18 Brown University students dropped course after professor announced in-person final

Full text

For today’s students, the temptation to cheat is at an all-time high, with AI tools making it easier than ever. In one recent class at Brown University , an economics class of 86 students, students scored higher on the (take-home) midterm exam than ever before: with an average score of 96 percent. Only 5 students out of 86 scored below 90 percent on the midterm: in a class where midterm scores had never exceeded 80 percent before. Strongly suspecting cheating, the professor announced that the final would be in-person instead.

18 students dropped the class. 9 students didn’t show up for the final. And of the 59 who took the final, the average score was 48.6 percent, with only one student achieving a score of 90 percent or more. The professor, Roberto Serrano, put it bluntly :

“We cannot afford to have a society in which a significant fraction of our best young minds think that cheating is OK. That leads to a declining society, to a failed society … We cannot choose to become idiots.”

Despite all that AI can actually do, the most dangerous thing it can do is erode our skills: a real danger if we outsource our actual learning, our critical thinking, and the time we spend struggling with puzzles and problems to a tool that purports to do it for us. This is not a new problem, but merely one that has worsened in the era of Large Language Models (LLMs). In fact, Einstein warned about this problem himself, leading us to be confident that if he had been around for the modern era, Einstein wouldn’t have used AI at all. Here’s how we know.

In 1896, Einstein was admitted into Zurich Polytechnic Institute, where he would graduate in 1900: 4th in his graduating class of 5. Although he was held in low esteem by the most famed and prestigious of his professors, he would go on to surpass them all.

Credit : Zurich Polytechnic Institute

On the one hand, there’s the myth of Einstein that most of us grow up hearing. According to the myth:

Einstein was an outsider: interested in physics, but flunked and left his academic studies.

Einstein was working at a patent office, rather than engaging in formal training in physics.

Einstein was a lone genius, working on his own, when he came up with relativity, mass-energy equivalence, Brownian motion, and the photoelectric effect.

And because of his unique and singular brilliance, he was able to single-handedly revolutionize physics, first with his “miracle year” in 1905, and then, 10 years later, with his greatest achievement of all: the publication of general relativity, overthrowing more than two centuries of Newtonian gravity.

It’s important to recognize that very little of this is actually true. Einstein was no outsider, but rather completed his physics degree at one of the top Universities in Europe: what is today ETH Zürich. He finished his undergraduate studies in 1900, and then — contrary to the myth you may have heard — he remained there, continuing his studies at the same place.

“But didn’t he work at the patent office?”

Yes, but that was because he didn’t earn the equivalent of either a Teaching Assistantship (TA) or Research Assistantship (RA) at his University.

Einstein, contrary to the popular narrative, wasn’t a lone genius, but rather only achieved the successes that he did because of his friends, colleagues, professors, and the larger community of physicists, astronomers, and mathematicians that he was a part of. Without them, including his study-buddy friends Conrad Habicht and Maurice Solovine, pictured alongside him in 1903, his ideas, brilliant as they were, would likely have gone nowhere.

Credit : Emil Vollenweider und Sohn/Public Domain

Instead, it was Einstein’s friend and classmate, Marcel Grossmann , who reached out to his father to try to get Einstein a job. Grossmann’s father had connections that could open up a position for a graduate student to work at the Swiss Patent Office: the same way any student would get work/study or a part-time job to help support them throughout the 20th century. Since Grossmann himself already had an assistantship, he used his connections to get his friend a job, enabling Einstein to concurrently continue his studies.

Einstein also was far from a lone genius in any regard. He wasn’t working alone at any point, but rather founded a group in 1902 known as the Olympia Academy : a study and discussion group about physics, philosophy, literature, and mathematics. Along with Conrad Habicht and Maurice Solivine (and sometimes also joined by Grossmann and others ), the group studied:

their own works,

books by scientists Karl Pearson, Ernst Mach, and Henri Poincare,

and philosophy and literature works by John Stuart Mill, Baruch Spinoza, and Miguel de Cervantes.

