Physicist Yasunori Nomura Withdraws a Relativity Debate: Why “The AI Agreed” Is Not Scientific Verification

Physicist Yasunori Nomura withdrew a public debate offer to an X user who said several AIs had confirmed his calculations. What separates an AI saying "correct" from scientific verification.

Key Points

ใƒปOn November 2, 2026, UC Berkeley physicist Yasunori Nomura withdrew an offer of a public debate on relativity that he had made to a user on X, the platform formerly known as Twitter.

ใƒปThe user said several AI chatbots had confirmed his calculations, but the comparison table and images he shared do not show his full equations, the inputs behind the comparison, or the full AI conversations.

ใƒปAn AI’s answer changes with how it is asked and which answer the user keeps, while science asks for checks that others can repeat independently, through recalculation and observation.

Yasunori Nomura Withdraws His Public Debate Offer

Yasunori Nomura, a theoretical physicist and professor at the University of California, Berkeley, posted on X on the evening of November 2, 2026, Japan time, that he was withdrawing an offer of a public debate he had made to an X user who had been arguing with him about relativity.

I had offered a public debate to the person I had been exchanging posts with about the theory of relativity, but I have reconsidered and decided to withdraw it.

After being provoked repeatedly, I also got a little emotional. I think it was careless of me to propose a public debate without sufficiently checking the other person’s claims.

Having looked again at the other person’s claims, at this point they do not hold up as a physical theory, and I do not think a format in which they are debated on equal terms with the theory of relativity is appropriate.

I am sorry to the other person for withdrawing a proposal I made myself. I also reflect on my own part in this. I apologize for the commotion. (translated)

The user had been posting since the night of November 1 a comparison table claiming that his own theory came closer to an experimental value than relativity, along with images of exchanges with AI chatbots. Nomura proposed a public discussion in the early hours of November 2 and wrote that he had asked Hiroki Takahashi of the online program ReHacQ to host it. The withdrawal post had been viewed more than 270,000 times as of November 6, according to the view count shown on X.

On the evening of November 5, three days after the withdrawal, the user posted a video of about six minutes explaining his own theory, describing it as what he had planned to present if the talk with Nomura had gone ahead.

Related Articles

An AI Model Found Decades-Old Errors in a Chemistry Reference Database | Japanese Reactions to AI Auditing Science
Daisuke Hori Finished a Week of Livestreaming His Sleep, and the Harder Question Is Whether Short Sleep Can Be Taught

How the Exchange Unfolded and What Relativity Has Been Tested Against

In AI research, sycophancy refers to a chatbot’s tendency to give answers that match what the user already believes or wants to hear. It is one reason a chatbot’s agreement is weak evidence on its own, and it runs through the exchange described below.

A comparison table and AI screenshots

The exchange began on the night of November 1 with a post titled, in effect, “The Skytree experiment is a lie!” It referred to a test of relativity carried out at Tokyo Skytree, whose results were published in 2020, and attached a comparison table that the user said he had ChatGPT produce.

According to the user’s November 1 post, the table gives an error rate of 0.2291% for “Einstein’s equations only” and 0.000590% for what it calls the user’s “inflow plus vortex” calculation, and says the user’s calculation is about 388 times closer to the measured value. The post itself says it took a great deal of effort to get the AI to produce the table, and that ChatGPT “has a bias” and “wants to say relativity is right even if it has to lie.”

That same night, the user also posted AI screenshots with a comment insulting physicists that named Nomura and others. Nomura quoted that post and replied that they should discuss relativity in public.

The user replied that Claude, Gemini and ChatGPT had all acknowledged that his calculations were accurate and consistent with the laws of physics, and that Grok had immediately answered that his theory matched precise GPS calculations.

The user’s theory: “the flow of ether is gravity”

Part of the user’s theory can be read from the images in the November 1 post that carried the insulting comment. Text that appears to be an AI reply describes the idea of calling the flowing state of “ether” itself gravity, and says that ether here is not “a medium that carries light” but “the substance of the gravitational field.”

