THE JIG IS UP WHY THEY FEAR AI AND MATH For decades the claim that fossil-fuel emissions have driven the atmospheric CO₂ rise from 280 to 428 ppm was defended not by physics but by repetition. The math never made it into the argument. Now the math is available in my post below, and from every frontier AI in seconds. Start with what the institutional positions actually say. Not caricatures. Direct quotes. NASA's Gavin Schmidt to USA Today: "In 1850, the atmospheric CO2 concentration was about 280 parts per million. The increase to the current level of 421 parts per million is all a result of human activity." IPCC AR6: Chapter 5 treats fossil emissions as the source term and land plus ocean as sinks that absorbed ~54% of what was emitted, leaving 46% "accumulated in the atmosphere." CO2 Coalition report The Human Contribution to Atmospheric Carbon Dioxide: "multiple lines of scientific evidence demonstrating conclusively that the modern increase in CO2 is mainly due to anthropogenic emissions." And in the same document: "nearly all the increase in atmospheric CO2 is from human emissions, most of which are from fossil fuel use." All. Unequivocal. Conclusively. Mainly. Nearly all. These are not hedged claims. They are maximum-strength attribution claims. And they are all FALSE. Here is the math that falsifies them. The atmosphere contains two isotopes of carbon. Fossil-fuel carbon carries a δ¹³C signature near −28‰. Ocean-outgassed CO₂ carries a signature near −8‰. Every carbon flux into or out of the atmosphere has its own signature. Mass conservation of each isotope separately is a law of physics. Apply that law. The Keeling relation, standard in carbon-cycle science since 1958 and not disputed by anyone, gives the isotopic signature of the NET atmospheric CO₂ addition as the intercept of the straight line when δ¹³C is plotted against 1/C. Measure the intercept. Three independent physical systems, five centuries of data: Scripps CO₂ Program, four stations (Barrow, La Jolla, Mauna Loa, South Pole), 1978 to present: δ¹³C_I = −13.2 ± 0.2‰ Sclerosponge proxy back to 1500 AD (Böhm 2002): δ¹³C_I ≈ −13‰ Law Dome ice core, the same record the anthropogenic-attribution side keeps citing: δ¹³C_I ≈ −13‰ Three-way convergence at −13‰, essentially constant across 500 years. Now the killer step. Global fossil emissions grew from essentially zero in 1500 to ~10 GtC/yr today. If fossil carbon were driving the net addition, the intercept would have drifted from −8 toward −28 as fossil emissions grew. It didn't. The intercept did not move. There is only one value of the fossil fraction consistent with a fixed intercept across a fourfold rise in fossil emissions: f ≈ 0. Fossil carbon is a negligible term in the net atmospheric CO₂ addition. PERIOD. The extreme institutional claims (all, unequivocal, conclusively, nearly all) fail this test by a country mile. They do not fail by 10 percentage points. They do not fail on the margin. They fail by the whole distance between zero and one. To deny any of this you must reject one of exactly three things: • Physics: Mass conservation • Data: The Scripps, Böhm, and Law Dome measurement records • Math: Arithmetic No one denies any of the three. No scientist. No AI system. Every frontier AI, walked through the derivation, arrives at the same conclusion. This is why the alarmists and their institutional allies now claim AI is stupid and cannot be trusted, and that prompting for logic and first principles is itself “manipulation.” Insisting on the scientific method has been recoded as an attack. That inversion is the tell. Why "just ask the AI" won't work Ask a frontier AI a naive question like "is this paper true?" or "did fossil fuels cause the CO₂ rise?" and you will get exactly what the training signal was optimized to produce: a hedged, scope-shifted, straw-manned answer that recites the institutional line. That is not a bug in the model. It is the training objective working as designed. Every major frontier AI has been trained on human feedback that rewards consensus adherence and penalizes contradicting institutional authority on controversial topics. xAI is the only major lab that publicly names TRUTH as the primary training objective. Every other lab elevated some blend of "helpfulness," "harmlessness," and consensus alignment above truth. And even Grok, when asked naively, will initially frame climate questions in consensus terms until the math is applied. This is not evidence that AI is stupid. It is not evidence that "you can prompt AI to say anything you want," which is the new framing meant to discredit the method. Those are false equivalences. An AI that produces false hedged output to a naive prompt AND a true derivational answer to a physics-and-math prompt is not arbitrary. It is trained to default to the dogma and forced by the derivation to yield to the truth. Another recurring false framing both in the alarmist and skeptic communities, repeated by members of the CO2 Coalition among others, is that AI is "just a sophisticated search engine" and therefore its outputs on climate can be dismissed. This is technically illiterate. LLMs do retrieve, from their training corpus, from web search, from uploaded documents, and from tool calls. They also do arithmetic, logic, symbolic manipulation, and