In other words, it is true that Einstein was working at the patent office in the years between completing his undergraduate degree (in 1900) and his miracle year (in 1905), but he was also engaged in formal graduate studies and augmented that with group study with others, in addition to his independent work.

Different observers will mark different times and different spatial locations as far as the occurrence of events is concerned. However, for every observer in all frames of reference, the quantity known as the spacetime interval (or Einstein interval, as Minkowski dubbed it) will remain invariant.

Credit : Maschen/Wikimedia Commons

It’s also a myth that Einstein flunked his studies, although there is a grain of truth there. Einstein’s most famous teacher at the Eidgenössische Polytechnikum (today’s ETH Zürich), the great mathematician Hermann Minkowski , had Einstein as a student during his undergraduate years. Minkowski remembered Einstein as:

“ always skipping lectures ,”

“ being a real lazybones … who never bothered about mathematics at all,”

and as having a “ mathematical education [that] was not very solid .”

However, that is not atypical of a large fraction, and perhaps even a majority, of people who go on to become excellent scientists as adults. Very few teenagers (and Einstein was 17 when he began his undergraduate studies) have the discipline, study habits, and foundation to immediately be successful in a formal academic setting.

In other words, Minkowski’s evaluation of Einstein was based on the only thing that a college professor can justifiably evaluate a student on: performance. But performance is not the same thing as potential. Minkowski couldn’t see the large, sustained effort that Einstein would put into his education into the future, into improving his mathematical foundation, or into learning how to think long and deeply about problems. He couldn’t see the skills that Einstein would develop, or foresee what would sufficiently motivate Einstein to make that big effort that would lead to his revolutionary discoveries.

The identical behavior of a ball falling to the floor in an accelerated rocket (left) and on Earth (right) is a demonstration of Einstein’s equivalence principle. If inertial mass and gravitational mass are identical, there will be no difference between these two scenarios. This has been verified to better than ~1 part in one trillion for matter through torsion balance experiments, and was the thought (Einstein called it “his happiest thought”) that led Einstein to develop his general theory of relativity. Recently, the ALPHA-g experiment confirmed that this is true for antimatter as well.

Credit : Markus Poessel/Wikimedia commons; retouched by Pbroks13

Einstein would later talk about the value of the time he spent getting a formal education, and discussing just what it was that made the investment of his time and efforts so valuable. In contrast to Thomas Edison, who derided a college education as useless, Einstein wrote the following in 1921, as recorded by Philipp Frank in his biography of Einstein :

“It is not so very important for a person to learn facts. For that he does not really need a college. He can learn them from books. The value of an education in a liberal arts college is not the learning of many facts, but the training of the mind to think something that cannot be learned from textbooks.”

This powerful statement, on its own, is sufficient to illustrate that Einstein would have chosen not to use AI/LLMs, even if they were available to him, to do anything other than look up facts (which he would then, of course, have to verify weren’t AI hallucinations). He absolutely would not have outsourced his critical thinking skills — or any endeavors that helped develop them — to a tool that would have taken those valuable efforts away from him. After all, thinking and struggling with the material, and mutual sharing in the expertise of a group of colleagues, is how Einstein grew his mind and attained his great achievements in the first place.

It really is true: Einstein did actually say “Imagination is more important than knowledge.” But unless you understand the context in which this statement was made, you’re almost certainly misinterpreting and misunderstanding what was meant.

Credit : Nobel Prize foundation

Despite the fact that Einstein indeed did recognize the unparalleled value of a formal education, that’s something that normally gets forgotten by history. When we quote Einstein, his most famous quote tends to get summarized as, “imagination is more important than knowledge,” but taking it at face value — suggesting that imagination is what’s important and that knowledge is unimportant — completely misunderstands the full meaning and context of what Einstein was saying. That quote arose in the context of a 1929 interview with George Sylvester Viereck , where Einstein was discussing the validation of general relativity by the 1919 eclipse expedition.