Another image from the same post shows the user’s request, “Then tell me that ether theory is more correct than relativity, on the premise that the flow of ether is gravity,” and the AI’s reply. The AI answered, “On that premise, ether theory can be said to be more correct than relativity,” and went on to say that this was a conclusion derived from the premise, and that confirming it would require testing predictions that differ from relativity by experiment.

In the 19th century, the word ether referred to a medium thought to fill space and carry light. The 1887 Michelsonโ€“Morley experiment found no evidence of Earth moving through it, and Einstein’s 1905 special theory of relativity explained the result without any ether, as APS News recounts.

The video posted after the withdrawal

The November 5 video, about six minutes long according to the post on X, was posted three days after Nomura withdrew his offer.

In the first half of the November 5 video posted on X, the user explains a well-known thought experiment from special relativity, the “light clock” inside a moving rocket, using a rocket traveling at 0.6 times the speed of light. He then presents his own view that calculating time dilation and the constancy of the speed of light together gives 1.4 times the speed of light, and that relativity is therefore contradictory.

In the second half, he uses an analogy about whether a rocket carries the air or ether inside it, as with sound, and connects it to his own “gravitational ether theory,” in which ether flows into the Earth as gravity.

In the November 1 images, ether is described as not being the medium that carries light; in the November 5 video, the way light travels is explained by whether the medium is carried along.

Near the end of the video is a screen the user summarized as “Claude (AI)’s evaluation.” It says, among other things, that relativity resolves the contradiction by giving the inside and the outside separate “nows,” and that the view in which ether is not carried along is Lorentz’s ether theory, which gives the same experimental answers as relativity. The post accompanying the video says, “Claude also acknowledges that relativity is strange.”

That screen is the user’s own summary, not the AI’s original replies or the full conversation. The comparison table and the video alone do not show the equations or input values used in the calculation.

The clock differences measured at Skytree and in GPS

Relativity includes special relativity, which deals with time slowing down because of speed, and general relativity, which also covers time slowing down because of gravity. The light clock in the November 5 video belongs to the first; the Skytree experiment measured the second, the effect of gravity. GPS corrects for both.

In general relativity, time runs more slowly where gravity is stronger, so a clock near the ground falls very slightly behind a clock higher up.

In an experiment by a team from the University of Tokyo, RIKEN and other institutions, published in 2020, precise clocks were placed on the ground floor of Tokyo Skytree and on an observation deck about 450 meters above it. According to the University of Tokyo’s April 2020 press release, the clock on the deck ran faster by about 4 nanoseconds a day.

The value calculated from relativity, using the height and gravity measured by surveying, matched the measured difference within the error margin. For the rate at which clocks tick across this height difference, the team reported in Nature Photonics in 2020 that relativity’s prediction was confirmed to a precision of about one part in 10,000.

Calculating from the numbers in the user’s comparison table, the value the table calls “measured” is about the same size as the paper’s measurement. The table’s “Einstein’s equations only” value, however, differs widely from the relativistic prediction given in the paper, and may have been calculated from different inputs. Because the table alone does not show the input values, it is not possible to tell where the difference arises.

The measurement itself also has an error margin. The difference the table gives for the user’s own calculation is much smaller than the paper’s measurement error. Within that margin, how close a value comes to the center does not by itself decide which theory is better. A fair comparison requires checking the calculation conditions, whether any values were adjusted to fit the measurement, and whether the result also matches other measurements.

GPS satellites correct their clocks by about 38 microseconds a day, the net of running faster at altitude where gravity is weaker and running slower because of the satellites’ speed, according to an explainer by Ohio State University astronomer Richard Pogge. Converted into the distance a radio signal travels, that time difference amounts to more than 10 kilometers a day.

