Innocent errors are more commonplace than fraud in research papers, so they need to be identified and corrected promptly. We call for more journals to accept their responsibility to ensure that this happens.

We searched for errors in 107 papers in the fields of engineering, materials and computer science, which were based on existing small data sets (see G. Taguchi et al. Quality Engineering Handbook, Wiley; 2004). Our search revealed an alarming number of errors. Ten papers had one or more mistakes that were substantial enough to affect the findings and conclusions. There were errors that were not so significant in almost one-third of the papers.

We notified the journals that published substantial errors. Only four of the ten formally corrected their mistakes. These corrections were published 4–8 months from initial notification. The remaining journals declined to publish a correction; some even had a policy not to publish criticisms of their papers.

Journals that were willing to print general corrections (through errata, letters and notes) were found to have corrected 0.11–0.71% of their existing papers. This represents less than 10% of the error rate exposed here (see also R. D. Chirico et al. J. Chem. Eng. Data 58, 2699–2716; 2013).

Although our selection of journals is narrow, our findings hint that many crucial errors may go forever uncorrected. The risk is compounded by the rarity of attempted replication studies (see go.nature.com/dstij3).

It should be standard practice for journals to insist on full data provision by authors and on declarations of individual author contributions (see, for example, go.nature.com/lwkkqo). Prolonged investigation may be necessary in cases of suspected fraud — but honest errors can be corrected relatively quickly. And this demands a consistent correction policy among journals.