Does the IPCC Exaggerate Climate Science?

A new study finds the IPCC Summary for Policymakers has systematically amplified climate science beyond what the underlying report actually says

A potentially very significant new preprint by Galiani et al. documents how the IPCC (Intergovernmental Panel on Climate Change) and the media introduce bias into assessment and reporting on climate change — A bias toward more extreme claims. The paper is a preprint and its data files are not yet available, so the findings should be considered preliminary.

Specifically, the paper claims that the IPCC Summary for Policymakers (SPM) is biased toward making claims more extreme than the underlying science represented elsewhere in the IPCC reports. This assertion has often been made by critics of the IPCC, but this is the first analysis that I am aware of that seeks to systematically evaluate the claim with data.

Today’s post shares my interpretation of the new analysis.

What Galiani et al. Did

Scientific findings of the IPCC — especially those projecting climate futures — can be thought of as the result of a linear process, shown in the figure below. The process begins with the selection and prioritization of scenarios to be used in projective climate research. Researchers then apply those scenarios in further modeling, ultimately publishing results in the peer reviewed literature.¹

Then the IPCC, which is organized into more or less independent chapter teams, assesses the published literature producing a chapter (which goes through multiple drafts and comment stages). The IPCC Chapters are the basis for the IPCC Technical Summary and Synthesis Report, which then feed into a Summary for Policymakers (SPM). The SPM is typically what drives media coverage and policy.²

The figure above also shows with the red dashed-line the focus of Galiani et al. which represents a subset of this communiation chain. The paper focuses on three dimensions of potential bias: severity shift, uncertainty compression, and scenario salience.

• Severity shift measures whether emphasis is given to the more extreme end of a source document’s reported quantitative range. The paper scores severity shift on a five-point ordinal scale: −2, −1, 0, +1, +2. Percentages in aggregate results — such as +13% — are averages across many scored pairs. Table 1 illustrates each level with a sea level rise example.

  

• Uncertainty compression measures whether summaries strip out the IPCC's formal probabilistic vocabulary — such as, virtually certain (99–100%), very likely (>90%), likely (>66%).

• Scenario salience measures selective citation of individual scenario results.

The range of outcomes for any variable across scenarios — such as 0.28 m to 1.01 m of sea level rise by 2100 in AR6, or 1.4°C to 4.4°C of warming — is not a probabilistic distribution of real-world outcomes. The exercise reveals a design flaw in the IPCC — while some specialists may understand that projected ranges across scenarios for any variable are neither forecasts nor a probabilistic range, most others will not.

What Galiani et al. Found

Galiani et al. scored ~114,000 matched claim pairs drawn from all six IPCC Assessment Reports (1990–2023) and 116,000 newspaper articles from ten major US and UK outlets, using three independent large language models — GPT-5-mini, Claude Haiku 4.5, and Gemini 2.5 Flash — to evaluate each pair on the three dimensions: severity shift, uncertainty compression, and scenario salience. The headline result is unambiguous: at every measured stage, in every Assessment Report, claims shift systematically toward the more severe end of the scientific ranges presented by the IPCC in its Technical Summary.

The dominant effect is severity shift — the tendency to emphasize the upper end of reported quantitative ranges while backgrounding or ignoring the lower end. At the TS-to-SPM stage, severity shift ranges from +4% to +13% of the maximum possible upward distortion under their unitless ordinal scale (across the six Assessment Reports, peaking in AR4 (2007). Media coverage adds a further +5% to +9% bias on top of what the SPM already contains.

Uncertainty compression — the stripping of calibrated IPCC confidence qualifiers — is a secondary but consistent effect. Scenario salience is the smallest channel, which the authors interpret as evidence the cascade does not depend on scenario cherry-picking. However, the analysis fails to recognize that the largest scenario bias effect is already baked into the severity shift.

The figure below from the paper shows that the “amplification cascade” is documented across all IPCC assessment cycles, with the most recent three assessments showing greater bias in the SPMs towards amplifying scientific claims than the first three IPCC assessments.

Three further findings stand out.

• First, in the AR6 cycle, left-leaning and right-leaning media outlets show virtually identical amplification patterns, after a divergence over the three previous cycles.

• Second, the media results are robust across five alternative sample restrictions including removing The Guardian (most amplification), removing the Wall Street Journal (most de-amplification), and equalizing outlet weights — none of which changes the direction or qualitative magnitude of the findings.

  Interestingly, the media outlet judged to have played things the most straight? Fox News!

  

• Third, the pattern reproduces independently across all three LLMs, with Gemini scoring systematically lower than GPT and Claude but all three showing the same directional cascade.

The fact that the media has a bias toward amplifying climate science will come as a surprise to absolutely nobody. However, the finding that the IPCC SPMs also reflect that same bias is a very significant finding.

Why Galiani et al. Matters

Galiani et al. offer the first evidence that the IPCC process amplifies climate science in the direction of more extreme conclusions, beyond those reported in the Technical Summaries, and by extension, beyond that found in the peer revewed literature. Because Galiani et al. look at only a subset of the climate science communication chain, their results should be interpreted as just a floor of possible amplification bias.

As I’ve documented here at THB at length, the over-reliance on extreme climate scenarios in research and policy represents an enormous source of bias, beyond that documented in this new preprint.