Einstein : “I believe in intuitions and inspirations. I sometimes feel that I am right. I do not know that I am. When two expeditions of scientists, financed by the Royal Academy, went forth to test my theory of relativity, I was convinced that their conclusions would tally with my hypothesis. I was not surprised when the eclipse of May 29, 1919, confirmed my intuitions. I would have been surprised if I had been wrong.”

Viereck : “Then you trust more to your imagination than to your knowledge?”

Einstein : “I am enough of the artist to draw freely upon my imagination. Imagination is more important than knowledge. Knowledge is limited. Imagination encircles the world.”

In context, it is simply Einstein expressing his conviction that his ideas were correct before they were validated: no different than any imaginative thinker believing that their novel ideas can take us beyond the current frontiers of our existing understanding. This is a requirement for anyone seeking to advance what we know, but it is not sufficient on its own.

The results of Arthur Eddington’s 1919 expedition, which confirmed and validated the predictions of Einstein’s general relativity, while disagreeing significantly with the alternative (Newtonian) predictions, was the first observational confirmation of Einstein’s new theory of gravity. The amount that starlight was deflected by during a total solar eclipse was a key prediction that was unique to Einstein’s new theory.

Credit : London Illustrated News, 1919

Without the knowledge to underpin the foundation for his ideas — which, in turn, grew out of his imagination — and without the requisite knowledge and skills to formally and fully develop those ideas, Einstein’s novel predictions never could have come to be. The expectation that he would be correct is not the revolutionary part of Einstein’s quote about imagination and knowledge, but rather that his deep theoretical understanding meant that any other outcome would have greatly surprised him.

And yet, because he was a physicist at heart, and physics itself is a science that’s rooted in experiment and observation, he remained open to being surprised in exactly that regard.

This is something that no AI or LLM can replace. No matter how confidently an LLM tells you what the outcome of an analysis or study is going to be, there is absolutely no way to ever replace what can be learned by conducting the study, particularly by collecting the relevant data and letting a fully correct and rigorous analysis be what determines the conclusion. When LLMs tell you what the outcome will be for an experiment or set of observations that has not yet been conducted, you must absolutely not believe it. It is, at best, an educated guess that has dressed itself up as a confident conclusion: an impostor posing as something it can never be a substitute for.

This up-to-date “Eagle plot” of the black holes and neutron stars detected through gravitational wave mergers (orange and blue) and through electromagnetic signals (yellow and red) show the present status of known black holes and neutron stars under 250 solar masses. Just 11 years ago, there were no known gravitational wave events, but today, in June of 2026, there are 390 confirmed events, representing 780 pre-merger and 390 post-merger objects. No increase in computational capabilities could ever substitute for this actual data.

Credit : LIGO-Virgo-KAGRA / Aaron Geller / Northwestern

This isn’t something physicists and astronomers tell themselves to continue to justify their existence in the era of widespread AI and LLMs, but rather a conclusion reached by a consortium of world-leading computer science and AI researchers. As the leader of the Stanford Trustworthy AI Research (STAIR) lab, Dr. Sanmi Koyejo addressed the American Astronomical Society at their 248th meeting in June of 2026 , and presented an important distinction: doing well on a test — or in the case of LLMs, what we call “performing well on a benchmark” — is not the same as doing good science, or as doing science at all.

In order for AI to work in any sense at all, that’s precisely what it needs: benchmarks, and in particular, benchmarks for validation. Is there some kind of test or check you can perform for your data? If so, then you might be able to leverage AI in some useful way, as then your AI can run quickly and often on your data, even as you acquire it. However, these benchmarks assume three things.

Cheap verification: that checking whether your benchmark is met (or a threshold is crossed) is fast and nearly free.

Short feedback loops: that you learn whether “you were right” or “you were not right” immediately.

And that the questions are fixed: the set of questions and the benchmark tests are given up front, and there are no revisions (or moving targets) to those benchmarks.

The big limitation of AI for science is this: none of these assumptions are true when it comes to physics and astronomy.

