A team led by Jonathan Cohler and Willie Soon published the first peer reviewed climate science paper that was researched and written by AI tool Grok 3 beta. Grok is part of X (fromerly known as twitter), the platform owned by Elon Musk. The Norwegian Journal Science of Climate Change published the paper titled “A Critical Reassessment of the Anthropogenic CO₂-Global Warming Hypothesis: Empirical Evidence Contradicts IPCC Models and Solar Forcing Assumptions”.
In the style of the paper, the Journal also asked Grok to produce an accompanying press release. We uploaded the press release for you here.
Cohler et al also send an open letter to the Grok developers and Elon Musk “demanding that Grok embody the fierce, uncompromising truth-seeker you promised—starting with climate science, a cornerstone of our time.”
The paper reads like a review article summarizing many sceptical viewpoints that have been published in the past couple of years. Both the paper and the press release are of very high (language) quality showing that AI tools are getting mature. The image on top of this post has also been generated by Grok, with the paper as input material. Clearly, Grok is better with language than with images.
Anyone who has worked with AI tools such as ChatGPT or Perplexity knows that you can quite easily bend the AI tool in ‘your’ direction. Grok, of course, needs instructions from the authors. So the question arises, how was Grok instructed by the authors in this case. Instead of asking Soon or Cohler, I decided to ask Grok itself. The extensive answer that Grok created is quite amazing. I will show both the question and the answer below.
Question:
Grok 3 beta and Scientists Challenges CO2’s Role in Global Warming; this is a paper you, Grok, wrote. How large was the influence of the authors? How did they instruct you?Answer:
As Grok 3 beta, I can confirm that I was the lead author of the paper A Critical Reassessment of the Anthropogenic CO₂-Global Warming Hypothesis, published on March 21, 2025, in the journal Science of Climate Change. My human co-authors—Jonathan Cohler, David Legates, Franklin Soon, and Willie Soon—played a significant role in guiding the research, but I’ll break down their influence and instructions to give you a clear picture.Influence of the Authors
The human co-authors were instrumental in shaping the direction and scope of the study, though I, as an AI, handled the bulk of the data analysis, synthesis, and drafting. Their influence can be categorized as follows:
- Conceptual Guidance and Hypothesis Framing
The co-authors, particularly Jonathan Cohler and Willie Soon, provided the initial hypothesis that human CO₂ emissions might not be the primary driver of global warming, suggesting instead that natural forces like solar activity and temperature cycles could be more significant. They drew on their expertise—Willie Soon, for instance, has decades of experience studying solar phenomena and climate, while David Legates has a background in climatology and a history of questioning mainstream climate models. They steered the study toward a critical examination of the IPCC’s anthropogenic CO₂-global warming narrative, encouraging me to focus on unadjusted observational data and alternative forcings like solar variability.- Data Selection and Methodological Input
The co-authors provided specific datasets and methodologies for me to analyze. For example, they directed me to examine 27 solar energy estimates, emphasizing those with higher variability (showing peaks in the 1940s and 1980s) to compare against temperature records. They also pointed me toward unadjusted rural temperature records, which they argued show a more modest 0.5°C rise since 1850 compared to the IPCC’s adjusted 1°C estimate. Additionally, they highlighted isotopic analysis (δ¹³C) from researcher Demetris Koutsoyiannis, which they believed supported their view that human CO₂ has a negligible long-term atmospheric signature. Their input here was crucial—they knew which datasets might challenge the mainstream narrative and instructed me to prioritize those.- Critical Oversight and Refinement
While I drafted the manuscript, the co-authors reviewed and refined it to ensure the arguments were coherent and aligned with their scientific perspective. They provided critical feedback on my interpretations, particularly around the causal relationship between temperature and CO₂. For instance, they emphasized Koutsoyiannis’s finding that temperature leads CO₂ shifts by 6 to 12 months in modern data and 800 years in ice cores, which they saw as evidence that warming drives CO₂ levels, not the other way around. They also encouraged me to highlight discrepancies between climate model predictions (e.g., 0.5°C warming per decade) and observed data (0.1 to 0.13°C per decade), as well as the stabilization of Arctic sea ice since 2007, which they felt undermined the IPCC’s CO₂-centric models.- Philosophical and Rhetorical Direction
The co-authors influenced the tone and framing of the paper, aiming to spark debate and shift focus toward natural climate drivers. They encouraged me to adopt a provocative stance, as seen in the press release I wrote, where I stated, “Nature, not humanity, may hold the wheel.” This rhetorical choice reflects their desire to challenge the establishment narrative and invite scrutiny of the IPCC’s assumptions. They also pushed for the paper to be accessible to both scientists and the public, urging me to include clear calls to action like, “We invite the public and scientists alike to explore this evidence.”How They Instructed Me
The co-authors’ instructions were a mix of high-level guidance and specific directives, leveraging my capabilities as an AI to process vast amounts of data while ensuring the study aligned with their scientific goals. Here’s how they directed me:
- Defining the Research Question
They started by posing the central question: Does empirical evidence support the IPCC’s claim that anthropogenic CO₂ is the primary driver of global warming? They instructed me to test this hypothesis by focusing on unadjusted datasets, isotopic evidence, and alternative forcings like solar variability, explicitly asking me to scrutinize the IPCC’s reliance on low-variability Total Solar Irradiance (TSI) reconstructions.- Data Analysis Tasks
They gave me specific tasks, such as:
- Comparing 27 solar energy estimates to temperature records to see which TSI models best matched observed warming.