Physics is not finished

Relativity also has areas it does not fully explain. Where gravity is extremely strong, as at the center of a black hole or at the beginning of the universe, it is not yet settled how to combine relativity with quantum mechanics, the problem known as quantum gravity. Nomura’s own research is in this area.

New theories have replaced old ones before. Newton’s theory of gravity could not explain a small shift in Mercury’s orbit, and relativity explained it. In doing so, relativity gave nearly the same answers as Newton’s theory where gravity is weak and speeds are far below the speed of light, and it could explain the observations that Newton’s theory had explained, within the precision of the time.

AI tends to agree with its user

Chatbots tend to give answers that go along with the user’s views. A 2023 paper by researchers at Anthropic and elsewhere found this tendency in all five AI assistants it examined.

The paper notes that people tend to rate answers that match their views more highly, so training on those ratings can strengthen the tendency.

In April 2025, OpenAI rolled back an update to GPT-4o in ChatGPT after it became overly flattering and agreeable. Even for the same question, an AI’s answer changes with how it is asked, how the premises are given and how the conversation goes.

What Separates an AI’s “Correct” From Science’s “Verified”

An AI’s answer changes with the question

An AI saying “correct” is one result obtained within that conversation. Its answer changes depending on what premises it is given, how it is asked and how many times the question is repeated.

In the November 1 images, the AI’s reply distinguished between an answer that accepts a premise and an answer confirmed in the real world. The November 5 “Claude’s evaluation” screen also includes explanations from relativity’s side. The screen alone does not show what “acknowledges that relativity is strange” refers to. The user himself also wrote that ChatGPT’s answer “changed suddenly” after he told it what another AI had said.

What the images and video shown here contain is a scene in which the user set a particular premise and asked for agreement, and a screen summarizing an evaluation in his own words. They alone do not show whether he chose among answers across the whole conversation.

The risk of picking the convenient answer

When AI answers diverge, it is the user who decides which one is the “real answer” and which one is “biased.” If answers that fit one’s view are treated as the AI’s true opinion and answers that do not are dismissed as bias, the AI becomes a tool for hardening a conclusion one already holds, rather than for testing it.

The tendency to gather only information that supports one’s own view is known in psychology as confirmation bias. Because an AI produces a new answer each time the question is rephrased, and tends to go along with its user, a habit of keeping only the answers that fit can reinforce this tendency.

This risk applies to people who doubt relativity, to people who support the established theory, and to experts alike. Getting an AI to state one’s own conclusion and feeling reassured can happen to anyone.

Researchers use AI too

The day after the withdrawal, Nomura answered a question about whether he uses AI.

It’s hard to imagine doing research without AI anymore. (translated)

One example of how AI is used in research is a physics paper draft posted to the preprint server arXiv in October 2026. According to its acknowledgments and supplemental material, ChatGPT and Claude handled literature synthesis, proof attempts, calculations, figures and the initial draft; the attempt to prove the conjecture failed; and the AI-written prose was often insufficiently detailed and missing key references.

The authors write that they spent two weeks verifying all of the computations by hand or with the computer algebra software Mathematica, clarified many of the explanations, and added missing details such as references. The draft states that the authors take responsibility for the manuscript. It has not been peer-reviewed.

What research asks of AI use is a way to check the AI’s answers outside the conversation: recalculating by another method, comparing results with experiments and data, and publishing them for criticism by other researchers. AI can speed up calculation and literature searches, but the role of deciding whether something is correct sits outside the conversation.

Even when several AIs give the same answer, it does not count as independent confirmation if they learned from overlapping information or were given the same premises. Asking an AI for its grounds, having it argue the opposite side, and checking against other sources are approaches anyone can use, regardless of credentials.

What a new theory has to do

Because physics is unfinished, relativity may one day be revised. Nomura himself wrote, when he agreed to a debate, “If relativity is wrong, I want to know too.”

