Pre-trained language models (PLMs) achieve high accuracy on standard benchmarks for sentiment analysis. However, this performance can hide systematic weaknesses in determining the sentiment of negated sentences, for example when the phrase “not good” is still classified as positive. In this study, we use sentiment classification of English movie reviews in the Stanford Sentiment Treebank 2 (SST-2) as a case study to specifically examine and improve how BERT handles negated sentences. We perform a brief additional fine-tuning of the existing BERT model on a small, automatically constructed set of lexicon-based counterfactual examples that target simple lexical negation. Experimental results on carefully paired original-negated sentences show that this procedure substantially reduces prediction errors on negated inputs while leaving overall performance on SST-2 almost unchanged.