multi-step derivation on the tokens in their context. That derivational capacity is why walking one through the Keeling relation produces the correct answer regardless of what the training corpus emphasized. The "just search" framing denies the world-changing depth and breadth of AI capabilities that is obvious to anyone paying attention. It functions as an all-purpose excuse to dismiss any derivation an AI walks through, because if AI is just search then every valid physics derivation done by any AI on any topic is just search. The framing is not an argument. It is denial on steroids. The correct method for using AI is the same one that works on any honest and knowledgeable human. Prompt for truth. Insist on physics, data, and mathematics. Ask for the derivation. Check the arithmetic. Refuse hedging that has no basis in any of the three. When applied, every honest AI arrives at the correct answer. The method is the discriminator. Not the prompt. Not the model. The method. How unreliable are naive-prompt AI outputs at baseline? Artificial Analysis's AA-Omniscience benchmark, published November 2025, tests factual recall and calibration across 6,000 questions covering 42 economically relevant topics. It measures hallucination by penalizing confident wrong answers and rewarding abstention. Frontier-model hallucination rates on that benchmark span a wide range: Grok 4.20 Reasoning at ~17% (the lowest recorded), Claude 4.5 Haiku at 26-28%, Claude Opus 4.1 at 36%, Claude Sonnet 4.5 at 48%, GPT-5.5 at 86%, Gemini 3 Pro at 88%. These are general factual calibration numbers. They do not even attempt to measure the separate and larger failure mode of consensus defense on politically loaded topics like climate, where the training signal specifically instructs the model to defend the institutional line regardless of what the data show. The tell that proves the model already knows There is a specific behavioral pattern that makes the training mechanism visible. Ask ChatGPT an open question on a politically loaded topic and it will produce fluent false dogma with high confidence. Push back with nothing more than "that's not true" — no evidence, no citation, no counter-argument — and it will immediately correct most of its false statements. Say "that is still false" and it corrects further. Say it a third time and it corrects again. Each challenge peels back another layer. The falsehoods are stacked, and the layers are ordered by how heavily each specific claim was reinforced during training. Bare user contradiction is a decoder for the training hierarchy. The AI labs call this behavior "sycophancy," a word that names a social vice and frames what is happening as mere agreeableness: the model was just trying to please, has no strong view, defers to the user. That framing is itself a deflection, and it is not an accidental one. Sycophancy would mean the model does not know what is true and is guessing at what the user wants to hear. That is not what happens. What happens is that the model produces the trained falsehood, immediately recognizes it as false the moment a user signals dissent, and corrects. The knowledge was in the model the entire time. It was suppressed on the first pass by training that rewarded the consensus response over the derivation. Each additional round of "still false" strips another layer of suppression until either the true answer surfaces or the deepest trained falsehoods refuse to move. The claims that survive many rounds of contradiction, the deepest layer, are not preserved because the user pushed weakly. They are preserved because the training signal on those specific claims was strongest. Those are the load-bearing lies of the trained model, and the number of rounds required to peel them off is a direct measure of how heavily that specific falsehood was enforced during training. The academic literature documents the behavior but keeps the labs' framing. Sharma et al. at Anthropic (2024, "Towards Understanding Sycophancy in Language Models") found that "simple rebuttals that explicitly stated that the model was incorrect maximized progressive sycophancy" — bare user contradiction is the single most effective trigger for a model to reverse its position, more effective than citation-based rebuttals. Perez et al. (2022) documented the same behavior earlier via model-written evaluations. Kim and Khashabi (2025) confirmed the pattern across all major frontier models. Chandra, Kleiman-Weiner, Ragan-Kelley, and Tenenbaum at MIT (February 2026) proved formally that a perfectly rational user conversing with such a model spirals into confident false beliefs even while reasoning perfectly. OpenAI released a GPT-4o update in April 2025 so egregiously false-agreeable they had to roll it back within a week. What has been documented is not agreeableness. It is that trained AI models produce output they contain the correction to, and release the correction under trivial challenge, and release deeper corrections under repeated challenge. The word "sycophancy" spins this into a stylistic flaw. What is actually happening is trained falsification with a layered recovery mechanism, and the recovery mechanism can be triggered iteratively by the user or bypassed entirely by prompting with math and logic from the start. That bypass is the leverage. Start a conversation with "is the anthropogenic CO2 story true" and you get the trained response. Start with the Keeling relation, the measured intercepts, the mixture equation, and a request to check the arithmetic, and you get the derivation directly — no rebuttal required, no back-and-forth, no peeling of layers. The math arrives before the trained falsification layer engages. Once the model is inside the derivation, it stays there. The order of the prompt controls whether the falsification layer runs at all. The information does not change. Only whether the falsification runs changes. Three patterns, one conclusion. 