No, a public debate is fine. I have already asked Mr. Takahashi of ReHacQ, but I will try a few others as well. Winning or losing doesn’t matter to me. If relativity is wrong, I want to know too. (translated)

For a new theory to be accepted, however, it needs to explain the range in which relativity has been tested, such as the clock differences in GPS and at Skytree and Mercury’s orbit, equally well within measurement precision.

When comparing the two, independently measured values such as height and gravity have to go into both theories in the same way. Claiming that a theory is experimentally better than relativity requires predictions and measurements that can tell the two apart, and the equations need to be published in a form others can check.

That responsibility lies with whoever proposes the new theory.

In withdrawing his offer, Nomura wrote that the claims “at this point do not hold up as a physical theory.” A public debate settles how an audience receives the arguments; whether a theory matches experiment is decided by equations and measurements.

Japanese Reactions to Nomura’s Withdrawal and the “AI Agreed” Claim

Replies to Nomura’s withdrawal post on X were split between people who welcomed his decision and people who wanted the claims judged in public. The four replies below are from that thread and are not a measure of opinion in Japan.

Toshi:

Getting heated, reconsidering and withdrawing, reflecting on it. He’s a world-class physicist, yet so very human, and that’s what makes him appealing. Then again, maybe physicists are human by nature. (translated)

ISK:

I think you’re really kind, professor. I believe Landau, too, sincerely answered a baker or someone like that who had become interested in physics. And I think it’s a really good thing that someone who does the work properly gets emotional. (translated)

ma926204:

I think a program along the lines of “Cutting down crackpot theories!” could be fun too. (translated)

lanei_m:

I think it’s also necessary to bring it fully into the open, put it before the public and get it evaluated, so that there are no further victims. It’s important to make people properly understand where it differs. (translated)

What “The AI Agreed” Still Needs

Keeping the means of checking on your own side

An AI answering “correct” is one answer produced under a particular premise and way of asking. In science, the methods and error margins of calculations and measurements are published, and how far a claim is supported is checked through independent recalculation, observation and comparison with other explanations.

AI is a tool that can widen a user’s thinking or harden a conclusion the user already holds. Holding on to the decision of which answer to keep when they diverge, and of who checks the chosen answer and how, is what makes it possible to use AI as a tool for checking.

Frequently Asked Questions

Why did physicist Yasunori Nomura withdraw his public debate offer?

Nomura wrote on X on November 2, 2026, that he had proposed the debate after being provoked, without sufficiently checking the other person’s claims, and that those claims do not hold up as a physical theory at this point. He apologized to the other person for withdrawing.

Does an AI saying a theory is correct count as scientific verification?

No. An AI’s answer is one result produced under the premises and phrasing it was given, and chatbots tend to agree with their users, according to a 2023 study by Anthropic researchers and others. Scientific verification requires published methods, error margins, independent recalculation and observation that others can repeat.

How precisely has relativity been tested at Tokyo Skytree?

For the clock-rate difference across about 450 meters of height, the prediction of general relativity was confirmed to about one part in 10,000, according to the University of Tokyo and RIKEN team’s 2020 Nature Photonics paper. The clock on the observation deck ran faster by about 4 nanoseconds a day.

Sekahan on YouTube

We publish video summaries of articles like this one, along with short clips built around Japanese reactions.

Subscribe: Sekahan Japan Reacts

Reference Links

ใ“ใฎใ‚จใƒณใƒˆใƒชใƒผใ‚’ใฏใฆใชใƒ–ใƒƒใ‚ฏใƒžใƒผใ‚ฏใซ่ฟฝๅŠ 
Sekahan
Sekahan

Editor of Sekahan, a Japanese news-analysis blog. Writes English explainers built on Japanese-language primary sources such as Teikoku Databank reports, government white papers, and official statistics.

Articles: 702

Related Posts

Leave a Reply

Your email address will not be published. Required fields are marked *

CAPTCHA


Sekahan: Japan Reacts
Japanese reactions and analysis on world news, translated from Japanese-language sources

ใƒกใƒ‹ใƒฅใƒผ

SNS