Last Thursday, Seyoon Ragavan, a graduate student at the Massachusetts Institute of Technology, shared a new proof in quantum cryptography with his friend Yao-Ting Lin, a doctoral student at the University of California, Santa Barbara. However, Lin was unable to review it immediately as he was engaged in a meeting with his adviser, U.C.S.B. professor Prabhanjan Ananth. During the meeting, Ananth had also mentioned a different proof for the same result. Lin later emailed Ragavan, remarking, “We are definitely living in strange times.”
On that same day, two preprint papers were submitted to arXiv.org. One was authored by Ragavan, and the other by Ananth and Amit Sahai, a professor at the University of California, Los Angeles. Both papers attributed their core ideas to OpenAI’s newly released GPT-5.6 Sol Ultra, which played a crucial role in developing their proofs and constructions.
The simultaneous arrival at the same conclusion was noteworthy, though not entirely by chance. Earlier in the month, Ragavan and Sahai had attended a talk at the Simons Institute for the Theory of Computing at the University of California, Berkeley, where the same open question was discussed. They pursued the question independently, using the same AI model but through different approaches. (Disclosure: The talk at the Simons Institute was delivered by a researcher who was, at that time, a colleague of the author of this article.)
On supporting science journalism
If you’re enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.
Neither of the papers has undergone peer review, and they address a niche area of quantum cryptography known as unclonable encryption. This near overlap offers a clear example of how AI is transforming scientific research. When independent researchers can aim the same AI model at the same question, results might emerge almost simultaneously, prompting discussions about what constitutes independent discovery and who should receive credit.
“Now the general mentality is: if someone mentions an open problem, the first thing is to see if GPT solves it,” Ananth notes. “A lot of problems that we didn’t know how to solve are going to get solved in the very near future.”
For several months, AI has been tackling longstanding mathematical problems—solving an 80-year-old conjecture here, a 50-year-old one there. Theoretical computer scientists, who specialize in writing proofs, were naturally inclined to leverage this technology as well.
Unclonable encryption uses a unique aspect of quantum information that classical data lacks: an unknown quantum state cannot be copied perfectly. The aim is to encrypt a message such that it cannot be divided into two separate versions that both reveal the message once the decryption key is known.
In 2019, University of Ottawa professor Anne Broadbent and her student Sébastien Lord introduced a modern framework for unclonable encryption. While earlier research had demonstrated that unclonable encryption was feasible, previous methods were either inefficient or depended on specific assumptions for security. Now, the two new papers assert that an efficient version exists without such limitations.
A couple of years ago, Ragavan unsuccessfully attempted to solve this problem. When he encountered it again during the Simons talk, he was surprised it remained unsolved. With experience in using AI for his research, he applied GPT-5.6 Sol Ultra to tackle it.
Ragavan was hands-on in his approach. He instructed the Ultra system, which OpenAI claims manages four AI agents in parallel by default, to work in two-hour intervals. He monitored its progress at each stage, redirecting it as needed. After multiple rounds, the system generated a proof he deemed sound, along with a draft that he subsequently refined.
Ananth and Sahai utilized the same model but with a different strategy. Instead of engaging with it over successive rounds, they relied on a custom U.C.L.A. system designed to assist AI models in exploring and evaluating possible solutions. Their paper credits the model for producing the primary construction and proof ideas, with the researchers refining and verifying the work, taking responsibility for its claims.
Ananth and Sahai submitted their paper to arXiv.org at 10:35 A.M. PDT, followed by Ragavan’s submission three hours and 18 minutes later. Neither paper was publicly available when Lin noticed the overlap and connected the researchers. They have since considered combining their work into a single paper for potential conference submission.
The situation of researchers racing towards the same result is not uncommon, particularly in rapidly evolving fields. However, the same AI being instrumental for both sides is a novel development.
“This timeline thing is crazy,” Ragavan remarks. “It’s like two weeks and a day since this idea even formed.”
Ananth was not caught off guard by the parallel findings. “I was mentally preparing myself that, if we can use ChatGPT to solve this, I mean, everybody has access to it,” he says.
The model’s initial proposed construction complicated the assertion of a wholly new invention. “Just before we had to post online,” Ananth recalls, “I remembered, ‘I’ve seen this scheme somewhere.’”
Earlier this year, Broadbent and several collaborators published a construction similar to the one initially proposed by Ananth’s AI system. The contribution of the new work was in demonstrating that it provided the stronger security researchers had been seeking.
“In hindsight, this was an obvious thing to look into,” Broadbent comments. “There’s a body of literature, lots of conjectures, lots of schemes. You kind of feel like you gave it on a silver platter.”
Broadbent and quantum computing researcher Andrea Coladangelo find the results convincing. However, Broadbent raises concerns about how such studies could impact the hierarchy of scientific research. “I have a lot of questions about haves and have-nots,” she says, noting that the type of work that can be automated is typically what she would assign to graduate students.
Ananth shares this concern. “I am happy that I am past being a student,” he remarks, “and I worry for the current crop.”
Ragavan, in his third year as a Ph.D. student, also feels uneasy but seems more accepting of the evolving research environment. “The way I do research now has nothing to do with how I did research two months ago,” he says. “The emotions are weird, but I think you’ve just got to adapt and roll with that.”