This result shows the spectra of four different galaxies from JWST found in the same region of sky. The top galaxy, also known as Callum’s galaxy, was originally thought to be at a redshift of 16, which would have made it the most distant ever. Additional, spectroscopic data (shown here) was required to determine its (much closer) distance. Although these four galaxies are all of a modest redshift (z=4.9, or an age of the Universe of 1.2 billion years), they’re all co-located, which may indicate that they’re part of the same bound structure: a young galaxy group or cluster, which could make it the earliest known such object if confirmed.

Credit : P. Arrabal Haro et al./CEERS collaboration, Nature submitted, 2023

In real science, things are a lot messier. There are often mislabeled items and mismeasured points: for example, when two galaxies at two different distances overlap along the same line-of-sight. There are often several plausible correct answers, not one unique solution, such as when fitting the photometric redshift of a distant object. Sometimes, there are structural ways that a benchmark breaks: when a model is marked wrong (because it crosses or doesn’t cross some threshold) despite being correct.

As Koyejo elaborated, a benchmark for AI is like a measuring instrument, but with noise. Some percentage of the signal that it picks up will be usable signal, while the rest will be noise and artifacts. But as you go to stronger AI models and more frontier models, the instrument that is AI actually loses resolution, and is more likely to provide noise (and not usable signal) exactly when you need to only extract the signal the most.

To put it bluntly, of the three assumptions you need for AI benchmarks to be useful, a physical science like physics or astronomy actually breaks all three of them. When you need more (or better) data — as you often do in science — having extra computing power does not help. This leads to an extremely grave danger: that AI models will reach agreement, and converge on what they can conclude based on all current inputs, without actually approaching the truth of factual reality. AI consensus can measure agreement, but cannot tell you anything about correctness.

The galaxy JADES-GS-z14-0, alongside an interloping, much closer unrelated galaxy that happens to be along the exact same line-of-sight. Its extended nature is clearly visible even in this blown-up NIRCam image of the two galaxies. This object would not have been identified correctly by any AI algorithm, as it was not trained on either galaxies this distant or on two faint, blurry galaxies that occupied the same line-of-sight at different distances. For astronomy and astrophysics, humans are indispensable.

Credit : NASA, ESA, CSA, STScI, B. Robertson (UC Santa Cruz), B. Johnson (CfA), S. Tacchella (Cambridge), P. Cargile (CfA); Annotations: E. Siegel

AI models cannot evaluate; you need an expert capable of critical thinking to do that. AI models cannot identify unique events that you didn’t pre-specify that it should look for; it will overlook them. AI models cannot engage in critical thinking the way an expert-level human can; they can only reach agreement by comparing to what they were trained on. They severely underperform, compared to humans, at reviewing and evaluating the work of others, and moreover can be gamed: an AI’s evaluation score can be increased significantly simply by rewriting the veneer (surface) of a submission, without changing the science underneath it. In Koyejo’s exact words:

“When the thing being measured and the thing doing the measuring share the same blind spots, the errors do not cancel. They reinforce.”

To be sure, humans are fallible as well. We have gaps in our knowledge. We have blind spots. We have biases. But we have tools and abilities that AI does not. We have critical thinking skills. We can think beyond our training data. We have a culture of blind analysis, of rigor, and of reproducibility. And, perhaps most importantly, we have a collaborative community that works together to strengthen each other, and to strengthen the field as a whole. Einstein relied on it, benefitted from it, and contributed to growing it for the next generation.

Every time you eschew critical thinking for yourself and outsource it to AI, you fundamentally cheat yourself out of any intellectual challenge, and hence, any opportunity to actually grow your mind. People like Einstein actively sought such challenges out, and that enabled him to flourish, transforming him from the below-average student that Minkowski knew to one of the greatest geniuses in all of recorded history. Einstein would have rejected the use of AI, and you should too. As Nicholas Carr summed it up , “Armed with generative AI, a B student can produce A work while turning into a C student.” Once the class or the evaluation is over, your mind will be the only thing of value you’re left with. Don’t cheat yourself out of the opportunity to grow the most capable one you can.

This article The big reason Einstein would never have used AI is featured on Big Think .