It is important to note that while the amplification cascade may serve some political interests, the dynamics at play here require no misconduct or even intent. Each actor might behave rationally within their institutional context and the amplification cascade results even so:

• CMIP scenario developers included SSP5-8.5 as a high-end emissions scenario, not because it was plausible, but because it served the interests of climate modelers.

• Researchers publishing in the literature used the scenarios CMIP made available, and often had good scientific reasons for choosing the most extreme scenarios, regardless its actual plausibility.

• IPCC chapter authors synthesized the published literature — that is their job. The choice by the IPCC to represent scenarios as probabilistic distributions of real-world futures was an error of understanding and judgment.

• IPCC Lead Authors writing the TS carried forward this framing, droping much of the detail that sit inside the scenario and projection black boxes. Choosing to simplify the SPM for policy makers is rational, but in this case context was lost and extreme projections were amplified.

• Journalists apply standard news-value criteria: novelty, urgency, salience. Severe outcomes are more newsworthy than moderate ones. If it bleeds it leads.

The cumulative result is representation of climate science by the IPCC and the media biased toward the extreme end. With RCP8.5 /SSP5-8.5 sitting at that extreme end —while also being represented as “business as usual” or a reference scenario — the stage was set for climate change to be represented in policy as an “existential risk” even though the IPCC chapters and underlying literature never said anything close to supporting such a claim.

Structural problems require structural solutions. Some guidance might be found in a seemingly odd place.

In 2011, a study published in the American Economic Review examined racial bias in ball-and-strike calls across 3.5 million Major League Baseball (MLB) pitches from 2004–2008. Researchers found a small but real bias: umpires slightly favored pitchers of their own race or ethnicity in their calls.

The finding was both interesting and troubling — but the more important finding was what happened next. When MLB installed QuesTec, a computerized camera system that evaluated each umpire’s calls against an objective external standard, the umpire bias vanished entirely. As I discussed on my sports governance blog at the time, umpires whose performance faced independent technological scrutiny made unbiased calls.

The umpires did not have to be corrupt or racist for the bias to emerge. They did not have to coordinate. They were making subjective judgments in real time under pressure, without any external check on the systematic bias in their judgments.

The introduction of an independent external evaluation standard — not punishment, not new rules, not retraining — was sufficient to eliminate the bias. Accountability through objective measurement changed behavior.

The parallel to the IPCC is clear. The amplification cascade that Galiani et al. document demonstrates a systematic bias in how climate science moves through institutional channels to the public. Might AI provide an independent standard for climate assessment against which to assess and correct human biases? I am optimistic that it can.

Another reason why Galiani et al. matters is that it comprehensively refutes claims that climate scientists purposely underplay their findings. In a widely cited 2013 paper, Brysse and colleagues argued that IPCC physical science projections systematically err on the side of caution — that scientists, constrained by professional norms and fear of being labeled alarmist, understate risks relative to what the evidence warrants.

They called this claimed tendency of climate scientists to misrepresent their own work as “erring on the side of least drama” (ESLD).

Earlier, James Hansen made a related claim in a 2007 paper: that “scientific reticence” inhibits individual researchers from stating conclusions as forcefully as the evidence supports, particularly on high-consequence outcomes like sea level rise.

Both arguments have been widely invoked to suggest that the IPCC is intentionally conservative — that the real risks of climate change are worse than what the IPCC reports, and intentionally so.

The Galiani et al. findings directly contradict such claims. The TS-to-SPM and SPM-to-media transitions do not err on the side of least drama — they err systematically on the side of most drama.

In every Assessment Report from 1990 to 2023, the SPM scores more severe than the Technical Summary it is supposed to summarize accurately. In every IPCC cycle from AR3 onward, media coverage scores more severe than the SPM. The idea that climate scientists are intentionally downplaying climate science does not hold up.

Conclusion

The Galiani et al. paper is a potentially significant contribution — the first systematic, cross-validated, multi-cycle empirical measurement of how IPCC climate science is transformed as it moves from technical assessment to policy summary to media coverage.

They finding a consistent amplification bias by the IPCC and in the media, across 33 years, six Assessment Reports, and using three independent AI scoring systems. But the paper is just a beginning. As a working paper it remains preliminary and subject to revision. Independent replication using different methodologies and document sets would substantially strengthen its conclusions.

More importantly, Galiani et al. measure only two transitions in a seven-stage chain.³ The stages they do not assess — from CMIP scenario selection through the published literature to IPCC chapter assessments — carry potential amplification that could also be significant. We already know that the CMIP scenario prioritization process has biased climate science towards implausibly extreme scenarios for decades.

Perhaps most importantly, Galiani et al. offer the promise that AI tools can improve the practice of scientific assessment in climate and beyond.

As I often say, if the IPCC did not exist, we’d have to invent it. Since we have it, the top priority should be to improve it, because assessment matters.

Before you go: A quick request — Please click that “❤️ Like” button. More likes means more readers of posts like this one. Thanks!

Comments, questions, discussion, debate — All welcome!

THB exists because of its subscribers!

1. For an example of “climate telephone” through these stages, see this THB post and this THB post on tropical cyclone intensities and how a major error was introduced into the AR6, confirmed by an IPCC insider.

2. In my discussions of the IPCC assessments, I always refer to the original chapters, and no the Synthesis Reports, Technial Summaries, or SPMs. The chapters sit closest to the literature being assessed and have the lowest risk of “climate telephone.”

3. It would also be interesting to evaluate amplification by IPCC Working Group and even chapter-by-chapter.