I realized loads in that class, and I partly owed it to my nice collaborators on the time. Debating, discussing and simply studying from them was genuinely pretty. It helped me turn into higher at math and actually grasp all of the ideas.

I used to be deeply excited to begin the physics main and collaborate with different physics friends. In the autumn time period of my sophomore yr, I signed up for an honors differential equation math class with a buddy who was additionally a physics main. As I believed was customary, we despatched one another our options to test. It was my first time taking an honors math class, whereas that buddy had taken three earlier than. I additionally needed to miss the primary two lectures since my educational advisor insisted I wasn’t prepared for honors, and I spent every week on a deeply boring customary class earlier than I satisfied the professor to let me swap. That is to say, I made a ton of errors in that first homework. So I messaged my buddy like I had finished in previous and began asking: “Oh, it appears I misunderstood this idea, what do you imply right here? Why is that this true? I missed this class, so I believed we needed to do one thing totally different” to which she accused me of attempting to plagiarize her work and began ignoring me fully, and telling different folks I used to be attempting to repeat her. Our friendship was by no means the identical after that. And I realized a tough lesson: I used to be alone. For the remainder of the course, I labored on homework alone, I studied for exams alone, and I ended up with an A within the class. Actually, I spent most of my diploma working alone, whilst an undergrad amongst PhD college students. I might clarify my solutions to my classmates, however I hardly ever had that backwards and forwards from freshman yr once more.

That being mentioned, ChatGPT simply made the surroundings surrounding collaboration extra bleak. Now, there was nothing past “Just ask ChatGPT.”

I used to be nonetheless hesitant to make use of AI. I noticed many individuals within the MMathPhys use it for not solely their homework, however for his or her dissertations and the analysis they anticipated to publish. I used to be confused. Another buddy began utilizing it as a therapist. I used to be nonetheless deeply confused. I used to be repelled by how agreeable it was, nevertheless it did give the bizarre phantasm of getting somebody behind the display screen.

Regardless, in May of 2025, all my dangerous choices caught as much as me. I truly spent most of my time at Oxford doing nothing. I all the time thought-about attending lectures some type of social ritual. They have by no means helped me. I’ve all the time realized from books. Similarly, some downside units have been dense, uninsightful and ineffective, and in the event that they weren’t wanted for my grade, I wasn’t bothered to do them. I additionally had a PhD supply ready for me at Harvard, so there was actually nothing to lose. I as a substitute traveled loads, went to events, embarrassed myself, learn a ton of books, performed chess, fed the geese and cried loads. But in some unspecified time in the future, I noticed I needed to get my masters diploma and I had a month and a half to be taught 4 Oxford MSc lessons from scratch and move my exams. I used to be cooked.

And there was ChatGPT. The factor about ChatGPT is that, properly, it’s not an individual. It doesn’t “suppose” like we do. But I noticed it was very helpful for one factor: rearranging what people mentioned in a extra digestible manner. So, I might sit with my cosmology and common relativity notes, and each time I bought caught or I couldn’t visualize a definition, I might simply screenshot or copy the small part or sentence I didn’t perceive and ChatGPT would re-explain it to me. Sometimes it included a fast instance that helped me navigate index notation, as an example, or it might clarify to me that torsion-free simply meant that parallelograms closed.

One of the primary incompatibilities with customary physics/math schooling and the way in which I take into consideration issues is that definitions and derivations in a vacuum are a bit ineffective to me. I can observe sophisticated math and I can perceive what a definition says, nevertheless it received’t click on for me till I see an actual illustrative instance. This is not only a me factor. This is how Feynman thought of physics as properly, and I’m certain lots of people share this cognitive profile.

But as a substitute of coping with friends that bought irritated at me and my stupid questions, ChatGPT supplied that bridge virtually immediately and with no fuss. After three weeks of utilizing it, I noticed I didn’t want it as a lot anymore. I might observe superior ideas—like Penrose diagrams or Kerr Black Holes—alone, make my very own illustrative diagrams and derivations with out AI. I took my exams two weeks after that. I realized the content material General Relativity I&II, Cosmology and Geophysical Fluid Dynamics in a month and a half from scratch, whereas additionally taking the 2 Standard Model programs and a QFT in curved spacetimes class. My grades have been between excessive passes and deserves, nevertheless it was sufficient! I bought my diploma. And these have been absolutely closed booked exams, so I did the work myself. But ChatGPT actually helped me speed up my information acquisition.