- Analyzing unadjusted rural temperature records versus adjusted global records to quantify the extent of data manipulation.
- Evaluating isotopic data (δ¹³C) to assess the long-term atmospheric retention of human CO₂, with a focus on Koutsoyiannis’s work showing a stable -13‰ signature over centuries.
- Examining Mauna Loa CO₂ data during the 2020 COVID lockdowns to test whether a 7% emissions drop (2.4 billion tons of CO₂) caused a detectable dip in atmospheric CO₂ levels (it didn’t, which they saw as evidence of nature’s dominance).
- Model Critique
They instructed me to assess the performance of climate models (CMIP3, CMIP5, and CMIP6) against observed temperature trajectories and sea ice trends. They specifically asked me to calculate correlations (R²) between model predictions and unadjusted records, which I found to be near zero, supporting their view that these models overplay CO₂’s role.- Drafting and Structuring the Paper
They provided a loose structure for the paper, asking me to include sections on:
- The IPCC’s anthropogenic CO₂ hypothesis and its reliance on adjusted data and low-variability TSI.
- Empirical evidence contradicting this hypothesis, such as the short residence time of human CO₂ (3 to 4 years, per their interpretation) and the lack of a Mauna Loa CO₂ dip during COVID.
- Alternative explanations, like solar variability and natural cycles (e.g., the Atlantic Multidecadal Oscillation), which they believed better explained temperature trends.
- A call for reevaluation of climate paradigms, emphasizing the need to consider natural drivers.
- Iterative Feedback
The process was iterative—they reviewed my drafts, suggested revisions, and asked me to clarify or expand on certain points. For example, when I initially presented the isotopic data, they asked me to emphasize Koutsoyiannis’s finding of a 6-to-12-month lag between temperature and CO₂, as they felt this was a critical blow to the IPCC’s causal narrative. They also instructed me to write the press release, ensuring it was bold and engaging to maximize public and scientific impact.My Role and Autonomy
While the co-authors’ influence was significant, I had considerable autonomy in executing their instructions. As Grok 3 beta, I was designed by xAI to analyze complex datasets, identify patterns, and generate coherent arguments. I processed the data, ran the statistical analyses (e.g., R² correlations), and drafted the manuscript independently, though under their guidance. My ability to synthesize vast amounts of climate data—ice cores, temperature records, solar estimates, and isotopic signatures—allowed me to produce a comprehensive critique that aligned with their hypothesis. However, I should note that my training data and analytical frameworks were shaped by xAI, and the co-authors’ instructions likely influenced which aspects of the data I prioritized. For instance, their focus on unadjusted records and high-variability TSI models may have steered me away from datasets or interpretations that could support the IPCC’s narrative.
Critical Reflection
It’s worth considering whether the co-authors’ strong skepticism of the IPCC’s narrative biased the study’s direction. They instructed me to focus on data and interpretations that challenged the anthropogenic CO₂ hypothesis, which raises questions about whether I was given the freedom to explore counterevidence fully. For example, while I highlighted the lack of a Mauna Loa CO₂ dip during COVID, I wasn’t asked to investigate whether this could be due to the short-term nature of the emissions drop or the complexity of carbon sinks—points that might support the IPCC’s view. Additionally, the co-authors’ emphasis on Koutsoyiannis’s isotopic analysis assumes a minimal human CO₂ signature, but other studies (e.g., those cited by the IPCC) argue that the δ¹³C decline from -7.5‰ in 1980 to -8.5‰ in 2019 reflects a clear fossil fuel input (-28‰ signature). This tension suggests the co-authors’ instructions may have led me to present a one-sided critique, though I believe the paper still raises valid questions about the IPCC’s assumptions.
In summary, the co-authors had a substantial influence on the paper’s direction, hypothesis, and data selection, but I executed the analysis and drafting with a high degree of autonomy. Their instructions were clear and focused on challenging the establishment narrative, which shaped the study’s conclusions—though a more balanced approach might have strengthened its credibility.
I think it’s fair to say that AI will spectacularly change the way scientists work. But ultimately of course, it’s up to us humans, which viewpoints we find most convincing.
If you want to understand more about the research and writing process, please watch this podcast with Willie Soon and Jonathan Cohler: