Annual net ecosystem production (NEP) was estimated at three boreal forest sites where open-path infrared gas analysers were used in eddy covariance systems. The analysers showed apparent ecosystem uptake of carbon dioxide during frozen conditions where small net respiration was expected. Previous studies have shown this to be largely caused by instrument heating creating an additional heat flux that is not used in density corrections. A set of meteorological conditions where good data could be extracted in cold temperatures was not evident. Annual NEP was evaluated based on estimates that used a heating correction of Burba et al. (2008) or removal of data when air temperature <0 °C, with replacement using a regression function based on soil temperature. The methods agreed within ±0.5 t C ha−1 y−1 which is similar to the potential bias that could be caused by density corrections if accounting for energy balance closure at some sites. Annual NEP (±0.5 t C ha−1 y−1) for boreal forest sites that had experienced stand-replacing fire was: a 6-year-old site was losing 0.6, a 15-year-old site was gaining 0.9, and a 27-year-old site was losing 0.8, for 2–3-year averages. It is difficult to reduce the uncertainty in annual NEP estimates at these boreal sites because of the long cold periods. Cold-weather anomalies have been observed by other researchers, but this has not been a universal problem.