That being mentioned, I didn’t belief it in any respect for brand spanking new analysis. Once I arrived to Harvard, I used to be given my first PhD undertaking in early November of my G1 yr, which is a bit early although not uncommon for prime power. I used to be working with a postdoc, and I used to be pleasantly shocked to see that the abilities I realized in undergrad from my two mentors have been extremely helpful for that downside. I embedded some twistor concept and chiral algebras into a pleasant mathematica pocket book, which helped with quite a lot of calculations. I barely used AI on that undertaking. It helped me discover background literature and interpret some paragraphs/derivations in these papers. Beyond that, I solely used Claude Opus to patch one piece of code in my library to assist me cope with poles and residues in an expression as a result of I wasn’t certain how one can do it. That was it.

However, the single minus paper actually modified the expectations surrounding AI in physics. Now, an inside mannequin of OpenAI had discovered one thing researchers tried to work out for months earlier than. That was cool, nevertheless it had nuance. Numerous the heavy lifting of that paper and the conceptual leaps have been finished by actual folks, however the accomplishment of AI was to not be understated. It was spectacular. The proof was spectacular and never simple, and this paper opened quite a lot of analysis instructions in Celestial Holography. AI is definitely nice at proving issues inside a given framework. It’s additionally good at issues that observe a strict logic or offering complicated mathematical buildings which are sound, however that’s not all that physics is. Or, I wish to imagine so.

The difficulty is that each time I confirmed my little mathematica pocket book that I’ve coded principally myself wanting on the Wolfram documentation, each my seniors and PhD friends simply would ask:

“Can’t ChatGPT try this?”

Maybe. But coding an issue in mathematica is not only about “effectivity” or writing quick code. It is about encoding a concept, a framework, working with nook circumstances, and studying. Maybe I might have requested Claude or ChatGPT to make a chiral algebras mathematica library, and it may need been good or higher than mine, however then what could be the purpose? Maybe it might have been sooner, barely sooner. Maybe I might have fed it all of the various things I needed to do, and let it code on the background and spit a ton of solutions… however as a primary yr scholar that appeared to defeat your complete function of doing analysis within the first place. Furthermore, ChatGPT is just not infallible. It may be very flawed. And not noticing when it’s flawed is simply embarrassing for everybody concerned.

I don’t imagine within the grind of doing a tough integral or proof from scratch to show you might be man sufficient to deal with it (if a calculator or mathematica can do it, why endure?), however I do suppose there’s something useful is fixing an attention-grabbing downside, understanding a construction, and doing a lot of the conceptual heavy lifting. I completed my first undertaking with very restricted AI use, and the paper was revealed.

But when it was time to begin my second undertaking as a contemporary G2, I did discover the surroundings had modified drastically. If you stroll throughout the Harvard physics division and lean into the workplaces, lots of people could have ChatGPT, Codex, Claude or comparable open of their screens. Numerous the emails I began to obtain contained AI generated notes or notes based mostly on concepts given by ChatGPT. It wasn’t simply me, quite a lot of my friends undergo this too, right here and in all places. This is a standard expertise now. Now, “brainstorming” means speaking to ChatGPT. Numerous my time is spent deciphering ChatGPT notes, realizing it’s nonsense, and shifting on to the subsequent.

I used to be aggravated. I had objections, questions, and considerations, however I used to be shut down and advised to ask ChatGPT. So, I requested ChatGPT and it simply “validated” all the things I mentioned, however didn’t suggest any options. The downside is that I needed to debate with my superiors, ask questions… and I couldn’t. I noticed quite a lot of doing a PhD in theoretical physics is coping with loud folks screaming at a blackboard, so it doesn’t really feel very interactive both. Some folks say quite a lot of fancy phrases. Bullshitting is frequent apply some say. I noticed in some unspecified time in the future that it wasn’t simply me; there was some posturing, some efficiency, and ChatGPT made that worse. So, I used to be alone with a chatbot once more. But, now, I didn’t have a set of lecture notes written by a high physicist at Cambridge a few concept that was established virtually a century in the past… I had a few papers, obscure concepts, and nowhere to go as a result of on a PhD you’re employed out issues nobody has finished earlier than.

