The Great Innovation Myth: How Data Errors Created a False Decline in Scientific Disruption

The 2026 Correction: Where the Errors Were

Holst and his team traced the apparent decline to two main sources :

• A software bug in the visualization library. A bug in the plotting software used by the original authors caused papers with the maximum disruption score to be hidden from the published histograms. These papers were still included in the main statistical analysis, but the visual representation—and the unusual concentration of maximal scores—was invisible to readers and reviewers. The Belgian team discovered this when they attempted to reproduce the study using the open data .

• Missing reference data in older publications. The bibliometric databases used in the 2023 study had more complete reference lists for newer papers than for older ones. Older publications, lacking full reference records, were misclassified by the CD index as if they had zero backward citations. The CD index then assigned these papers the maximum possible disruption score. As database quality improved over time, fewer recent papers received this artificial boost, creating a false impression of a downward trend .

When the Belgian researchers filtered out papers with missing reference data—a standard quality-control step—93% of the reported decline between 1945 and 2010 disappeared in the largest dataset from the 2023 study .

A Nearly Three-Year Publication Delay

Holst and his colleagues submitted their critique to Nature in October 2023, just months after the original paper appeared. It was not published until August 2026—a delay of 32 months .

During that time, the original study accumulated over 1,000 additional citations . Nature itself published an editorial and a news feature that treated the decline as established fact . The journal's policy also prevented the critics from speaking to the press while their critique was under review, effectively silencing correction efforts during the period when the flawed finding was most influential .

The delay highlights what the researchers call a systemic problem in scientific publishing: incentives favor eye-catching results over verification, and corrections move too slowly to keep up with the spread of questionable findings .

What the Correction Changes—and What It Doesn't

The Belgian team is careful about what their reanalysis proves. They conclude that "there is currently no convincing evidence that scientific research is less groundbreaking than it used to be" based on the CD-index metric used in the 2023 study . They argue that the methodology and conclusions of the original paper do not provide a sound basis for science policy .

The researchers also caution that even small errors in large databases can drive conclusions that influence global debate. "Our findings serve as a cautionary tale about the reliance on large-scale bibliometric analyses without careful quality control," they write .

The Ongoing Debate

The original authors, Park, Leahey, and Funk, published a peer-reviewed Reply in the same issue of Nature. They argue that the decline remains statistically significant under alternative specifications, including when using Holst et al.'s own dataset and exclusion criteria. They point to their own regression model, which they say yields large and significant declines for both papers and patents .

For their part, the Belgian team notes that when using a stricter filter (papers with at least 10 references), they actually found a slight increase in disruption over time. However, they caution that the CD index itself may not be robust enough to support definitive conclusions either way .

Why It Matters

This debate is not merely academic. The claim that science is becoming less disruptive has been used to justify changes in research funding, evaluation systems, and even the narrative about whether we are living in an era of stagnation . If that claim is largely the result of data artifacts, then policies built on it may be addressing a problem that does not exist.

More broadly, the case underscores the fragility of big-data science. The 2023 paper analyzed tens of millions of records with automated tools, but a single bug in a visualization library and routine database incompleteness were enough to produce a headline-grabbing conclusion that took three years to correct—and that some researchers still defend.

For now, the question of whether scientific disruption is truly in decline remains unresolved. What is clear is that the confidence placed in the 2023 finding was unjustified by the evidence it actually contained.