The observed temperature record, which combines sea surface temperatures with near-surface air temperatures over land, is crucial for understanding climate variability and change^1–4. However, early records of global mean surface temperature are uncertain owing to changes in measurement technology and practice, partial documentation^5–8, and incomplete spatial coverage⁹. Here we show that existing estimates of ocean temperatures in the early twentieth century (1900–1930) are too cold, based on independent statistical reconstructions of the global mean surface temperature from either ocean or land data. The ocean-based reconstruction is on average about 0.26 °C colder than the land-based one, despite very high agreement in all other periods. The ocean cold anomaly is unforced, and internal variability in climate models cannot explain the observed land–ocean discrepancy. Several lines of evidence based on attribution, timescale analysis, coastal grid cells and palaeoclimate data support the argument of a substantial cold bias in the observed global sea-surface-temperature record in the early twentieth century. Although estimates of global warming since the mid-nineteenth century are not affected, correcting the ocean cold bias would result in a more modest early-twentieth-century warming trend¹⁰, a lower estimate of decadal-scale variability inferred from the instrumental record³, and better agreement between simulated and observed warming than existing datasets suggest².