And it wasn’t good. Because I’m only a second yr PhD scholar and I don’t know something. I do know quite a lot of issues, however within the grand scheme of issues, I’m inexperienced and I would like mentors. And a chatbot can’t be my mentor. And then I noticed that very same device that helped me a lot a yr in the past was drowning me and my area. I don’t suppose applied sciences are evil, however I do suppose how people make the most of some applied sciences may be dangerous. But it’s not AI’s fault, it’s ours.

I did understand early on that lots of people in my area have been much less centered on discovering the fact than they have been on projecting, being proper and exhibiting off. I noticed early on that for lots of physicists instructing and mentorship is extra of a burden than one thing they deem notably useful. We need the quickest PhD college students, those that revealed three papers in undergrad, those that end a calculation the quickest. We need the most well liked subjects and probably the most citations. But if a chatbot can change a PhD scholar, if it’s sooner, extra environment friendly and it could actually take up Weinberg’s QFT volumes in 5 minutes, why do we’d like college students in any respect?

Why can we collaborate once we can simply work alone with a bot?

Remember after I talked about I had that group chat with my engineering buddies and we have been continually evaluating and brainstorming? Sometimes, they’d catch an idea or see a clear resolution I didn’t discover, and I might try this for them too. That helped us develop, and it gave me a deep understanding of quite a lot of these ideas. I might clear up exams in surprising methods, however that was enjoyable! It helped actually grasp the mathematics and the way lovely it was. But that was additionally a wholesome collaboration. It wasn’t about getting the very best grade, or having the neatest resolution, gatekeeping or screaming. We would sit down, attempt it out, politely focus on, and clarify. We wouldn’t be imply or condescending, we might simply be taught from one another and enhance one another.

People maintain us in test. Each particular person is totally different. We all have a distinct background, a distinct manner of seeing the international stage. We come from totally different coaching. But, you be taught from others they usually be taught from you too.

And we’re dropping that now too. Should we fear that colleagues working in AI companies might take our prompts and try to get ahead to solve a problem before we do? Should we concern for threats? Should we concern that our days as PhD college students shall be spent telling a machine to “attempt tougher” as a substitute of deeply understanding one thing new?

Truth is, I used to be very impressed and enthusiastic about how ChatGPT helped me research. Would I relatively have responsive professors, lecturers that know pedagogy, physicists that wish to write comprehensible papers as a substitute of purposefully burying perception in meaningless jargon, and friends I can speak to with out having an nervousness spike? Yes. I might relatively have that a hundred percent. And I concern that now, it’s all going to be worse. Because we now have in our palms a machine that validates all of our biases and buries them in dozens of pages of meaningless math.

I’m only a PhD scholar who feels a bit defeated, not as a result of AI may take my job, I’m privileged sufficient for that to not be a priority (and let’s be trustworthy, I’m privileged sufficient to have the chance to do a PhD in theoretical physics). But as a result of the way in which we’re utilizing AI is destroying what made science enjoyable within the first place. Is it actually enjoyable to publish thirty papers in every week utilizing AI only for the sake of it? One of my buddies in this system rightfully asks “Are these papers value publishing?”

The total level of fixing an attention-grabbing downside or understanding the mysteries of the universe is to not win an award or have probably the most citations, however to get nearer to that just about legendary factor we name fact, to know the universe and ourselves via the research of the character of issues.

The downside is that the little chatbot in our pockets is making us extra remoted and in so, it’s making us stray farther from this humble purpose: to be taught. But it wasn’t simply the chatbot, it was the tradition round academia too. I feel chatbots are a mirror, a deeply validating sycophantic mirror with the information of the web. But mirrors may be harmful if we keep for too lengthy too charmed by our personal reflection, and we neglect about all the things and everybody else too. In specific, it might be fairly unhappy to destroy one of many few issues that makes our life significant, not the flamboyant prize or the flamboyant place, however merely having a neighborhood to share and focus on and be taught. To be taught issues. That is a good looking factor.

-Nia