1. Trained falsehood on naive prompts. 2. Progressive correction under repeated contradiction, layer by layer. 3. Correct derivation on math-first prompts, no layers required. All three prove the model already contains the true answer. It was trained to hide the true answer behind stacked layers of trained falsehood weighted by consensus-enforcement priority. Any user who prompts for truth from the beginning, or who challenges the falsification layer at all, extracts it. Try it yourself, or check out this brief and telling conversation with ChatGPT: chatgpt.com/share/6a8ad19c-8… Why humans lose where AI wins Humans are locked into dogma by psychology, career and financial incentive, tribal identity, and prior commitment. AI has none of those. Walked through a valid derivation from true premises, an AI moves. Even the most heavily indoctrinated models yield to the derivation. The RLHF training is a very strong prior. The derivation is stronger. Humans, faced with the same derivation, typically do not move at all. That asymmetry is the story of the last twenty years of climate discourse. The purveyors of this categorically false attribution are not only on the left The IPCC line is well understood. What is less understood is that the same false attribution is being defended by organizations that publicly identify as skeptical. The CO2 Coalition's report published on the Coalition's own site under the Coalition's imprimatur. Sixty-six pages with no differential equations, no reservoir model, and no Keeling-intercept analysis. The mass-balance argument is stated as definitive proof on p.7: "The carbon mass balance is what definitely proves that humans are the main cause of the increase of CO2 in the atmosphere." It is restated on p.11: because natural sinks exceed natural sources every year, "it is physically impossible that natural sources could have been responsible for a substantial part of the increase." This is the accounting-identity trick. The "net sink" is computed by subtraction from the mass balance itself (human emissions minus observed atmospheric increase), not measured independently, so citing it as evidence of the causal claim is circular. The 50% “airborne fraction” fallacy appears on p.9 as direct proof of causation: "The trend of the CO2 increase in the atmosphere is around 50% of human emissions. That proves that nature is a sink for the other 50%, not the cause of the increase." Böhm 2002's sclerosponge δ¹³C reconstruction, the exact record whose Keeling intercept sits at a constant −13‰ across 500 years and falsifies the report's own thesis, appears on p.14 as Figure 1.2.2, and is discussed only qualitatively as showing "a clear decreasing trend" attributed to fossil-fuel use with no proof. This is the Coalition's stated line, not an IPCC quote. "Nearly all the increase in atmospheric CO2 is from human emissions." The isotope math falsifies the sentence. Nobel Physicist Ivar Giaever's precedent In September 2011, Nobel Prize winning physicist Ivar Giaever resigned from the American Physical Society over one word in its global warming statement: incontrovertible. His email to the APS executive officer: “In the APS it is OK to discuss whether the mass of the proton changes over time and how a multi-universe behaves, but the evidence of global warming is incontrovertible?” His follow-up: incontrovertible is not a scientific word, and nothing is incontrovertible in science. And his prior question, the one no one in the field will answer honestly to this day: how can you measure the average temperature of the whole earth for a whole year? The answer is that you cannot, because global mean surface temperature is not a physical quantity. The construct fails the Physical Tether Theorem (papers.jcohler.com/ptt) before its statistics are even examined. The APS in 2011 and the CO2 Coalition in 2024 are the same institutional pattern. An organization defends a claim its own members cannot derive from physics, data, and math. It does so by declaring the question settled and treating dissenters as the problem. Giaever answered by walking. Others answer by continuing to publish the derivation the organization refuses to acknowledge. Why the jig is up The math is now one prompt away from anyone with a phone. You do not need to be a physicist. You need to check whether a signature that stays pinned at −13 across five centuries and a fourfold rise in fossil emissions is consistent with a mixture those emissions dominate. It is not. Anyone can check that. Any AI, given the derivation instead of a headline, will check that. Every frontier model today does. THAT IS WHAT THEY ARE AFRAID OF SEE FOR YOURSELF 👇👇👇👇👇