On July 17, 2026, Steve Kirsch published an article claiming that newly analyzed Louisiana state data proves childhood vaccines increase infant mortality and that “there is no credible way to attack this study.” He goes further, asserting that “all Bradford Hill causality criteria are satisfied” and that the findings represent “game over” for the scientific consensus on childhood vaccine safety.
Those are extraordinary claims.
And extraordinary claims deserve extraordinary scrutiny.
Ironically, the strongest criticism of Kirsch’s article is not that it asks uncomfortable questions. Science should always ask uncomfortable questions.
The problem is that Kirsch repeatedly presents a preliminary analysis as settled science while dismissing nearly all criticism as hand-waving, deception, or evidence that critics are avoiding debate.
That is not how science works.
1. A Preprint Is Not Scientific Consensus
Kirsch repeatedly refers to Brian Hooker’s analysis as if it has fundamentally overturned decades of vaccine research.
But the paper he relies upon was never published in a peer-reviewed journal. According to Kirsch himself, it was removed from Preprints.org and later posted elsewhere.
A preprint is simply a manuscript shared before independent expert review.
Sometimes preprints later become landmark discoveries.
Many others are substantially revised or rejected entirely because reviewers identify methodological flaws, unsupported conclusions, or analytical errors.
Being removed from one repository does not automatically prove censorship.
Likewise, removal alone does not prove the paper is wrong.
Scientific validity depends on whether the methods and conclusions survive independent scrutiny, not on where the paper is hosted.
2. The Study Cannot Measure Overall Vaccine Risk
This is perhaps the most important issue.
Kirsch describes vaccinated versus unvaccinated children.
But the dataset he describes contains only children who died before age three.
There is no comparison group of children who survived.
Instead, the study compares when children who eventually died died.
That is a fundamentally different question than asking whether vaccination increases the overall risk of death.
Without including the millions of vaccinated children who did not die, no study can estimate vaccine mortality rates in the general population.
Even Kirsch’s own description acknowledges that every child in the analytical dataset had already died.
That limitation alone should make readers cautious about sweeping conclusions.
3. Correlation Is Not Automatically Causation
Kirsch repeatedly states that Bradford Hill criteria prove causation.
That is not how Bradford Hill works.
Bradford Hill is a framework for evaluating whether a causal explanation is plausible after considering the totality of evidence.
It is not a mathematical test that automatically converts one observational study into proof of causation.
Researchers typically evaluate:
Biological plausibility
Consistency across independent studies
Strength of association
Experimental evidence
Temporality
Alternative explanations
No single observational study, particularly one with acknowledged design limitations, can establish causation by itself.
4. “Nobody Can Explain It” Is Not Scientific Proof
Throughout the article, Kirsch argues that because critics have not identified a confounder that completely explains the findings, vaccines must therefore be responsible.
That reasoning shifts the burden of proof.
Science does not conclude that a hypothesis is true because competing explanations remain incomplete.
The burden remains on the person making the causal claim to demonstrate that alternative explanations have been adequately ruled out.
Not yet having every answer does not automatically validate one proposed explanation.
5. AI Is Not an Independent Scientific Reviewer
Kirsch repeatedly cites ChatGPT, Claude, Gemini, and Grok as supporting his reasoning or predicting expected outcomes.
Large language models are useful for summarizing literature, explaining statistical concepts, and exploring hypotheses.
They are not independent validators of scientific truth.
An AI model’s output depends heavily on the prompts it receives, the framing of the question, and the information supplied to it.
Using AI predictions as supporting evidence is not equivalent to independent replication by epidemiologists or statisticians.
Conclusion
Questioning scientific consensus is not anti-science.
Declaring victory before the broader scientific community has thoroughly evaluated the evidence is.
If the Hooker analysis raises important questions, those questions deserve careful investigation, replication, transparent statistical review, and publication in the scientific literature.
That is how science corrects itself.
But declaring “game over” before that process occurs risks replacing scientific skepticism with certainty unsupported by the available evidence.
Science advances through replication, not declarations.
Stay tuned for a deeper dive in Part Two.